Are AI Tools Stealing Credit for Your Intellectual Property? Reclaiming Human Intellectual Property from the AI Labs
The statistical patterns and watermarks AI labs leave in human work, and detectors like Pangram that read them, are shifting credit for human ideas to the tools that helped write them. How that happened, what it means legally, and the path for authors, artists, musicians and developers to take their work back.
By Brian James Curry, Founder of Vector1 Research, Kansas City
In May 2023, an instructor at Texas A&M University-Commerce pasted his students’ final essays into ChatGPT and asked whether it had written them. ChatGPT said it had written every one. The instructor withheld grades for the class, and some graduating students temporarily had their diplomas held. Students produced Google Docs timestamps showing their drafting history, and the university later said several had been cleared (Rolling Stone). ChatGPT had no way of knowing who wrote those essays. It claimed them anyway, and a person treated the claim as evidence.
The episode was early and crude, but the pattern has held. Machines are now routinely asked whether a person did their own work, and their answers get treated as findings about authorship. A detector answers one narrow question, whether a model produced a particular sequence of words, and the answer gets read as a much bigger one: whether the author did the thinking. For writers, analysts and researchers who use AI to gather sources and put their own ideas on the page, those are different questions. Treating them as one creates an attribution problem with ethical, legal and economic consequences.
This piece covers what detectors and watermarks actually measure, documented cases where the labels went wrong, why a tool’s fingerprint ends up standing in for the author, the ethical and legal problems with labeling original work as AI-generated, where the trend leads if it continues, a proposal for fixing it, a path for reclaiming human work, and how to build an assistant that keeps the author’s ideas and voice primary.
Fingerprints and Watermarks Are Different Things
Two separate mechanisms get lumped together as “AI patterns,” and they work very differently.
The first is a fingerprint. Every language model has statistical habits: which words it tends to pick, how predictable its next word is, how evenly its sentences and paragraphs are paced. Nobody puts these there on purpose. They come from how the model was trained and how it samples words. Detectors such as Pangram learn to recognize those habits. Pangram’s technical report describes training a classifier against “synthetic mirrors,” AI-written counterparts of human documents matched on topic, length and style, and reports catching more than 99% of text from seven open models it never saw in training, with similar rates for Claude 3, GPT-4o and Llama 3 in a 2024 update (Pangram Labs, arXiv).
The second is a watermark, and that one is deliberate. Google DeepMind’s SynthID Text adjusts the probabilities of specific words as text is generated, leaving a pattern a detector can test for later. Google tested it on about 20 million chatbot responses, reported that users didn’t notice a difference in quality, published the method in Nature in October 2024 and open-sourced it (HPCwire). OpenAI built a text watermark of its own, reportedly 99.9% effective in internal documents, and held it back. Among the reasons it gave: almost 30% of surveyed ChatGPT users said they’d use the product less, and the company worried about stigmatizing non-native English speakers who rely on these tools (Search Engine Journal).
The distinction matters for authors. A fingerprint is a side effect you can reduce by doing more of the writing yourself. A watermark is a policy choice made by the company that runs the model, and it travels with the words whether or not you know it’s there.
Instructions Don’t Switch the Fingerprint Off
A reasonable author assumes that a tool told not to leave its mark won’t leave it. That assumption is wrong, and most tools don’t say so up front.
The typical sequence goes like this. An author brings an assistant original research, a thesis and a direction, and instructs it explicitly: no AI patterns. The assistant complies with what it can see. It strips out em dashes, stock phrases and familiar constructions, runs a style check, and reports a clean result. The author publishes, and a detector flags the piece as AI-generated anyway. The instruction was followed on the surface and ignored underneath, because the fingerprint was never in the surface features. It’s in the statistics of how the model chooses each word, and no instruction in a prompt reaches that layer.
The same thing happens when an author hands over writing that is entirely their own and asks only for an edit. Every sentence the assistant rewrites comes back carrying the assistant’s signature. A light edit introduces a little of it; a heavy rewrite of a human draft can introduce enough to make the whole piece read as machine-made. The author’s original work goes in clean and comes out marked, and the more helpful the edit, the stronger the mark.
Watermarks are worse on this point, because the user has no control at all. A watermark is applied by the provider during generation. No prompt, setting or polite request from the author removes it, and the author may not even know it’s there.
None of this means the fingerprint is permanent. Nothing proves it can’t be removed, and there is evidence it can be changed substantially. A model fine-tuned on one author’s writing takes on that author’s statistical habits instