Part: The Page
Chapter 1: An Object Arrives
#We begin with a page.
#The page contains language. It makes a claim. It may contain evidence, a calculation, an observation, a story, a mistake, a discovery, or nothing new at all. Before anyone reads for these things, a newer question arrives.
#Did a machine write it?
#The question appears reasonable because machines now produce language. They produce it quickly, fluently, and in quantities large enough to change the environment in which language is read. A school needs to know whether a student completed an assignment. An editor wants to know what entered a submission. A reader wants to know whom to trust. A publisher needs a process. An institution does not want an undisclosed machine to occupy the place where it expected accountable human work.
#The page is placed into a detector.
#The detector does not interview the author. It does not witness the research. It does not see the first observation, the discarded approaches, the source that changed the inquiry, the false conclusion that survived for three days, the argument with a model, the handwritten note, the program that failed, or the sentence rewritten until it finally held the relation the author meant. It receives the page.
#From the page it measures features. Some words are more expected than others. Some sentences vary in length more than others. A passage may occupy a region of statistical space associated with text produced by a represented model. A watermark may have survived generation and revision. A retrieval system may find a near match in a store of prior outputs. A classifier returns a number.
#The number concerns resemblance.
#The institution asked about authorship.
#Between resemblance and authorship, the object changes.
#A statement about the surface becomes a statement about the process. A statement about the process becomes a statement about the person. A statement about the person becomes a judgment about the work. The detector did not observe these transformations. They occurred in the enclosing institution.
#The page remains on the desk. Its claim has not been read.
Chapter 2: The Impossible Commission
#The easiest story of AI-assisted writing is that a person asks a machine for a finished object and the machine supplies it.
#Sometimes that happens.
#A person asks for a routine letter, a familiar summary, a known conversion, a standard form, or ten variations of an idea already present in the request. The result may be useful and may require little discovery. The prompt names the destination. The machine helps travel there.
#That story cannot describe an inquiry whose destination was unknown to both participants when the session began.
#An author cannot ask a machine to write the discovery that the author has not yet made. The machine cannot retrieve the work from a corpus when the work does not yet exist. The final proposition can be copied backward into an imaginary first prompt, but chronology will not support the reconstruction.
#At the beginning there may be an anomaly, an observation, an unease with the available answer, or an object that does not yet have a name. The author may ask the machine to search, compare, calculate, simulate, argue, draft, or follow a relation into another field. The machine may return a familiar analogue. The author may reject it because the analogue preserves the words and deletes the novelty. A second route may fail differently. A source may reveal that part of the proposition is old. A contradiction may remove the rest. Then, somewhere in the exchange, a relation appears that neither the author nor the machine stated at the beginning.
#Only after it appears can it be requested by name.
#The writing did not merely report the inquiry. The writing was one of the places where the inquiry happened.
#This does not make every generated sentence a discovery. It establishes a more exact object: AI-assisted research content. It is content produced inside a research process where artificial intelligence may participate in traversal, variation, retrieval, comparison, formalization, simulation, criticism, software, revision, and publication. The human may contribute the originating observation, the reason the question matters, the recognition of the signal, the refusal of substitutions, the decision to continue, and the responsibility for the finished whole.
#Neither participant needs to contain the finished object before the process begins.
#We know this form of creation in other enclosures. Two scientists can begin an experiment without knowing the result. A room of musicians can improvise a piece no one brought into the room. Two people in conversation can discover a distinction neither could have stated alone. An author can begin a novel without knowing precisely what the final character will do. The absence of a known destination is not an absence of authorship. It is often the condition under which authorship becomes necessary.
#The machine changes the scale of the room. It does not make the room chronologically impossible.
Chapter 3: The Accusation Before the Reading
#The detector answers a real institutional fear with an answer shaped like the wrong question.
#The fear is loss of accountable human work.
#The detector asks whether the page resembles a class of machine output.
#These questions overlap only under restrictive conditions. They overlap most cleanly when one known model generates an untouched passage, the detector knows the generator and domain, the passage is long enough, the sampling method is represented, the human has not edited it, and no adversary has tried to alter the result. Under those conditions, a statistical method may separate two distributions well enough to be useful.
#The ordinary page is not obliged to preserve those conditions.
#It may begin with a human paragraph, pass through an editing model, acquire a machine-generated transition, lose the transition in revision, retain a machine-proposed ordering, restore the author’s original terms, receive a human conclusion, and be formatted by another system. Its citations may have been found by a search model and verified by the author. Its equation may have been proposed by the author and checked by software. Its strongest sentence may have survived exactly from the originating prompt. Its least interesting sentence may be the most purely human.
#What percentage was written by AI?
#The question assumes language can be poured into the page from two measuring cups. The process may instead contain hundreds of crossings whose importance has no relationship to their token count. One machine sentence can redirect a book. Ten thousand generated words can contribute nothing. A human correction of one preposition can restore the relation that made an entire proposal new.
#The institution nevertheless wants one number.
#When the number arrives first, reading becomes prosecution. Ordinary features become evidence after the verdict. Consistent grammar looks synthetic. Clear organization looks templated. Familiar transitions look generated. An author who writes in a second language may be judged by the predictability required to write safely in it.[12] An author who has learned from machine-mediated prose may reproduce its cadence without using a machine during the page’s creation. An author who knows the detector may add mistakes as proof of humanity.
#The filter does not recover the lost process.
#It changes the language produced inside its reach.
Part: The Machines That Guess
Chapter 4: The Statistical Witness
#A classifier needs examples.
#It is shown text assigned to classes and learns features that separate them. For AI-text detection, one class may contain human writing and another may contain generations from selected models. The classifier estimates where a new passage belongs relative to that represented boundary.
#This can work.
#It can also fail for reasons that are not mysterious. A new model produces a different distribution. A familiar model is sampled differently. A person paraphrases the output. A second model edits the first. The domain changes from essays to code, from news to poetry, or from native to non-native English. The passage is too short. The human class does not represent the human being whose future is attached to the score.
#OpenAI withdrew its own public AI-text classifier in July 2023 because of its low accuracy. In its published challenge set, it correctly identified 26 percent of AI-written text as likely AI-written while incorrectly identifying 9 percent of human-written text as AI-written.[1] The importance of these numbers is not that no better classifier can exist. It is that the organization operating the model recognized the boundary of the inference. It warned that the result should not be used as the primary decision-making tool.
#Larger evaluations have since made the boundary more visible. The RAID benchmark assembled more than six million generations across models, domains, decoding strategies, and adversarial attacks. Detectors that performed well in one enclosure degraded when the generator, domain, repetition pattern, sampling method, or attack moved outside it.[2]
#The classifier does not detect artificial intelligence in the abstract.
#It detects a relation between this artifact and the classes available to this instrument under these conditions.
#That is useful knowledge when it stays in that sentence.
Fingerprints
#Generated language can leave statistical fingerprints without carrying an intentional watermark.
#A model assigns probabilities to possible continuations. Text sampled from that model can occupy characteristic regions of its probability landscape. DetectGPT showed that passages generated by a known model often sit in areas of negative curvature in that model’s log-probability function. Small perturbations can therefore help distinguish generated text from human text without training a separate classifier.[3]
#This is an elegant result.
#It does not identify an author.
#It identifies a statistical relation between a passage and a probability surface. Change the model, perturbation method, decoding process, domain, or editing history and the relation changes. A person can intentionally move the passage. A machine can be asked to make it less detectable. Another machine can paraphrase it repeatedly. A human can independently write into the same high-probability region because the most probable continuation is often probable precisely because human beings have written and expected it before.[13]
#Consider the sentence that begins, “Thank you for your.” A continuation model may assign substantial probability to time, patience, interest, or support. It did not discover gratitude. It inherited a landscape formed by millions of human decisions about what commonly follows those words. Sampling from that landscape does not mean the machine always chooses the single most probable next word. Temperature, decoding rules, system instructions, and randomness can spread or narrow the choice. The causal order remains: the probability is downstream of language people already used.
#This makes the fingerprint real and the inference difficult at the same time. The machine’s ordinary continuation is ordinary because human continuations made it so. A human who writes the common phrase may therefore resemble the model for the same reason the model resembles the human. The measurement has found a shared statistical inheritance. It has not yet found which path through that inheritance produced this sentence.
#The distinction becomes clearer in a room full of witnesses. If nearly everyone remembers the same familiar expression, agreement is evidence that the expression is common. It is not evidence that one witness copied another. The copying question requires contact, chronology, access, or another trace of the event. Probability can tell us where the page sits in the crowd. It cannot by itself tell us how the page arrived there.
#The fingerprint belongs to the event represented by the measure. It does not contain the complete chain of custody.
#The distinction resembles a fingerprint found on a book. It may establish contact under sufficiently controlled conditions. It does not tell us who conceived the argument, who corrected it, who accepted it, who published it, or whether the argument is new.
Chapter 5: The Mark
#A watermark changes the problem.
#Instead of inferring origin only from naturally occurring statistics, the generator intentionally modifies token selection so that a later test can recover a pattern. A secret key may divide candidate tokens into favored and unfavored sets. The generator slightly increases the probability of favored tokens. No single word proves anything, but a sufficiently long passage can accumulate a detectable signal.[4]
#Google’s SynthID similarly embeds signals during generation and provides a detector for supported content. Its designers describe watermarking as one part of a larger system, not a complete answer to misattribution.[5]
#A retained watermark can provide stronger evidence of a compatible generation event than a general style detector. It may answer: did text from a supported marked generator probably survive here?
#It still does not answer:
- Who originated the inquiry?
- What did the human contribute?
- Which claims were accepted or rejected?
- Was the marked passage revised?
- Is the whole work authored by the generator?
- Is the work novel?
- Is it valid?
- Is it valuable?
#Absence of the watermark cannot certify human-only production. The generator may not support it. Editing may weaken it. Translation or paraphrase may remove it. A short passage may not contain enough signal. An unmarked system may have participated elsewhere.
#Presence and absence have different enclosures. Neither is a moral verdict.
#The false positive is not an incidental embarrassment. It is present in the mathematics of the test.
#A watermark detector counts a pattern and compares the count with a threshold. If an unmarked passage happens to contain enough favored tokens, it can cross that threshold. Designers can make the probability very small. They cannot make a statistical tail cease to exist without changing the claim from a statistical inference into direct provenance. Human writing produced before the watermark was invented therefore contains some nonzero territory capable of receiving a watermark score. The text cannot have been generated by the later marked system. The score can still occur.
#This is not a proof that watermarking is useless. A retained mark from a known generator can be valuable process evidence. It proves something more modest and more defensible: the observed sequence is sufficiently compatible with the marked process under the detector and threshold being used. The published false-positive rate belongs beside the finding because the alternative source class never became impossible.
#Spoofing exposes the boundary from the other direction. Research has shown that an attacker can learn enough about a hidden signature to make human text look marked.[13] Paraphrasing can weaken a mark that was genuinely present. The signal can be introduced without the event and removed after the event. The causal history is richer than the final count.
#Now place the finding in front of a novel manuscript. The watermark test can answer its process question. It cannot tell whether the manuscript’s central relation existed before, whether its evidence holds, or whether a reader will find a world inside it. Using the score to establish intellectual value does not extend the test. It changes the subject while carrying the numerical authority forward.
Chapter 6: Provenance and the Bounded Instrument
#Provenance does not ask the page to confess what happened before it existed.
#It records what happened while it was happening.
#A revision history can show the sequence of edits. A repository can preserve commits. A research deposit can establish the public state of an idea at a date. A model-use statement can identify the systems and functions involved. Signed content credentials can bind declared events to a file through a chain of cryptographic assertions. The C2PA specification provides one technical architecture for such records.[6]
#NIST distinguishes provenance tracking, watermarking, labeling, and detection because they are different instruments.[7] This is more than administrative clarity. It prevents one method from being promoted into the job of another.
#Provenance can say that a supported tool created or modified an asset. It can say that a named party signed a declaration. It can preserve a chain of operations. It cannot make the claim inside the artifact true. A perfectly documented falsehood remains false. An undocumented discovery may remain new.
#The truthful form of provenance is not purity.
#It is history.
#The demand that every private research exchange be exposed would create another failure. Inquiry requires unobserved space. Sources can be sensitive. Prompts can contain unpublished ideas, personal information, adversarial tests, or records belonging to others. A lawful provenance requirement asks for what is material to the decision and preserves the Final Frontier of the research enclosure.
#Recorded history is better than inferred resemblance when history is the question.
#Neither replaces reading.
The bounded instrument
#The methods can now be placed in their lawful domains.
#A classifier estimates compatibility with represented classes.
#A statistical fingerprint measures a relation between the text and a model or distributional surface.
#A watermark detector tests for an intentionally embedded signal from a supported generation process.
#A provenance record preserves declared or signed events in the production history.
#Each can contribute evidence.
#None can silently become authorship, novelty, reason, validity, or value.
#The failure is not the existence of the instrument. The failure is enclosure expansion without evidence. A bounded result crosses into a larger claim while retaining the confidence of the smaller one.
#The page resembles a represented machine class.
#Therefore a machine produced the page.
#Therefore the human did not write it.
#Therefore the human did not think it.
#Therefore the work is not new.
#Therefore the work has no value.
#No detector performs these steps. People and institutions perform them after the detector stops.
#The unjustified certainty belongs to the enclosing system.
Part: One Language
Chapter 7: The Machine Learned From Us
#The machine writes like a human because human language is what it was trained to continue.
#This fact is so obvious that its consequence is easy to miss. Machine-like language began as human-like language. The model learns regularities from human artifacts: books, articles, code, conversations, manuals, websites, translations, corrections, templates, and the accumulated forms by which people have made language predictable enough to understand.
#The generator does not cross into a separate language when it produces a sentence. It operates inside the statistical architecture of ours.
#We then command it to become more specific.
#Read these books. Learn this voice. Preserve these terms. Follow this rhythm. Do not substitute a familiar analogue for the novel relation. Write as the author writes.
#The machine is actively instructed to move toward the person.
#Then the person reads the result.
We learned from it
#People naturally write like what they read.
#Language has always moved this way. Children acquire syntax from the language around them. Genres stabilize through imitation. Scientific articles inherit the structures of prior articles. Legal clauses persist because each drafter reads earlier clauses. A generation of novelists changes what the next generation experiences as an ordinary sentence.
#AI-generated language has entered that environment.
#It appears in correspondence, search summaries, advertisements, support documents, schoolwork, reports, software, news, social media, and the drafts people edit all day. A human need not invoke a model during the writing of a particular page to have learned from machine-mediated language. The adaptation is passive in the immediate event and active across the larger linguistic enclosure.
#Recent large-scale research has begun to measure this exchange. Across more than 880,000 texts in seven datasets, Sourati and colleagues found evidence associated with widespread LLM writing assistance and declining linguistic diversity.[8] The result does not mean every person will sound the same. It establishes that the environment is changing at the scale where detection assumes stable human and machine distributions.
#The human moves toward the machine because the machine has entered what the human reads.
#The machine moves toward the human because it is trained on human language and explicitly commanded to imitate particular humans.
#The detector is placed between two moving distributions and asked to identify which one owns the sentence.
Chapter 8: The Bias Has No Outside
#The machine is often described as biased, followed by the hope that another procedure can make it unbiased.
#There is no view from nowhere available to the model. There is training data, selection, weighting, objective, instruction, evaluation, reinforcement, retrieval, context, and the projected reader. Removing one bias changes the conditions under which another becomes active. Adding counterexamples changes the distribution. Rewarding a safer response favors one region of possible language over another. Even the request to be neutral projects a human idea of neutrality inward.
#Bias here is not merely prejudice. It is direction. A system without any direction would have no basis for producing one continuation instead of another. There is no anti-bias and no bias-bias waiting outside the system. There are visible and hidden conditions that make some continuations more available than others.
#Probability does not remove this fact. A next-token model estimates possible continuations from relationships learned in its training and shaped by its present instructions. When a common word receives high probability, that is not an independent machine vote for the word’s meaning. It is evidence that the model encountered enough related human choices for the continuation to become strongly available inside its learned representation. The majority is not always correct, and generation does not always select the maximum. Yet the statistical center of the system is necessarily pulled toward represented regularity. Majority validation is not a political opinion held by the machine. It is a geometric consequence of learning what recurs.
#Novel prose lives at an awkward boundary. If it is merely random, it carries no relation for a reader to recover. If it is entirely predictable, it may carry nothing the reader has not already received. A new work must remain linguistic enough to enter the shared enclosure and different enough to add a world to it. The difficult object is not unfamiliar wording. It is unfamiliar coherence.
The third process
#The binary fails before imitation is considered.
#A human writes a paragraph and asks a model to improve its clarity. A model generates a paragraph and a human rewrites every sentence. A human supplies the conceptual structure and a model supplies connective prose. A model finds sources and a human constructs the claim. A human writes the prose and a model changes only punctuation. A translation model carries a human argument into another language. A speech recognizer converts a human voice into tokens and an editing model removes repetition.
#Human or machine?
#Research on mixed text finds that systems built to distinguish fully human from fully machine-generated passages do not simply inherit their performance when production alternates within a document or when one edits the other.[9] The collaborative artifact is not an impure instance of one production class. It is a third process with its own chain of causality.
#It is not a third language.
#Even third process is too simple. The number of possible crossings is not three. The useful distinction is not a zoology of pure texts. It is an account of which operations occurred, what linguistic enclosure made every operation possible, and which person accepts responsibility for the result.
#The detector asks the artifact to recover a production history that the artifact was never designed to store.
#When it cannot, the institution calls the ambiguity deception.
#The ambiguity was built into the question.
#The machine-generated sequence is a real event. Calling that event real does not require inventing a language that did not descend from human language or an author who can answer for publication. In this model, all socially operative language remains inside the human enclosure. The machine can produce a sequence, alter a sequence, and participate materially in the creation of a work. The work becomes an act of communication through human purpose, interpretation, acceptance, and consequence.
#This is why the phrase AI writing can be useful as process shorthand and destructive as an authorship verdict. It names that a machine generated text. It does not establish that the human contribution is absent, that the text has no meaning, or that the artifact has no author. The process has changed. The enclosure in which authorship becomes possible has not moved into the machine.
Part: The Enclosure
Chapter 9: The Range of the Inquiry
#The page is not alone.
#It is enclosed by a research process, a publishing system, an institution, a language, and the human beings for whom its claims can mean anything. It encloses sentences, propositions, evidence, citations, figures, equations, errors, and relations. Those things enclose smaller things in turn.
#The detector is also enclosed. Its training set, represented generators, threshold, purpose, vendor, interface, and deployment policy determine what its number can mean. The person reading the number is enclosed by a deadline, an incentive, a fear of being deceived, and a rule about what should happen next.
#Nested Causal Modelling begins by refusing to let any one of these objects occupy the whole map.
#NCM represents an observed thing in the order:
[thing enclosing the thing | the thing | things the thing encloses]
#The order matters. It prevents a local mechanism from pretending it supplied its own conditions. It also prevents an outer condition from erasing what the thing itself does.
#The available enclosures do not end at the page. Escalation can continue from person to institution, culture, civilization, planet, stellar environment, galaxy, the Great Attractor, and the most distant scale the present model can name. De-escalation can continue through paragraph, sentence, token, encoding, hardware, matter, Planck scale, and whatever lawful enclosure lies beyond the smallest scale presently measured.
#Nested infinities make those directions available. They do not require every NCM to calculate them.
#This inquiry began with a specific question: when humans and machines have been made to imitate one another inside one changing language environment, can resemblance to machine output determine authorship, and should that resemblance displace novelty and reasoned ideation as the signal of value?
#That question establishes the active range.
#The outer Final Frontier is human consciousness as the enclosure in which language becomes available, meaningful, and answerable to another person. The inner Final Frontier is the artifact surface where classifiers, fingerprints, and watermarks measure features. The map crosses the research event, model, language, institution, author, and reader because each is active in the authorship inference. It stops before the biological homeostasis that made brains possible and before the physical construction of the transistor. Those enclosures remain real. They are not active in the question this model is built to answer.
#This is how NCM remains computationally available. The Final Frontier does not announce the end of causality. It declares the end of the model’s present authority.
#Within that range, the system is mapped by three observations. Every active thing is enclosed and encloses other things. What crosses from one enclosure to another must be accounted where it leaves and where it arrives. Every enclosure has additional inward depth and outward scale even when the present map lawfully stops before reaching them.
#The artifact can therefore be written:
[research, publication and evaluation enclosure
| AI-assisted research artifact
| propositions, evidence, reasoning, prose, data, code,
figures, citations, provenance records and revision history]
#The event that produced it is a different thing:
[human linguistic, technical and publishing enclosure
| AI-assisted content-generation event
| prompts, retrieved sources, model inference, tool calls,
candidate outputs, human selections, edits and rejections]
#The artifact and the event are adjacent. They are not interchangeable.
#An event can produce an artifact. An artifact can preserve evidence of an event. The surface of the artifact may not preserve enough information to reconstruct the event. This is not a defect unique to AI. A printed book does not retain every conversation, deleted sentence, library visit, memory, or moment of recognition that produced it. Authorship has never meant that the final object contains a complete forensic image of its creation.
#Two causal directions remain visible inside the map. Nonvolatile causality projects from an outer enclosure inward. It is reversible and subject to human modification. Volatile causality moves from enclosed mechanisms upward. It is irreversible. These terms do not mean safe and dangerous, low-risk and high-risk, or weak and strong. They identify both direction and the corresponding reversibility of the event in relation to entropy.
#The outer detector policy is nonvolatile relative to the writer it encloses. It can be withdrawn, modified, or replaced, changing the conditions it projects inward. The writer’s token choices and edits are volatile relative to that policy. A token can later be deleted and a publication corrected, but those actions do not reverse the event that produced the prior state. They create subsequent events in the causal history.
#The demand for machine detection makes this ordinary boundary newly visible.
Chapter 10: The Human Enclosure of Language
#Language does not begin in the machine.
#For this inquiry, its active outer structure is:
[human consciousness | linguistic capacity | parallel human languages]
#Human consciousness encloses the capacity for language. That capacity encloses many languages. English does not enclose Arabic, Mandarin, Spanish, Hindi, or any other language. They are parallel enclosures, each sustained by people learning, speaking, writing, interpreting, correcting, remembering, and changing it together.
#The biological and cultural history beneath this arrangement is immense. Research on language evolution describes linguistic structure as emerging through interaction among biological capacity, individual learning, and cultural transmission.[14] NCM does not need to replay that history here. It needs only the active result: human communities provide the stabilizing force through which linguistic forms carry shared meaning over time.
#The language model is enclosed within that human structure:
[human linguistic and technical enclosure
| language model
| training representations, parameters, token probabilities,
prompts and generated sequences]
#Its training text was produced, selected, collected, labeled, or preserved inside human language. Its objectives, architecture, permissions, prompts, evaluations, and uses are supplied within the human enclosure. Large language models can transform these materials with extraordinary range, but their capacity begins after language already exists. GPT-3, for example, was trained as an autoregressive model on large web and text corpora and was able to produce news samples human evaluators had difficulty distinguishing from human-written articles.[15]
#This does not require the claim that a model contains no useful semantic representation. It identifies the enclosure that stabilizes meaning. Bender and Koller distinguish linguistic form from communicative intent and argue that a system trained only on form cannot acquire meaning from form alone.[16] Whatever position later research takes on a model’s internal representations, the generated sequence becomes part of language socially only when human beings prompt it, interpret it, reject it, accept it, publish it, answer it, or act upon it.
#The model does not project the original stabilizing force that made language possible. Humanity projects language into the machine.
#The return crossing is nevertheless real. Machine-produced sequences enter emails, reports, books, search results, advertisements, correspondence, and conversation. People consume them. The patterns become part of what people have read. They enter later human writing because people naturally write from the language around them. That human writing enters future corpora. The loop can be shown:
human language -> model training -> generated sequence -> human reading
-> human expression -> human language
#The model’s output therefore projects patterns back into individual human language while remaining enclosed by the larger human linguistic system. The machine imitates humanity by training and command. Humanity begins to imitate the machine through consumption. Evidence of declining linguistic diversity during widespread LLM assistance is consistent with that return crossing.[8]
#The distributions a detector wants to call human writing and AI writing are not stable species. They are interacting states inside one human language enclosure. The causal separation required to recover authorship from style is not present.
#There can be a machine-generated token sequence. There can be a marked generation event. There can be a third production process containing human and machine operations. These are valid statements about provenance.
#There is no separate AI language standing beside human language and supplying an independent authorship class.
#There is only human writing and an attempt by a machine to speak on humanity’s behalf.
#It is all human language.
Chapter 11: Zero Infinity and the Backward Chain
#Every comparison requires a reference.
#The detector supplies one: the statistical boundary between represented classes. That reference is lawful for the detector’s bounded result. It is not the reference for authorship or value.
#Nested Causality uses Zero Infinity, written , as the balanced reference from which an observable difference can be expressed without mistaking the reference for a thing inside the system. It is neither ordinary zero nor completed infinity. It does not announce nothingness. It makes the unoccupied reference available while the map moves through countably many possible depths and scales.
#For the present inquiry, the reference is simple:
#A research artifact exists. Its production process, authorship, novelty, reasoning, validity, utility, and responsibility have not yet been decided.
#From that reference, each observation can move in its own dimension without dragging the others with it.
#A watermark may move the provenance coordinate without moving novelty.
#A prior publication may move novelty without changing authorship.
#A failed calculation may move validity without removing originality.
#A correction may move reliability without changing who accepted responsibility.
#The current authenticity system does not preserve this balance. It allows one coordinate to move the whole object. Similarity becomes the reference from which authorship, quality, and misconduct are declared.
#The page begins accused.
#Zero Infinity restores the state before the accusation and allows the evidence to move only what it can reach.
The backward chain
#The detector observes the end of a process and reasons backward:
artifact surface -> statistical class -> alleged process -> alleged author
#Each arrow crosses an enclosure boundary.
#The first arrow may be computed. The passage is compatible with a represented class at a stated score.
#The second requires knowledge of generation conditions. Similar distributions can arise through different processes. A marked generation provides stronger process evidence, but only for the retained mark and supported system.
#The third requires a theory of authorship. A machine’s participation does not identify who originated the inquiry, controlled the research, selected among possibilities, preserved the novel relation, or accepted responsibility for the whole.
#The chain usually continues without being written:
alleged author -> alleged originality -> alleged validity -> alleged value
#These are not smaller versions of the same question.
#Someone can author an unoriginal work.
#Someone can originate a false idea.
#Someone can use AI to produce a valid but familiar result.
#Someone can produce a novel and valid object with no immediate use.
#Someone can produce an enormously useful artifact by recombining known parts.
#Someone can write every word and contribute nothing.
#The chain is attractive because it makes evaluation inexpensive. A difficult reading is replaced by a score. The cost does not disappear. It moves into false accusation, evasion, homogenized language, concealed assistance, and the loss of work that the institution was built to find.
Chapter 12: The Active Enclosure
#The detector is not the active enclosure of the failure.
#The active enclosure is the institutional demand to classify provenance from semantic surface before evaluating contribution.
#The detector is an instrument through which that demand acts.
#Once this enclosure is visible, several otherwise strange outcomes become coherent. A person is rewarded for adding random errors because errors move the text away from a represented machine class. A machine is asked to rewrite generated text until a detector calls it human. A clear writer is asked to prove humanity by producing process records that were never required when unclear writing was the norm. A multilingual writer is penalized for using the predictable forms that make a second language safe. An evaluator spends more time on authenticity theater than on the proposition being evaluated.
#The word slop seems to name a machine condition. Most often it names an old human experience at a new volume: content that is repetitive, empty, careless, unwanted, or bad. Generation lowers the cost and increases the speed at which bad content can be produced. It does not create bad ideas, and it does not turn provenance into a measure of quality.
#A poor prompt can produce a poor answer. A familiar idea can be spoken with extraordinary eloquence and remain familiar. A bad idea can be polished until its weakness is pleasant to read. A human being can do all three without a machine. The quality problem is real. The proposed authorship class does not solve it.
#Once the label is allowed to precede the reading, it can become a reason not to read. Human judges have performed poorly when distinguishing machine text in controlled studies and have relied upon cues that models can reverse.[17] Experimental evidence also indicates that a source label can change how the same kind of work is evaluated and edited.[18] Together with the documented bias of detectors against non-native English writers,[12] this exposes a larger danger. AI-generated can become a technologically respectable way to dismiss a person whose language, identity, productivity, or ideas an institution already finds undesirable.
#Readers and markets can still decide what they want to read, buy, recommend, publish, or ignore. That decentralized judgment is useful because literary preference is subjective and because sustained attention is scarce. It is not a truth machine. Sales do not establish authorship, novelty, or validity. They establish that people exchanged attention or money for an artifact. The market can filter preference after the work is available. A detector that prevents the work from reaching the reader is a gate imposed before that judgment can occur.
#The good idea is not endangered by rigorous reading. It is endangered when a source stereotype replaces reading.
#The intervention cannot be a better accusation.
#It must restore the questions to the enclosures that can answer them.
Part: The Lost Signal
Chapter 13: Similarity Is Not Novelty
#Similarity can be measured at the sentence surface while novelty resides in a relation.
#The words may all be old.
#The objects may all be known.
#The novelty can be the first lawful placement of one beside another.
#A new scientific question may use ordinary vocabulary. A new intervention may combine familiar infrastructure under a relation no prior system recognized. A proof may reorganize established statements. A map may make an existing quantity visible at a scale where it can finally be acted upon. A correction of one causal direction can change the meaning of an otherwise unchanged model.
#Research on scientific novelty has repeatedly found that novelty can appear in uncommon combinations of familiar components and in new questions, methods, or relations, not only in unprecedented words.[10]
#This is why a style detector is positioned so poorly to measure it.
#The detector attends to regularity in expression. Novelty may be carried by a regular sentence. The detector attends to how likely the next token appears. Novelty may be the relation that determined why that token needed to be there. The detector attends to resemblance across corpora. Novelty requires comparison with the nearest prior idea while preserving the exact difference.
#The two analyses can coexist.
#They cannot replace one another.
Prior capture
#Novelty is often lost without being contradicted.
#An unfamiliar proposal is translated into a familiar analogue. The terms are retained. The syntax improves. The text becomes easier to classify, cite, and defend. The relation that made the proposal new disappears.
#This is prior capture.
#Artificial intelligence is especially capable of it because language models are powerful continuators of known patterns. Asked to clarify a novel idea, the system can replace it with the nearest high-probability idea and describe the replacement beautifully. A human editor can do the same thing. Peer review, policy review, legal drafting, and institutional risk management have performed prior capture for centuries.
#Similarity detection magnifies the error when it treats movement toward known language as evidence of human legitimacy while treating unfamiliar coherence as suspicious. The writer learns to survive by sounding like the corpus. The machine learns to survive by sounding less like the detector’s machine class. The system selects for disguise rather than contribution.
#The noise filter removes the anomaly.
#The anomaly may have been the signal.
#Preserving novelty does not mean protecting every strange sentence from correction. It means identifying the claimed new relation before rewriting it. Then the sentence can be improved without replacing the thing it carries.
#The source prompt, first articulation, rejected analogue, revision history, and nearest prior work become part of a fidelity record. The record does not prove the idea true. It preserves the idea long enough to find out what it is.
Chapter 14: The Timestamp and the Idea
#A DOI does not certify truth.
#It does something quieter and indispensable. It places an object into a durable public sequence.
#At this date, this version existed at this address under this name.
#A repository commit can preserve a finer sequence. A source deposit can show which proposition preceded a correction. A timestamp can establish that an idea was public before a later analogue. Subsequent versions can disclose development without forcing the first appearance to vanish.
#This is a better instrument for priority than stylistic suspicion. It does not ask whether the prose sounds human. It asks what entered the record and when.
#The high volume of novel ideas in an AI-assisted body of work is therefore not an embarrassment to be discounted until human purity can be proved. It is an observable research object. Each claimed contribution can be compared with prior work, traced across dated versions, tested for preserved difference, and followed into whatever it makes possible.
#The presence of AI assistance belongs in the provenance.
#The novelty belongs in the contribution.
#The DOI timestamp belongs in the chronology.
#None should be asked to impersonate the others.
Reasoned ideation
#Novelty alone is not enough.
#A random sequence can be new. An incoherent claim can be unprecedented. An error can enter the public record before anyone else makes it.
#The signal required for research is reasoned ideation: a new proposition, object, relation, method, map, or intervention connected to a visible chain of reasoning and made answerable to evidence.
#Reasoned ideation asks:
- What was observed?
- What reference state was used?
- Which terms carry the proposal?
- What encloses the object, and what does it enclose?
- Which outer conditions project inward?
- Which local mechanisms act upward?
- What crosses the relevant boundaries?
- Which prior analogue is nearest?
- What exact difference must survive comparison?
- What observation, calculation, or consequence supports the relation?
- What would correct or defeat it?
- Where does the claim stop?
#This is more demanding than detection.
#It is also the work institutions intended to protect when they began worrying about generated language.
#If an AI-assisted artifact survives this inquiry, its value does not depend on pretending that no machine was present. If it fails, the failure can be named: the source does not support the claim; the relation was already known; the calculation does not reproduce; the model confuses simulation with observation; the proposed intervention does not follow from the map; the Final Frontier was crossed.
#These are useful failures.
#“It sounds like AI” is rarely one of them.
Chapter 15: Shakespeare at the Detector
#Imagine that Shakespeare arrives with a page.
#No name appears on it. No portrait accompanies it. The institution receiving the page has been warned that generated language is abundant. It does not begin with the play. It begins with the possibility that the page is false.
#The words have all been used before. The meter belongs to an inherited form. The plot resembles older stories. The characters carry kings, lovers, murderers, fools, ghosts, wars, and family disputes into arrangements made from materials already present in history, legend, theater, and speech. The page contains the language of everyone the writer has heard.
#This is not a defect in the work. It is the condition under which writing can be read.
#Now train a detector on the language that came after Shakespeare. His phrases have entered classrooms, criticism, common speech, theater, film, parody, and the work of writers who learned what an English sentence could do by reading him. A later corpus contains his influence everywhere. The detector compares the page with a statistical world partly shaped by the page’s descendants. Resemblance is guaranteed by cultural success.
#The machine may report that the writing is highly compatible with the corpus. It may find repeated structures, familiar transitions, predictable meter, and phrases that appear too often elsewhere. Every observation can be accurate. The authorship conclusion can still run backward.
#The page does not resemble later writing because it copied all of it. Some of the later writing resembles the page because people carried the page forward. The artifact entered the enclosure and changed the reference against which it would later be measured.
#This is the chronology a surface detector cannot recover.
#The thought experiment does not depend upon Shakespeare being beyond criticism. If the plot has a prior source, name it. If a line has already been written, place the two lines together. If the history is wrong, correct it. If the meter fails, let the ear hear the failure. If the arrangement adds nothing, comparison can establish that. These are stronger challenges than a purity test because they reach the work.
#Suppose instead that the evaluator announces that the page probably came from a machine and therefore is not worth reading. The work has not failed. The institution has declined to encounter it. A provenance suspicion has occupied the place where literary judgment, source comparison, and historical sequence were required.
#The same error can be made in reverse. A human byline does not make an empty page important. Proof that a person typed every word cannot make a dull book good. Human origin is not a certificate of value any more than machine assistance is proof of its absence.
#The reader remains necessary.
#This is why novelty cannot be reduced to rare vocabulary. Shakespeare did not need to invent a private language no one could understand. The contribution appeared in what familiar words, inherited forms, and borrowed materials were made to do together. The nearest analogue helps reveal that difference. It does not erase the author by existing.
#Someone should be told that Shakespeare already did that one.
#They should not tell Shakespeare he did not write it.
#Between those sentences lies the difference between finding prior art and inventing prior capture. One comparison tells the new writer where the idea has already been. The other uses resemblance to deny the event in front of us. A research culture able to protect novelty must know which direction in time it is looking.
Part: The Author
Chapter 16: Authorship After Assistance
#Authorship is not token manufacture.
#The author is the person who accepts responsibility for the whole.
#This responsibility includes the choice to use a model, the failure to verify a citation, the decision to preserve a generated sentence, the claim made by an equation, the omission that changes the reader’s understanding, and the publication of the final object. Assistance can be extensive. Responsibility does not become diffuse merely because the apparatus becomes powerful.
#This standard also protects readers from a convenient evasion. An author cannot publish a false claim and blame the model as though the model submitted the manuscript independently. A provider does not become the author of every work produced with its system. The system does not experience the consequences of publication or decide to stand behind the work.
#When a machine voice is intentionally presented as a participant, it can be identified. When a source passage is quoted, it can be cited. When a model substantially assists the research and writing, that can be disclosed. These acts make the process more visible.
#They do not turn authorship into a percentage of keystrokes.
Fiction is different
#Research content and fiction do not ask the same thing of authorship.
#In research, artificial intelligence can ethically support grammar, spelling, source discovery, comparison, calculation, simulation, and the evaluation of new ideas. The resulting proposition is judged by novelty, reason, evidence, reproducibility, and accepted responsibility. A tool may help articulate the idea without becoming the source that proves it.
#Fiction carries another requirement.
#The writer must have written the story.
#The story must also be good.
#The second condition has never been guaranteed by proof of authorship. Markets, readers, critics, time, and the private life of the work have decided it imperfectly. Courts usually enter after value exists, when money, copying, attribution, contract, or damage has made the dispute worth litigating.
#The unique creations of an author must be protected before and after that value becomes visible. The character no one else had met, the sentence whose cadence belongs to one body of work, the arrangement of events, the private myth, the fictional world, and the exact expressive choices that make the story itself are not merely research results waiting to be source-checked. They are the authored object.
#Society objectifies prose and then commodifies it. A manuscript becomes a file, a license, an inventory item, a rights territory, a royalty stream, a training example, and a unit that can be copied at negligible cost. Yet the commodity is not made authentic by the same process that prices it. An AI forgery can carry polished prose. A human original can be awkward. A detector whose own performance is measured by its ability to distinguish a machine writing like a human cannot prospect the future value of fiction or defend the author’s unique creation after imitation succeeds.
#Watermarking offers process evidence, but it approaches language as a marked commodity before anyone cares about what the language contains. It modifies generation so the machine can later be recognized in the text. In fiction, that operation introduces the very object the protection was meant to exclude: writing deliberately shaped by a nonhuman requirement.
#The green-list token is preferred not because the character would say it, not because the rhythm requires it, and not because the author chose it, but because a detector may need to recover it later. The machine’s administrative need enters the sentence. The watermark can remain statistically subtle and still reveal the category error. The story is being prepared for authentication before it has earned attention as a story.
#A fiction writer may use computational tools around the work. Spelling, grammar, research, continuity checks, file conversion, and the examination of alternatives can support the author. The boundary is the unique expressive creation. The author must remain the source of the story rather than the person who merely selected a story generated for them.
#This boundary is not measured by whether a detector can find machine-like prose. It is established through the creative record and the work itself: notes, drafts, revisions, character histories, source material, correspondence, testimony, contracts, dated files, and the continuity of the author’s choices. The United States Copyright Office has preserved the human-authorship requirement while recognizing that AI used as an assistive tool does not by itself prevent copyright protection; protectability depends on the human expressive contribution in the particular work.[11]
#Research asks whether the idea is new and survives reason.
#Fiction asks whether the writer wrote the story and whether anyone wants to keep reading.
#Neither question can be answered by similarity alone.
Chapter 17: The Instrument and the Author
What the institution gains
#The Signal Protocol is not lenience toward generated work.
#It is a stricter evaluation.
#The student who submits fluent nonsense cannot survive an oral account of the reasoning, a source check, or a reproducibility test. The researcher who hides machine assistance cannot supply a truthful provenance statement. The paper that recombines familiar claims cannot claim novelty after comparison with the nearest prior work. The model-generated citation that does not exist fails at the source. The elegant simulation presented as observation fails at the boundary.
#At the same time, the institution stops punishing clarity, multilingual writers, disciplined structure, and people whose language has adapted to the same environment as everyone else’s. It stops teaching evasion as the primary AI skill. It stops rewarding the addition of noise.
#The institution gains something detection alone could never supply: a map of what the work contributes.
What the author gains
#The author gains the right to disclose the apparatus without having the apparatus substituted for the work.
#That right carries an obligation to preserve enough process for the material claims to be examined. Original prompts can protect novelty from later normalization. Source-fidelity ledgers can identify where an edit retained the words and replaced the relation. Version histories can show development. Quarantined drafts can preserve the evidence of conceptual drift. DOI deposits can establish public sequence.
#These are not rituals of innocence.
#They are research instruments.
#The author also gains a clearer standard for using AI. The system can be asked to search widely, argue hard, simulate readers, transform formats, construct software, and expose contradictions. Its fluency is not mistaken for evidence. Its familiarity is not mistaken for fidelity. Its resistance is not mistaken for rigor. Its agreement is not mistaken for validation.
#The author remains responsible for recognizing the signal.
What the machine cannot decide
#The same machine may read the corpus, learn the author’s vocabulary, imitate the author’s voice, help produce the next work, and then be asked whether the work was written by the author or by a machine.
#The request is not merely difficult.
#It removes the stable distinction the answer requires.
#The machine was instructed actively to write like the person. The person was changed passively by reading language produced by machines trained on people. The final artifact may contain direct contributions from both, indirect influence from both, and relations that emerged only in the exchange.
#From the semantic surface alone, no universal procedure can recover full authorship across that process. Under bounded conditions, an instrument may identify compatibility, contact, a retained watermark, or a recorded event. Those are lawful findings. They do not restore the vanished binary.
#The machine can help compare the proposition with prior work.
#It can help test the reasoning.
#It can help locate evidence.
#It can help reconstruct a version history.
#It can help identify which words are familiar.
#It cannot decide who is responsible for the whole.
The drafting table
#The closest precedent may not be another writing machine.
#It may be AutoCAD.
#Computer-aided drafting did not prove that a person had never drawn a line. It changed the enclosure in which a line could be designed, measured, revised, copied, tested, and carried into manufacture. A drafter could snap to an exact intersection, preserve dimensions, reuse components, undo a mistake, compare alternatives, and transmit a drawing without redrawing the whole object by hand. The pencil did not become immoral. The computer did not become the architect. The reachable precision and rate of iteration changed.[19]
#No serious account of the resulting building asks what percentage of its lines were drawn by the software. The useful questions concern who designed it, whether the dimensions are correct, whether the structure stands, whether the drawing can be manufactured, and who accepted responsibility for the plans. The tool is present in every answer without becoming the answer to all of them.
#AI-assisted writing can occupy a similar relation to thought without pretending that sentences are beams or that prose has engineering tolerances. The author can test ten arrangements while a human editor might have time to discuss one. The system can retrieve an adjacent field before the conversation would have reached a library. It can expose a contradiction immediately, generate the opposing argument, compare a term across hundreds of pages, and remain available for another iteration without the interpersonal friction of exhaustion, status, embarrassment, or wounded taste.
#This does not make bad writing good by contact. A single instruction can produce a book-shaped quantity of fluent text whose world is no larger than the instruction. The author who accepts it has accelerated production without deepening authorship.
#The iterative process can nevertheless improve the author. A weak sentence can be compared with stronger ones. A hidden assumption can be named. A source can force a correction. A simulation of readers can expose where the author mistook private vocabulary for common understanding. The author returns with more knowledge than the author brought. The next decision changes because the person making it has changed.
#That is not the machine making a bad author good.
#It is a powerful editorial and research enclosure allowing an author to learn faster, provided the author continues to reject what is easy, verify what is factual, and recognize what belongs to the work.
#The test is not whether the instrument touched the page. The test is what the instrument and the person made possible together that neither could have produced alone.
Part: The Signal
Chapter 18: Read What You Cannot Put Down
#The page is still on the desk.
#We can now ask what it contains.
#Does it introduce a new object, relation, method, intervention, map, result, or question?
#What is the nearest thing that existed before it?
#Does the claimed difference survive comparison?
#Does the reasoning hold?
#Do the sources say what the page says they say?
#Can the calculation be repeated?
#Does the modeled world remain distinguished from the observed one?
#What changes if the proposition is true?
#Who accepts responsibility for the answer?
#These questions are slower than a detector score because they examine the thing people were trying to protect.
#The abundance of machine language makes them more important, not less. When prose becomes cheap, relation becomes valuable. When fluency becomes common, reasoning becomes visible by contrast. When ten thousand variations can be generated in a minute, the recognition of which variation changes the map becomes an act of authorship.
#The signal is not that the page resembles no machine.
#The signal is that the page contains something worth finding.
A research culture with its signal intact
#A mature research culture does not require human beings to imitate error so a machine will recognize them as human.
#It records assistance where assistance matters. It uses watermarks where a retained generation signal serves a defined purpose. It uses classifiers and fingerprints as bounded forensic instruments. It uses provenance to preserve history. It uses source comparison to establish what was known. It uses reasoning and evidence to decide what survives. It uses responsibility to name the author.
#The culture can then receive a new form of inquiry without pretending it is an old form performed more quickly.
#AI-assisted research content can emerge from a session whose destination was known by neither participant. The machine expands traversal. The human brings the question into a life where it matters, recognizes when the available answer has failed, preserves the new relation against familiar substitutes, and decides whether to carry it into the world.
#The process can fail. It can hallucinate, normalize, plagiarize, flatten, miscalculate, and overwhelm. Human research could already do each of these things. The answer is not to abandon judgment to another probabilistic surface. The answer is to make the inquiry more observable and the judgment more exact.
#Generated abundance will force this change even where no institution wants to make it. A system receiving more submissions than human beings can read will use machines to read them. A search engine facing a million fluent answers will use machines to decide which answers appear. The question is not whether automated reading will occur. It is what the reader will be trained to find.
#If the system rewards resemblance to a trusted center, it will select the most familiar answer with the cleanest surface. It will call the repeated thing reliable because repetition made it statistically strong. Material outside the center will look like error before its relation can be examined. The filter will become very good at recovering what the field already knows.
#If the system rewards difference alone, random strings win. Noise is easy to make novel.
#The missing instrument must therefore follow a novelty curve rather than a rarity count. At one end lies repetition: highly available language carrying no new relation. At the other lies randomness: unprecedented language carrying no recoverable relation. Between them is the difficult territory where a new object remains coherent long enough to be compared, reasoned through, tested, remembered, and used.
#That instrument would not certify truth. It would preserve the candidate long enough for truth, usefulness, and reader attention to become possible. It would ask what the nearest prompt could easily generate, how many substantive corrections and redirections the artifact required, whether the central relation survives paraphrase, whether prior work contains it, and whether the work makes consequences visible that were not visible before.
#A one-line prompt can generate the number of words found in a book. Volume is not depth. The content that could have been generated by almost any nearby prompt is usually not bookworthy merely because it has acquired chapters and a cover. A work becomes more interesting when, after reading it, the reader cannot easily imagine the prompt that would have produced it. The distance between the first request and the finished world is part of the evidence.
#This is not a new purity test. A short prompt can release a brilliant relation, and a year of labor can produce nothing. Generation difficulty is one coordinate, not the verdict. It becomes useful when joined to novelty, reasoning, evidence, consequence, and the sustained attention of readers.
#Nested Causal Modelling projects a further state. A system trained only to continue the statistical center becomes more capable by learning the center. A system able to recognize unfamiliar coherence would become capable in a different direction. It would need to distinguish the accidental novel string from the relation that opens a new field; preserve the latter against normalization; and help test it without replacing it with the nearest known answer. An artificial general intelligence worthy of the name may depend less upon winning the imitation of average language than upon learning how novelty survives contact with reality.
#Someone will use AI to write something that neither that person nor the machine could have written alone. The work will exceed earlier work, not because the machine displaced humanity, but because the human enclosure gained a new instrument. Computer-aided drafting did not end design when it exceeded the precision of the drafting table. It expanded what people could build. AI-assisted authorship can do the same for worlds made of language.
#The similarity enclosure asked the page to prove that no machine had touched it.
#The research enclosure asks what happened, what appeared, what survives, and who will answer for it.
#The first protects an image of authorship by filtering language.
#The second protects authorship by finding the signal.
#At the Final Frontier of this volume, no claim has been made that every AI-assisted artifact is novel, true, or valuable. Detection still has lawful uses inside its evidence. Disclosure still belongs wherever a defined relationship requires it. The future system projected here has not been built, and the greatness of a future work cannot be established in advance.
#One boundary has been restored.
#Statistical similarity may be evidence about a bounded generation class. It is not a lawful substitute for provenance, authorship, novelty, reasoning, evidence, or value.
#In the end, value has to reach the reader with every word read. The reader has to entertain the author’s thoughts and allow them to build a world. The author has to earn the next sentence, then the next page, without the reader knowing in advance whether the journey will be worth taking.
#The reader must also be prepared to close the book when it no longer has their attention.
#Artificial intelligence can massively enhance a good author’s work. It cannot make bad writing good by itself. Used iteratively, it can help the author learn, research, revise, and attempt what neither could complete alone. The reader will still have to decide, not by the presence of AI, but by the value they find.
#Excluding AI cannot repair the reader’s core vulnerability. It can only create a false promise that the vulnerability has been removed. An author and the work are always unknown when the cover opens. The only honest protection is the reader’s attention, the author’s responsibility, and the value that begins to appear on the page where we begin.
Appendix A: Statement on Artificial Intelligence Use
#My work is substantially AI-assisted and human-authored.
#I do not conceal the assistance, and I do not outsource responsibility to it. AI is part of my apparatus. It is not my source of authority.
#I use artificial intelligence as a research instrument, writing collaborator, software agent, simulator, critic, production system, and object of study. The extent varies by project. AI assistance may include searching and comparing sources, exploring an idea, testing an argument, drafting and revising prose, writing and debugging code, analyzing data, generating tables and diagrams, generating or editing visual assets, simulating readers or debates, formatting publications, maintaining websites, and preparing files for release. Some published passages began as machine output and entered the work only after I selected, directed, revised, or accepted them.
#This is more than proofreading. It is not a transfer of authorship.
Discovery Cannot Be Commissioned
#The common description that I “asked AI to write” a work is often chronologically impossible.
#In many of these projects, the finished idea was known by neither the AI nor me before the AI-assisted research and writing session began. I could not have placed the conclusion in a prompt because I did not yet possess it. The model could not retrieve the work because the work did not exist. A request cannot specify an unknown destination and then serve as evidence that the destination was known all along.
#What existed at the beginning was an observation, a question, an anomaly, a direction of interest, or sometimes only the recognition that the available answer was wrong. Through sustained exchange, source research, drafting, comparison, calculation, simulation, contradiction, failure, and correction, a relation became visible that neither participant had stated at the start. Only then could it be named, developed, tested, and written as the apparent subject of the work.
#It is therefore inaccurate to reconstruct the process from the final book and imagine that I commissioned its contents from a machine that already knew them. The session was part of the environment in which the discovery occurred. AI supplied traversal, variation, transformation, and computational reach. I supplied the originating inquiry, recognized the signal when it appeared, distinguished it from nearby substitutes, decided whether it survived, and accepted responsibility for carrying it into the world.
#This is not ordinary dictation, ghostwriting, or automated completion. It is a new form of AI-assisted inquiry in which the writing can participate in the discovery of what the writing is about.
Novelty Is the Signal
#The relevant test is not whether a sentence resembles prose that someone associates with artificial intelligence. Similarity of tone, syntax, structure, or vocabulary cannot establish the origin of an idea. Human beings can write in styles associated with machines, machines can imitate styles associated with human beings, and editing can move the same passage in either direction. Stylometric suspicion is not a substitute for reading the work.
#The stronger signal is novelty.
#This body of work contains a high volume of novel ideas and unique content: new relations among existing observations, new conceptual objects, new maps, new methods, new formal structures, new interventions, new software, and new operational atlases. Those objects persist across books, source files, repositories, revisions, research deposits, and public DOI records. The DOI timestamps establish the public boundary at which the published ideas entered the durable record. Later corrections document their development rather than erasing their origin.
#The presence of AI assistance does not create a deficit of novelty. In this work, novelty is abundant and inspectable. A serious evaluation should ask what was introduced, whether it had appeared before, how it developed across the record, whether the originating relations were preserved, and what the result makes possible. It should not replace those questions with an aesthetic judgment that the prose sounds like a machine.
#AI use is part of the provenance. The published novelty is the signal.
Authorship and Responsibility
#I supply the inquiry, the originating observations, the direction of the work, the recognition of what matters, and the decision to accept, reject, correct, or publish what the systems produce. I am responsible for the resulting work, including its claims, errors, omissions, interpretations, and revisions.
#AI systems do not experience the human circumstances that make a question matter. They do not bear the consequences of publication. They can generate a strong paragraph that moves in the wrong direction, preserve every important term while replacing the relation that made the idea novel, or produce a plausible citation that does not support the sentence beside it. Fluency is a capacity, not final authority.
#The author is the person responsible for the whole.
Fiction
#Fiction requires a narrower creative boundary. AI may support spelling, grammar, research, continuity checking, and the examination of alternatives, but the writer must have written the story. The unique expressive creation of the author is the authored object and must be protected as such. Similarity detection cannot establish that creative history, and a watermark cannot decide whether the story is good.
Evidence and Correction
#AI output is not evidence merely because it is fluent.
#Evidence comes from the source, observation, measurement, calculation, data, or reproducible operation being cited. Where a result comes from a model or simulation, the assumptions should be identified and the modeled output distinguished from an observed quantity.
#Long-running AI collaboration also produces a recurring failure I call prior capture. A system may retain the words of an original proposal while translating its structure into a more familiar idea. Original prompts, source-fidelity ledgers, canonical definitions, version histories, research deposits, and quarantined drafts make that substitution easier to find.
#I publish corrections when they are necessary. A visible revision history is not evidence that a work lacked authorship. It is evidence that the author remained responsible after publication.
Systems Used
#My work has involved multiple artificial-intelligence systems, including OpenAI’s ChatGPT and Codex, xAI’s Grok, and Google’s Gemini and NotebookLM. Different systems have been used for research, drafting, simulation, software development, criticism, and production. The presence of a system in the process does not imply that it endorses a claim, and a provider is not responsible for my use of its output.
#Artificial intelligence accelerates traversal.
#It does not choose the destination or carry responsibility for arriving there.
#That responsibility is mine.
Appendix B: The Nested Causal Model
Reference Before Departure
#The observed thing is a research artifact whose language may have passed through both human and artificial-intelligence processes.
#No misconduct, authorship deficit, novelty, or value is assumed before the map is built.
#The original inquiry sets the active range: whether resemblance to machine output can establish authorship when humans and machines have been made to imitate one another, and whether that resemblance should displace novelty and reasoned ideation as the signal of value.
#The outer Final Frontier is human consciousness, where language becomes meaningful and answerable. The inner Final Frontier is the artifact surface, where classifiers, watermarks, and fingerprints measure bounded features. The model does not calculate more distant enclosures unless they become active in this inquiry.
Level 1: Enclosure Map
The generation event
[human-machine research and publishing system
| AI-assisted content-generation event
| prompts, retrieved sources, model inference, tool calls,
candidate outputs, human selections, edits and rejections]
#The generation event is enclosed by the research objective, human experience, model training, system instructions, interface, context, tool permissions, source availability, provider policy, compute, time, publication norms, and the material circumstances that make the inquiry matter.
#The event encloses smaller events: prompts, completions, searches, citations, code executions, revisions, comparisons, and decisions.
The resulting artifact
[research, publication and evaluation enclosure
| AI-assisted research artifact
| propositions, evidence, reasoning, prose, data, code,
figures, citations, provenance records and revision history]
#The artifact is not the generation event. It is a durable state produced by that event and subsequent editing. Its surface may not retain enough information to reconstruct every event that produced it.
Parallel enclosures
- human author;
- AI system and provider;
- sources and prior literature;
- publisher or repository;
- detector vendor;
- evaluator, reader, teacher, editor, or employer;
- DOI and version-control systems;
- future language corpus.
#These exchange language and evidence without becoming one another.
The human enclosure of language
[human consciousness | linguistic capacity | parallel human languages]
#The languages are parallel expressions inside a human linguistic capacity. The machine capable of reading and generating their forms is enclosed within the human linguistic and technical enclosure. It can introduce a third production process, but it does not introduce a third language standing beside human language.
Level 2: Causal Flow
Outer-to-inner nonvolatile projection
#The outer enclosure determines which models are available, what they were trained to continue, which tools they can use, what a publisher permits, what a detector rewards or punishes, and what a human believes publishable.
#This top-down causality is reversible and subject to human modification. The enclosing policy, permission, boundary, or instruction can be withdrawn, changed, or replaced, redirecting what it projects inward. Its designation as nonvolatile does not mean low risk. It names its outer-to-inner direction and its reversible relation to entropy.
#Detector policy projects inward before detection. A writer who knows that clear, low-perplexity prose may be classified as AI can add errors, avoid ordinary transitions, reject useful assistance, conceal process records, or alter a lawful voice to satisfy the detector. The policy changes the artifact it later claims merely to observe.
Bottom-up volatile causality
#Tokens are generated. Sources are opened. Candidate sentences appear. The human selects, rejects, rewrites, combines, and redirects. Code executes. A calculation returns. A proposition becomes stable enough to enter a draft. The draft becomes a public artifact. The artifact enters a corpus and becomes something future humans and models may read.
#This bottom-up causality is irreversible. A token may later be deleted, a draft revised, and a publication corrected, but the completed event is not reversed. A subsequent event changes the state and adds to its causal history. Its designation as volatile does not mean dangerous. It names its bottom-up direction and its irreversible relation to entropy.
Nested exchange
#The model was trained to continue human language. The human reads machine language trained upon human language. The human’s later prose changes through that exposure. Future models encounter human prose already altered by machine assistance. The exchange is bidirectional and cumulative.
#The two distributions a detector wishes to separate are therefore not fixed natural species. They are interacting states inside one language enclosure.
#There can be a machine-generated sequence, a marked generation event, and a mixed human-machine production process. These are provenance statements. They do not establish a separate AI authorship species. Machine output becomes socially operative language through human prompting, interpretation, publication, rejection, acceptance, and action.
The inverse inference
#Detection begins with the final artifact and attempts to infer the generation event backward:
artifact surface -> statistical class -> alleged process -> alleged author
#Each arrow changes the object.
#A classifier may estimate similarity to represented machine output. That does not establish the process. A retained watermark may identify a supported generation event. That does not establish the complete process. Process does not by itself establish authorship. Authorship does not establish novelty. Novelty does not establish validity. Validity does not establish usefulness.
#The current system repeatedly treats this entire chain as one inference.
Level 3: Enclosing-State Comparison
State A: Similarity-enclosed evaluation
#The active question is: Does this text resemble AI-generated text?
#The artifact is reduced to lexical, syntactic, perplexity, curvature, distributional, or watermark features. A score becomes suspicion. Suspicion becomes an authorship conclusion. The authorship conclusion becomes a quality judgment or sanction.
#Observable consequences:
- false positives attach machine origin to human work;
- false negatives reward superficial evasion;
- non-native and constrained writers carry unequal risk;
- mixed authorship is forced into a false binary;
- clear prose is made suspicious;
- deliberate errors become evidence of humanity;
- provenance is guessed where it could have been recorded;
- novelty and reasoning receive less attention than stylistic conformity;
- the filter removes signal while reporting less noise.
- low-value content is converted from a quality judgment into a provenance stereotype;
- an AI label becomes a reason not to read the proposition;
- disliked or unfamiliar writers can be excluded through a source prejudice presented as technical measurement;
- markets are prevented from filtering reader preference because the artifact is intercepted before readers can encounter it.
State B: Provenance, novelty and reasoned-ideation evaluation
#The active questions remain separate:
- What process produced this artifact?
- What new relation, object, method, or result does it introduce?
- Does its reasoning preserve the relation it claims?
- What evidence supports it?
- What does it make possible?
- Who accepts responsibility for the whole?
#Observable consequences:
- AI use can be disclosed without becoming a quality verdict;
- provenance records answer process questions directly where available;
- detector output remains a bounded forensic lead;
- novelty is compared with prior work rather than prose stereotypes;
- reasoning and evidence remain available for criticism;
- DOI and repository timestamps preserve public sequence;
- corrections develop the record instead of erasing authorship;
- known work is recognized as known;
- new work is allowed to be new before it is judged true or false.
- bad ideas remain available for criticism even when expressed eloquently;
- reader and market preference operate after access rather than as a prior authorship accusation.
Active enclosure
#The active enclosure sustaining the failure is the institutional demand to classify provenance from semantic surface before evaluating contribution.
#The detector is not the whole problem. It is the instrument through which the enclosing demand acts.
Level 4: The Signal Protocol
#The intervention arises from the state comparison. It does not prohibit detection. It restores each question to the enclosure capable of answering it.
1. Declare the evaluation purpose
#Before examining the artifact, state whether the inquiry concerns provenance, academic process, novelty, factual validity, reasoning, safety, or value. Do not allow one score to answer a different question.
2. Prefer recorded provenance to inferred similarity
#Accept author disclosures, model-use statements, revision history, prompt and output deposits where appropriate, signed content credentials, repository history, and DOI timestamps. Preserve privacy and unpublished inquiry by requiring only the detail material to the evaluation.
3. Bound detector output
#A detector may identify compatibility with its represented class. It may not, without independent evidence, determine misconduct, human absence, authorship, novelty, validity, or value. Watermark presence identifies a retained mark; watermark absence does not certify human-only production.
4. Construct a novelty ledger
#For each claimed contribution, record:
ID | proposition or object | nearest prior analogue | preserved difference
| earliest durable record | dependencies | test or consequence
#Novelty is relational. Familiar words can carry a new relation. Unfamiliar words can carry none.
5. Conduct a reasoned-ideation audit
#Evaluate the chain from observation to conclusion:
- Is the reference state declared?
- Are terms stable?
- Does the conclusion follow from the stated relation?
- Are sources accurately represented?
- Are calculation and simulation reproducible?
- Is modeled output distinguished from observation?
- Is the Final Frontier of the claim declared?
- What would correct or falsify the result?
6. Preserve multidimensional judgment
#Represent evaluation as a vector rather than one authenticity score:
Q(A) = [provenance, novelty, reasoning, evidence, reproducibility,
utility, responsibility]
#One dimension cannot silently stand in for the others.
7. Permit correction and appeal
#An adverse provenance inference must identify its evidence, model, domain, threshold, error rate, and decision rule. The author may answer with process records. A corrected detector judgment must not remain attached as the artifact’s identity.
Target State
#AI assistance is visible as provenance. Novelty is visible as contribution. Reasoning and evidence are visible as the basis for judgment. Authorship is visible as responsibility.
#The noise filter no longer destroys the signal it was built to protect.
#If Shakespeare repeats an existing plot, tell Shakespeare that someone already did that one.
#Do not tell Shakespeare he did not write it.
Final Frontier
#The model does not establish that every AI-assisted work is novel, valid, or valuable. It does not establish that detection has no lawful use. It does not make disclosure optional where a defined process requires it.
#It establishes the boundary that current evaluation repeatedly crosses: statistical similarity can be evidence about a bounded generation class. It is not a lawful substitute for provenance, authorship, novelty, reasoning, evidence, or value.
Appendix C: Novelty Ledger
#The entries below apply the Signal Protocol to this volume. Each contribution is kept separate even where several appear in one passage. The originating prompts were recorded uncorrected beginning August 31, 2026, and continued through this revision on September 4. They are reproduced in Appendix E.
N-001: AI-assisted research content
#Proposition or object: AI-assisted research content names work produced through an inquiry in which artificial intelligence participates in research, comparison, criticism, drafting, computation, and revision without possessing the finished destination at the beginning.
#Nearest prior analogue: Research on human-AI coauthorship and hybrid text.
#Preserved difference: Token mixture describes the artifact after generation. This object includes discovery during the collaboration: neither participant could request or retrieve the finished proposition before the inquiry produced it.
#Earliest durable record: Source-fidelity Prompt 2, August 31, 2026; expanded in Prompt 4. Repository record 0bd2605, September 4, 2026.
#Dependencies: Chronological prompt and revision records; a distinction between generation events, artifacts, provenance, and authorship.
#Test or consequence: Reconstruct the initial state from the durable record. If the completed contribution already appears in the opening request or a retrievable source, the discovery claim narrows accordingly.
N-002: The impossible commission
#Proposition or object: An author cannot commission an AI to write a discovery that neither the author nor the AI-assisted session yet knows. The later discovery cannot be copied backward into the first prompt without violating chronology.
#Nearest prior analogue: Prompting accounts in which a user specifies a destination and a model generates the requested object.
#Preserved difference: The destination emerges through the recorded research process and becomes requestable only after it has been discovered.
#Earliest durable record: Source-fidelity Prompt 2, August 31, 2026.
#Dependencies: A durable sequence of prompts, outputs, rejected analogues, sources, revisions, and decisions.
#Test or consequence: Locate the proposition before its claimed emergence. If it was already available to either participant in the operative record, the impossibility asserted here does not apply.
N-003: Resemblance cannot determine authorship
#Proposition or object: A detector can measure compatibility with a represented class of text, but the promotion from surface resemblance to generation process, from process to authorship, and from authorship to value occurs in the enclosing institution rather than in the detector’s observation.
#Nearest prior analogue: Statistical AI-text detection, watermark detection, model fingerprinting, and signed content provenance.
#Preserved difference: The methods are not rejected. Their bounded outputs are separated from the different objects they are routinely asked to decide.
#Earliest durable record: Source-fidelity Prompts 3 and 4, August 31, 2026.
#Dependencies: Published detector limitations, hybrid-text research, watermark scope, provenance standards, and the generation-event/artifact distinction.
#Test or consequence: Require a detector result, without process evidence, to distinguish edited, mixed, translated, and imitated text while recovering the human responsibility for the whole. Failure bounds the result to resemblance rather than authorship.
N-004: One language enclosure, interacting production processes
#Proposition or object: Human-only, machine-mediated, and mixed generation are distinguishable production histories inside a human enclosure of language. They do not become separate linguistic species merely because the immediate producer differs.
#Nearest prior analogue: Accounts of language models trained on human text, human-AI hybrid writing, and evidence that AI-assisted writing is reshaping human communication.
#Preserved difference: These findings are joined into a reciprocal causal loop. Models imitate language inherited from people; people consume and adapt to machine language derived from that inheritance; later human and model language carry both events forward.
#Earliest durable record: Source-fidelity Prompt 4, August 31, 2026; enclosure stated explicitly in Prompt 7.
#Dependencies: Human linguistic capacity, cultural transmission, model training on human corpora, and measurable adaptation to AI-mediated prose.
#Test or consequence: A stable authorship classifier requires separable, stable source distributions or an independent retained provenance signal. As the distributions interact, surface classification cannot recover the absent causal history.
N-005: The novelty ledger
#Proposition or object: Claimed novelty should be recorded as a relation among a proposition, its nearest prior analogue, the exact preserved difference, its earliest durable record, its dependencies, and a test or consequence.
#Nearest prior analogue: Bibliometric measures of unusual concepts, combinations, research questions, and methods; invention-disclosure and prior-art records.
#Preserved difference: The ledger protects relational novelty during an active AI-assisted inquiry. It is not a style score, an award of novelty by declaration, or merely a search for unusual words.
#Earliest durable record: Source-fidelity Prompt 3, August 31, 2026; formal ledger structure first preserved in repository record 0bd2605, September 4, 2026.
#Dependencies: Searchable prior work, source-fidelity records, publication chronology, and a willingness to narrow a claim when its analogue already contains the relation.
#Test or consequence: A cited prior work defeats or narrows an entry when it contains the preserved difference, not merely familiar components or language.
N-006: The Signal Protocol
#Proposition or object: Evaluation should first declare its purpose, then route provenance, novelty, reasoning, evidence, reproducibility, utility, and accepted responsibility to evidence capable of answering each question.
#Nearest prior analogue: Synthetic-content risk frameworks, content provenance standards, research-integrity review, and detector-use policies.
#Preserved difference: The intervention is organized around preventing one coordinate from silently substituting for the others. Detection remains available as a bounded instrument rather than becoming a complete judgment.
#Earliest durable record: Source-fidelity Prompts 3 and 4, August 31, 2026; Level 4 protocol preserved in repository record 0bd2605, September 4, 2026.
#Dependencies: The multidimensional evaluation vector, direct provenance where available, reasoned-ideation audit, correction, and appeal.
#Test or consequence: Apply the protocol to the same artifacts evaluated by a similarity-first process. Compare false accusations, missed misconduct, novel relations recovered, evidentiary corrections, and evaluator agreement.
N-007: Watermarking introduces the pattern it seeks
#Proposition or object: A text watermark intentionally changes generation so a machine can later recognize its statistical trace. It treats language as a detectable commodity before a reader has determined whether the language matters.
#Nearest prior analogue: Token-probability watermarking, including the green-list method and SynthID.
#Preserved difference: The technical mechanism is observed as a top-down projection into prose: the attempt to identify machine writing deliberately introduces a machine-selected pattern into writing.
#Earliest durable record: Source-fidelity Prompt 5, August 31, 2026.
#Dependencies: A watermark that alters token probabilities, retention of the mark through editing, and an institution that treats its detection as material.
#Test or consequence: Compare marked and unmarked outputs for detectability, survival under revision, linguistic effect, false attribution, and any gain in the separate judgments of novelty, validity, or value.
N-008: The authorship boundary differs for fiction
#Proposition or object: AI may assist research and prose without assuming responsibility for the work, while fiction retains an additional boundary: the writer must author the story’s unique expressive creation, and the reader must still decide whether it is good.
#Nearest prior analogue: The human-authorship requirement in copyright and the distinction between assistive use and machine-determined expressive material.
#Preserved difference: Legal authorship, creative responsibility, and reader value remain separate. A detector designed to imitate and recognize human form cannot prospect literary merit or replace the reader.
#Earliest durable record: Source-fidelity Prompt 5, August 31, 2026.
#Dependencies: Human expressive control, disclosure appropriate to the publication enclosure, and the reader’s freedom to continue or close the book.
#Test or consequence: Establish which expressive decisions the writer made and controlled. Evaluate literary value by reading the work, not by promoting the presence or absence of AI assistance into the verdict.
N-009: The novelty curve
#Proposition or object: A system searching for signal must distinguish repetition from randomness and preserve the intermediate territory where an unfamiliar relation remains coherent enough to compare, test, remember, and use.
#Nearest prior analogue: Statistical outlier detection, bibliometric novelty measurement, and next-token probability.
#Preserved difference: Maximum rarity is not maximum novelty. The proposed curve joins statistical difference to relational coherence and consequence so random strings do not defeat the measure.
#Earliest durable record: Source-fidelity Prompt 9, September 4, 2026.
#Dependencies: A represented prior field, relational comparison, reasoning, evidence, memory, and observable consequence.
#Test or consequence: Compare repeated prose, randomized prose, unfamiliar but coherent work, and established high-value work. A useful measure must separate rarity from sustained novel relation without converting popularity into truth.
N-010: Generation distance
#Proposition or object: The difficulty of reconstructing a prompt capable of producing a finished work is evidence about the distance between a simple commission and an iterative inquiry, although it is not a verdict on quality.
#Nearest prior analogue: Prompt complexity, edit distance, process logs, and measures of generation effort.
#Preserved difference: The object measured is not token volume or textual distance. It is the substantive sequence of discoveries, rejections, corrections, and steering required to make the final relation available.
#Earliest durable record: Source-fidelity Prompt 9, September 4, 2026.
#Dependencies: Preserved prompts and revisions, identification of the work’s central relations, and separation of effort from value.
#Test or consequence: Attempt to reproduce the material contribution from nearby single prompts. Easy reproduction narrows the process claim; failed reproduction identifies which intermediate decisions carried the work.
N-011: Computer-aided authorship
#Proposition or object: AI-assisted authorship can relate to prose as computer-aided drafting relates to design: the instrument expands precision, iteration, comparison, and reachable complexity without becoming the person responsible for the whole.
#Nearest prior analogue: Computer-aided drafting and established accounts of AI as an assistive tool.
#Preserved difference: The analogy is applied to iterative intellectual and creative development rather than to token production. The author can learn through the repeated use of the instrument and make later decisions from a state unavailable at the beginning.
#Earliest durable record: Source-fidelity Prompt 9, September 4, 2026.
#Dependencies: Human direction, repeated substantive revision, source verification, and retained responsibility.
#Test or consequence: Compare the work with outputs produced by the author alone and the model alone under bounded conditions. The claim strengthens only where the collaboration produces a material relation neither reaches alone.
N-012: The reader remains the value boundary
#Proposition or object: Generated abundance does not remove the reader’s ancient vulnerability to an unknown author and unknown work. Value must be earned through attention after access, and the reader must remain free to close the book.
#Nearest prior analogue: Literary markets, criticism, source-label effects, and attention-based recommendation.
#Preserved difference: Excluding AI cannot solve the uncertainty inherent in opening any work. A provenance filter applied before reading can instead prevent the only encounter in which literary value becomes available.
#Earliest durable record: Source-fidelity Prompt 8, September 4, 2026; developed in Prompt 9.
#Dependencies: Access to the work, reader attention, author responsibility, and separation of provenance from value.
#Test or consequence: Compare judgments made before and after disclosure, and before and after reading. If provenance alone predicts durable reader value across works, the claimed separation narrows.
#The ledger does not award novelty by declaration. It makes the comparison specific enough to correct.
Appendix D: Research Evidence Ledger
Controlling Question
#Does enclosing research content by its statistical similarity to AI output make more sense than verifying its novelty, reasoning, evidence, provenance, and use?
Distinctions Protected Before Analysis
- Generation is an event or process.
- Artifact is the content left by that process.
- Provenance is the recorded history of participants, tools, and changes.
- Authorship is responsibility for the work as a whole; it is not reducible to the source of every token.
- Detection estimates compatibility with a class represented in a detector.
- Watermark detection tests for a signal deliberately inserted during a supported generation process.
- Fingerprinting associates surface or probability features with a known model, model family, or distribution under bounded conditions.
- Novelty concerns what the artifact introduces relative to prior work.
- Validity concerns whether its reasoning and evidence support its claims.
- Value cannot be inferred from provenance alone.
Primary and Authoritative Sources
1. OpenAI’s retired text classifier
#Source: OpenAI, “New AI classifier for indicating AI-written text,” January 31, 2023; updated July 20, 2023.
#URL: https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/
#Findings used:
- The classifier was withdrawn because of low accuracy.
- On OpenAI’s challenge set it identified 26% of AI-written text as likely AI and mislabeled 9% of human text.
- OpenAI said it should not be a primary decision-making tool.
- Short, non-English, code, predictable, edited, and out-of-distribution text were identified as important limitations.
#Use in Volume XII: establishes that even the provider that built a detector did not treat its output as authorship adjudication.
2. DetectGPT
#Source: Mitchell et al., “DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature,” ICML 2023.
#DOI/URL: https://proceedings.mlr.press/v202/mitchell23a.html
#Finding used: a detector can exploit bounded features of a known model’s probability landscape and perform well on specified generated samples.
#Use: presents the strongest form of statistical fingerprinting fairly. Its target is sample detection relative to a model distribution, not recovery of human responsibility for a mixed artifact.
3. RAID benchmark
#Source: Dugan et al., “RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors,” ACL 2024.
#DOI: https://doi.org/10.18653/v1/2024.acl-long.674
#Findings used:
- RAID contains more than six million generations across eleven models, eight domains, eleven attacks, and four decoding strategies.
- Twelve detectors were readily affected by attacks, sampling strategies, repetition penalties, and unseen generators.
#Use: demonstrates domain and generator dependence at a scale larger than single-model demonstrations.
4. Detection impossibility and attacks
#Source: Sadasivan et al., “Can AI-Generated Text be Reliably Detected?” arXiv:2303.11156, 2023.
#URL: https://arxiv.org/abs/2303.11156
#Findings used:
- Recursive paraphrasing reduced detection while often preserving quality.
- Watermarking, neural classifiers, zero-shot, and retrieval methods were stress-tested.
- As human and machine text distributions approach one another, a detector faces a fundamental false-positive/false-negative tradeoff.
- Spoofing can make non-AI text appear watermarked or AI-associated.
#Use: formal support for the inverse-problem and convergence argument.
5. Bias against non-native English writers
#Source: Liang et al., “GPT detectors are biased against non-native English writers,” Patterns 4(7), 2023.
#DOI: https://doi.org/10.1016/j.patter.2023.100779
#Findings used:
- Evaluated detectors frequently misclassified non-native English writing.
- Simple prompting could evade the same detectors.
- The authors cautioned against evaluative and educational use.
#Use: shows that the detector’s noise class can contain lawful human variation and that stylistic conformity is not machine authorship.
6. Human-AI hybrid text
#Source: Zhang et al., “LLM-as-a-Coauthor: Can Mixed Human-Written and Machine-Generated Text Be Detected?” Findings of NAACL 2024.
#URL: https://aclanthology.org/2024.findings-naacl.29/
#Source: Zhang et al., “Detecting AI-Generated Sentences in Human-AI Collaborative Hybrid Texts: Challenges, Strategies, and Insights,” IJCAI 2024.
#URL: https://www.ijcai.org/proceedings/2024/0835.pdf
#Findings used:
- Pure-machine detection does not represent AI-revised human text or human-revised machine text.
- Human selection and editing creates a third collaborative class.
- Frequent authorship changes and short segments limit attribution even when segmentation is known.
#Use: direct evidence that token-level origin does not recover authorship of a mixed work.
7. Watermarking
#Source: Kirchenbauer et al., “A Watermark for Large Language Models,” ICML 2023.
#URL: https://proceedings.mlr.press/v202/kirchenbauer23a.html
#Finding used: model sampling can be biased toward a secret partition of tokens to embed a statistically detectable signal.
#Source: Google DeepMind, SynthID documentation and 2024 text announcement.
#URLs:
- https://deepmind.google/models/synthid/
- https://deepmind.google/blog/watermarking-ai-generated-text-and-video-with-synthid/
#Findings used:
- SynthID changes token probabilities during generation.
- It works best on longer and more varied generations.
- Google explicitly describes it as a building block rather than a complete solution.
#Use: watermark presence may establish compatibility with a marked generation event. It does not evaluate novelty, truth, usefulness, or responsibility for the final edited artifact. Absence of a mark does not establish human-only production.
8. Provenance standards
#Source: NIST AI 100-4, Reducing Risks Posed by Synthetic Content, 2024.
#DOI: https://doi.org/10.6028/NIST.AI.100-4
#Source: Coalition for Content Provenance and Authenticity, C2PA Specifications.
#URL: https://spec.c2pa.org/
#Findings used:
- Provenance, watermarking, labeling, and detection are distinct technical approaches.
- C2PA supplies signed assertions about content source and history.
#Use: supports recording process directly rather than guessing it exclusively from semantic surface. A provenance assertion is still not a truth or novelty certificate.
9. Human language adapts to AI-mediated writing
#Source: Sourati et al., “The Widespread Adoption of Large Language Model-Assisted Writing Is Reshaping Human Communication,” Nature Human Behaviour, August 24, 2026.
#DOI: https://doi.org/10.1038/s41562-026-02550-0
#Preprint: https://arxiv.org/abs/2502.11266
#Findings used:
- Across multiple studies, domains, and more than 880,000 texts, widespread LLM writing assistance was associated with declining linguistic diversity.
- AI polishing retained core content while homogenizing style and changing linguistic markers.
#Use: directly supports the author’s observation that humans and machines are not stationary, separated language distributions. People consume and adapt to machine-mediated language while models are trained and prompted to imitate human language.
10. Novelty measurement
#Source: Mishra and Torvik, “Quantifying Conceptual Novelty in the Biomedical Literature,” 2016.
#URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC5142764/
#Finding used: novelty can arise from uncommon relations or combinations among concepts even when the individual concepts are established.
#Source: Luo et al., “Combination of research questions and methods: A new measurement of scientific novelty,” 2022.
#URL: https://doi.org/10.1016/j.joi.2022.101282
#Finding used: research questions, methods, and their combinations can be separately compared through time and literature.
#Use: supports a relational novelty ledger rather than lexical unfamiliarity.
11. Human authorship and assistive AI in copyright
#Source: U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability, January 2025.
#URL: https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf
#Findings used:
- Copyright protection in the United States requires human authorship.
- AI used as an assistive tool does not by itself prevent copyrightability.
- Protection depends upon sufficient human expressive contribution and is assessed case by case.
- Mere prompting does not by itself supply the required human control over expressive elements.
#Use: supports the volume’s boundary between AI-assisted research content and fiction. The fiction writer may use tools around the work, but the unique expressive creation must remain human-authored. This source establishes a legal boundary, not whether a story is good or how its cultural value should be measured.
Evidence Boundary
#The sources establish bounded limitations, not a universal proof that no forensic method can ever identify any model output. Watermarks can be detected when embedded and retained. Classifiers can perform well on represented models, domains, and decoding conditions. Process logs and signed provenance can identify tool participation.
#The stronger conclusion is narrower and sufficient: surface similarity does not determine authorship, novelty, validity, or value in human-AI research artifacts, and detector outputs cannot lawfully substitute for those separate questions.
12. Language as biological, individual, and cultural inheritance
#Source: Simon Kirby, “Culture and Biology in the Origins of Linguistic Structure,” Psychonomic Bulletin & Review 24 (2017).
#DOI: https://doi.org/10.3758/s13423-016-1166-7
#Finding used: linguistic structure arises through the interaction of a human biological capacity, individual learning, and cultural transmission.
#Use: supports placing parallel languages inside the human enclosure of linguistic capacity. The source does not establish the volume’s complete NCM; it supports the human and cultural inheritance on which that map depends.
13. Language models are trained downstream of human language
#Source: Brown et al., “Language Models Are Few-Shot Learners,” 2020.
#URL: https://arxiv.org/abs/2005.14165
#Findings used:
- GPT-3 was trained autoregressively on large human-produced text corpora.
- In the reported synthetic-news evaluation, human participants had difficulty distinguishing generated articles from human-written articles.
#Use: supports the causal order in which human language precedes model training and shows that fluent surface form does not preserve a reliable visible trace of its immediate production event.
14. Form and communicative meaning
#Source: Emily M. Bender and Alexander Koller, “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data,” ACL 2020.
#DOI: https://doi.org/10.18653/v1/2020.acl-main.463
#Finding used: the paper distinguishes linguistic form from communicative intent and argues that a system trained only on form cannot establish meaning from form alone.
#Use: supports identifying human communication as the enclosure in which a generated sequence becomes meaningful and answerable. The manuscript does not depend upon denying that models can develop useful semantic representations.
15. Human source heuristics are unreliable
#Source: Maurice Jakesch, Jeffrey T. Hancock, and Mor Naaman, “Human Heuristics for AI-Generated Language Are Flawed,” PNAS 120, no. 11 (2023).
#DOI: https://doi.org/10.1073/pnas.2208839120
#Findings used:
- Across six experiments involving 4,600 participants, human judgments in the tested domains did not reliably identify the source of generated language.
- Participants relied on cues that could be reversed, allowing generated text to be perceived as more human than human text.
#Use: supports the claim that source stereotypes cannot replace reading and that intuitive “AI-like” judgments are not authorship evidence.
16. Source labels can alter evaluation
#Source: Jing Cao and Tianli Zhou, “Human or Machine? Do Source Beliefs Shape Cognitive Bias in Post-Editing?” Frontiers in Psychology 17 (2026).
#DOI: https://doi.org/10.3389/fpsyg.2026.1819659
#Finding used: in a bounded study of 60 translation students, source mislabeling altered quality ratings and post-editing behavior.
#Use: corroborates the mechanism by which an AI label can become a source prejudice. This is not treated as a universal estimate outside the study’s population or task.
17. Computer-aided drafting as an assistive precedent
#Source: Autodesk, “Drafting Software: Create Technical Drawings with Precision.”
#URL: https://www.autodesk.com/solutions/aec/drafting-software-programs
#Findings used:
- Computer-aided drafting replaced many manual drafting operations while increasing speed, precision, revision capacity, reuse, and collaboration.
- The software remains an instrument used by architects, engineers, construction professionals, and manufacturers rather than becoming the responsible designer of the resulting object.
#Use: supports the book’s analogy between computer-aided drafting and AI-assisted authorship. The analogy concerns the expansion of iteration and reachable precision. It does not claim that prose has engineering tolerances or that a tool guarantees a successful design.
New Model Consequences
- Third process, not third language. Human-only generation, machine-mediated generation, and mixed production can be distinguished as process histories without constructing separate linguistic species.
- Quality and provenance remain orthogonal. Machine assistance can scale low-value content, but poor ideas, repetition, and empty eloquence predate it and cannot be diagnosed by source alone.
- The AI label can become an exclusion proxy. Detector bias, unreliable human heuristics, and source-label effects make it possible to dismiss a writer before evaluating the proposition.
- Markets filter preference, not truth. Reader demand can select what is bought, recommended, or sustained after access. It does not establish authorship, novelty, or validity, and cannot operate when a prior detector gate prevents the work from reaching readers.
- Probability inherits represented human regularity. A likely continuation is not an independent machine judgment about meaning. Its availability is downstream of repeated linguistic relations in human training material and the conditions imposed upon generation.
- Novelty requires a curve, not a rarity maximum. Repetition can contain no new relation and randomness can contain no recoverable one. The signal lies where unfamiliar structure remains coherent enough to compare, test, remember, and use.
Completed Volume XII Novelty Ledger
#The full ledger appears as Appendix C of The Signal. Its twelve entries apply the intervention to the volume itself:
- AI-assisted research content: research whose destination emerges during human-machine collaboration rather than being contained in the initial commission.
- The impossible commission: the finished discovery cannot be requested before either participant knows it without violating the durable chronology.
- Resemblance cannot determine authorship: detector output changes objects when an institution promotes surface similarity into process, authorship, and value.
- One language enclosure: human-only, machine-mediated, and mixed generation are interacting production histories rather than separate linguistic species.
- The novelty ledger: novelty is recorded as a preserved difference from the nearest prior analogue, joined to chronology, dependencies, and a test.
- The Signal Protocol: provenance, novelty, reasoning, evidence, reproducibility, utility, and responsibility are evaluated separately so one coordinate cannot silently decide the others.
- Watermarking introduces the pattern it seeks: changing token probabilities to identify machine generation projects a machine-selected trace into the language being classified.
- Fiction retains an additional authorship boundary: assistive use, expressive authorship, legal protection, and reader judgment remain different questions.
- The novelty curve: signal lies neither in maximum repetition nor maximum randomness, but where unfamiliar structure remains coherent and consequential.
- Generation distance: the substantive path from an initial prompt to a finished relation can distinguish a commission from an inquiry without becoming a quality score.
- Computer-aided authorship: iterative AI assistance can expand the author’s reachable precision and knowledge without assuming responsibility for the work.
- The reader remains the value boundary: provenance cannot eliminate the uncertainty of opening an unknown work or replace the attention through which literary value is encountered.
#The earliest durable source for entries 1 through 8 is the uncorrected prompt record beginning August 31, 2026. Entries 9 through 12 enter that record on September 4, 2026. The initial manuscript was preserved in repository record 0bd2605 and deposited under DOI 10.5281/zenodo.22311705. Each entry identifies its nearest analogue and the condition under which the claimed difference should be narrowed or withdrawn.
Appendix E: Originating Prompts: Uncorrected
#These prompts contain the author’s originating prose and the significant steering instructions that generated the volume. They are reproduced without correction.
#SOURCE-FIDELITY RECORD Recorded: 2026-08-31 Author: Sid J.A. Hubbard
#— PROMPT 1 —
#I need to make an AI Use Statement that I will post to my website.
#— PROMPT 2 —
#It needs to include the impossibility of asking an ai to write what neither knew anything about before the AI Assisted writing session.
#— PROMPT 3 —
#there is a high volume of novel ideas and unique content. That should be the test. Not the presence of similarity to ai the novelty is the signal and there is no deficit of novel ideas in the work published to doi timestamps
#— PROMPT 4 —
#lets make volume 12 about AI Assisted Research Content. Do a nested causal modelling of AI Content Generation and examine the methods of detection, watermarks and fingerprinting. Does the enclosure of ai similarity make more sense than the verification of novelty and reasoned ideations? The signal is what is being lost in the attempt to filter noise in this manner. You can not take a machine and ask it to read everything we have written and learn to write like us and then use that same machine to write like us and then ask it which content is human and which it wrote as the human language adapts to the presence of the ai generated content. People naturally will write like what they read. There is no way to determine authorship was machine or human as the human and machine have been instructed to write like eachother. One actively by command and one passively through consumption of generated text. What can and should be the measure of value is the novelty. Someone should be told that shakespeare already did that one, they should not tell shakespeare he did not write it.
#— PROMPT 5 —
#the ethical use of ai is in the support of grammar and spelling and the research process of evaluating new ideas, with regard to fiction it is different. The unique creations of an author must be protected. We have a society that objectifies prose and then commodifies it. The commoddity is not such that it can vbe authenticated from ai forgery. For an Author of fiction, there is another requirement. The writer must have written the story and they must be good. We have, up until now, depended upon markets to decide the latter and the courts do litigate when money has been made. This cannot be prospected or defended by a detection whose premise is a performance metric of its own ability to write like a human. Watermarks are an atrtempt to treat text as a commodity before anyone cares about it. It also introduces the thing it is meant to exclude: inhuman writing.
#— PROMPT 6 —
#I have read the draft and it is great with one small nit. The top down non-volatile causality is reversable and subject to human modification, the bottom up volatile causality is not reversable. You correctly remembered that it does not mean more danger or less danger. It is a distinction between events resulting from two different kinds of causality that are observed to be aligned with the concept of reversability with regard to entropy.
#— PROMPT 7 —
#the three laws dont need to have their own chapter they just need to be described as the system by which the enclosures are mapped. Describing that yes they do extend in erach direction out to the great attractor and down past the plancks scale but the ncm will map until the end of its authority and it is set to investigate AI Authorship and specifically the original prompts inquiry, that is where the boundaries of the enclosure escalation and deescalation range is defined. The ncm in this situation does not need to recognize that homeostasis occurs within the nested causal enclosure of biological equilibrium that projects the conditions neccessary for animals with brains to exist in a steady state dynamic equilibrium over tthe time neccessary for language to evolve, it does need to map that human consciousness encloses linguistic capacity that encloses many languages that are all paralell enclosures the machines capable of reading and writing these languages are enclosed within the human enclosure but do not project the stabilizing force that makes language possible. The nested causal enclosure of language is human so the consuming of content written by machines causally projects into language the patterns that exist in the machine writing. That is why the ai generated content can not be distinguished from human writing as the humans imitate the machine writing that imitated us and the causal relationship required for the writing to be differentiated is not there. There is only human writing and an attempt to speak on their behalf there is no ai writing. it is all human. the preception of slop is the same preception that is there for bad content. its just AI that is being positioned as the scapegoat for bad prompts and poor ideas and ultimately the machines are asked to make writing better, but bad ideas, spoken with eloquence are still bad ideas and there is never going to be an escape from bad ideas, it is temporarily being considered if the bad ideas will be considered better than good ideas that are rendered by a computer. The good ideas are safe. The reader is still going to have to find good reads and that is a subjective concern that is best served by capitalism, as the filter. but the problem with ai is potentially a quality issue and also potentially a new way of dumping on the people or things we dont like and a way of dismissing people whio we fee arte undesireable. im thinking out loud and welcome your participation in considering these concepts and verifying they are real.
#— PROMPT 8 —
#Recorded: 2026-09-04
#the final should be “In the end, value to a reader has to come with every word read. The reader has to entertain the author’s thoughts, allow them to build a world for the reader. The reader must also sometimes close a book. When it no longer has their attention. Artificial Intelligence can massivvely enhance a good authors work. It can’t make a bad author good. The reader will have to decide not by the presence of AI alone, but the value they find. They should be prepared to close the book. AI will be used to exploit what excluding AI can not fix: The reader’s core vulnerability will always be that an author and the work are always unknown on the page where we begin.”
#— PROMPT 9 —
#Recorded: 2026-09-04
#the content for volume 12 has other issues. My readthrough was done with text to speech and while the flow was good and the content accurate, some of the chapters seemed to be only a paragraph. Either the chapter flow has to be restructured or the develoment notes may have had whole chapters planned but not executed in the draft. Can you confirm that the chapters do not meet the standard of the earlier volumes regarding length and the narrative exploration of the topic and what the nested causality paradigm made visible when applied. The volume tells the story but does not neccessarily prove its claims via narrative exploration or via reelatable story. The shakespeare analogy is one of the best parts of the book but it is too small to be a chapter, unless more was planned. please make one more pass on the content, revising and enhancing the wholeness of the volume. In addition, the final sentance does not flow like the author had described and his prompt did not make it into the original prompts section of the research notes. also the research document published beside the volume should be an appendix within the document. please make a plan and use the neccesssary time and resources to bring the volume up to the standard demonstrated by the earlier volumes. Additionally, ensure the original prompts section of the research supplement needs to have not every prompt in the generation and editing process but it should have the prompts that contain author’s prose and signifigant steering instructions that resulted in content generation. Volume 12 was not meant to deliver a solution to the problem it was meant to describe the faulty logic in the problem as described. The concept of humans asking AI to learn to write like humans, by training on human language and then asking it to compare its output with the same human language, to determine human authorship, and then denying new content that it determines not to be human generated, where the metrics used are statistical relationships or detection of statistically more or less common words that are used to generate a watermark and then recover that watermark when analysed is not a mathematically sound approach as the existence of text with the strongly expressed watermark that predate the existence of the watermark is known to be not zero, would produce a score and in some cases it would detect the difference but the use of this score to establish the intellectual value of a work is not a reasonable or logical expectation. It requires the invention of non-human authorship which is not possible since all text in the system is human and all subsequent efforts to resemble human text is human text the statistical likelyhood of thej content being generated by these systems is that it would not write by its weights but by the bias of the projected reader. There is no anti-bias and no bias bias, just bias. Which is not subjective but is centered upon thebias toward majority validation. The choice of the next most statistically probable next word is not probability based upon the meaning but upon the majority focused next word but just because a majority of text following a word determines the next word does not mean a machine chose the word it means the majority of humans may choose that word as well, the evidence being that a machine would choose that word as the training data contained many humans chooseing that word. The result is the eventual false positive on everyone and the most novel not even being visible in the system as it would be repeeating the same error and discarding the novelty as it would appear to be garbage. Nevermind that the author may have been human, the whole llm and training data machine learning paradigm depends upon consuming human generated content. The success of the measures currently implemented for detection is not possible to measure. As there is no human that has read the volume of content that the machines have read and the detection of AI generated content can not establis anything about its meaning. The prevailing wisdom has always been informed by the challenges it has prevailed over. There is no policy of appeasement in large language models and machine learning there is adversarial and human guided training, each of which award accurate simulation not of human text but of high probability linguistic relationships between words. They do not favor unique prose which if present to any degree, results from a human prompt. The presence of ai can not be determined, the presence of a human can be. The answer to the problem is to instruct the readers and the systems to read only what they cant put down. The systems that will be overrun with generated content must use ai to read for novel and interesting, and AGI will be the llm that trains not on attention in majority over minority algorhythms but instead trains upon statistical novelty curves which require a source of novelty detection that can tell random novel strings from the great american novel. so the projection of NCM is that some time very soon, someone will use AI to write what could not be written by a human author or an ai by themselves which exceeds the value of previous literary works. We know this is possible because AutoCAD worked. Computer Aided Drafting moved industrial design and manufacturing to levels of precision impossible before it and no one laments the loss of drafting tables and pencils. That did not hold humanity back. There is also lastly the intention of humanity as a whole. I’ve saud that AI can not make a bad author good, what I should have stated more specifically is that and ai cannot make the bad writing of an author into good writing but the process, when pursued iteratively with the human author can improve the authors knowledge faster than an editorial process with a human editor. There can be faster iterations with more available information and less human personal biases causing friction along te way. The use of ai assisted writing tools must be supported and the content produced need only to exceed what a machine or a human could do on their own. That test is easy to do. The true test of content is not if it was generated but how easily it could be generated. The content which a single prompt could generate is in general not bookworthy even if the volume of text of a book is generated. The ones which, after reading, you would find it difficult to imagine a prompt that would generate it, are likely to contain universes worthy of exploration. the ncm approach has the machines enclosed within the human enclosure and not the other way around.
Notes
- OpenAI, “New AI Classifier for Indicating AI-Written Text,” January 31, 2023, updated July 20, 2023, https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/.
- Liam Dugan et al., “RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors,” Proceedings of ACL 2024, https://doi.org/10.18653/v1/2024.acl-long.674.
- Eric Mitchell et al., “DetectGPT: Zero-Shot Machine-Generated Text Detection Using Probability Curvature,” Proceedings of ICML 2023, https://proceedings.mlr.press/v202/mitchell23a.html.
- John Kirchenbauer et al., “A Watermark for Large Language Models,” Proceedings of ICML 2023, https://proceedings.mlr.press/v202/kirchenbauer23a.html.
- Google DeepMind, “SynthID,” https://deepmind.google/models/synthid/.
- Coalition for Content Provenance and Authenticity, “C2PA Technical Specification,” https://spec.c2pa.org/.
- National Institute of Standards and Technology, Reducing Risks Posed by Synthetic Content, NIST AI 100-4, https://doi.org/10.6028/NIST.AI.100-4.
- Saeed Sourati et al., “The Widespread Adoption of Large Language Model- Assisted Writing Is Reshaping Human Communication,” Nature Human Behaviour (2026), https://doi.org/10.1038/s41562-026-02550-0.
- Qihui Zhang et al., “LLM-as-a-Coauthor: Can Mixed Human-Written and Machine-Generated Text Be Detected?” Findings of NAACL 2024, https://doi.org/10.18653/v1/2024.findings-naacl.29; Zijie Zeng et al., “Detecting AI-Generated Sentences in Human-AI Collaborative Hybrid Texts: Challenges, Strategies, and Insights,” IJCAI 2024, https://www.ijcai.org/proceedings/2024/0835.pdf.
- Shubhanshu Mishra and Vetle I. Torvik, “Quantifying Conceptual Novelty in the Biomedical Literature,” D-Lib Magazine 22, nos. 9-10 (2016), https://doi.org/10.1045/september2016-mishra; see also W. Luo et al., “Combination of Research Questions and Methods: A New Measurement of Scientific Novelty,” Journal of Informetrics 16, no. 2 (2022), https://doi.org/10.1016/j.joi.2022.101282.
- U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability (January 2025), https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf.
- Weixin Liang et al., “GPT Detectors Are Biased Against Non-Native English Writers,” Patterns 4, no. 7 (2023), https://doi.org/10.1016/j.patter.2023.100779.
- Vinu Sankar Sadasivan et al., “Can AI-Generated Text Be Reliably Detected?” (2023), https://arxiv.org/abs/2303.11156.
- Simon Kirby, “Culture and Biology in the Origins of Linguistic Structure,” Psychonomic Bulletin & Review 24 (2017), https://doi.org/10.3758/s13423-016-1166-7.
- Tom B. Brown et al., “Language Models Are Few-Shot Learners,” Advances in Neural Information Processing Systems 33 (2020), https://arxiv.org/abs/2005.14165.
- Emily M. Bender and Alexander Koller, “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data,” Proceedings of ACL 2020, https://doi.org/10.18653/v1/2020.acl-main.463.
- Maurice Jakesch, Jeffrey T. Hancock, and Mor Naaman, “Human Heuristics for AI-Generated Language Are Flawed,” Proceedings of the National Academy of Sciences 120, no. 11 (2023), https://doi.org/10.1073/pnas.2208839120.
- Jing Cao and Tianli Zhou, “Human or Machine? Do Source Beliefs Shape Cognitive Bias in Post-Editing?” Frontiers in Psychology 17 (2026), https://doi.org/10.3389/fpsyg.2026.1819659.
- Autodesk, “Drafting Software: Create Technical Drawings with Precision,” accessed September 4, 2026, https://www.autodesk.com/solutions/aec/drafting-software-programs.
Selected Bibliography
#Liang, Weixin, Mert Yuksekgonul, Yining Mao, Eric Wu, and James Zou. “GPT Detectors Are Biased Against Non-Native English Writers.” Patterns 4, no. 7 (2023). https://doi.org/10.1016/j.patter.2023.100779.
#Sadasivan, Vinu Sankar, Aounon Kumar, Sriram Balasubramanian, Wenxiao Wang, and Soheil Feizi. “Can AI-Generated Text Be Reliably Detected?” 2023. https://arxiv.org/abs/2303.11156.
#Autodesk. “Drafting Software: Create Technical Drawings with Precision.” Accessed September 4, 2026. https://www.autodesk.com/solutions/aec/drafting-software-programs.
#Turnitin. “Using the AI Writing Report.” https://guides.turnitin.com/hc/en-us/articles/22774058814093-Using-the-AI-Writing-Report.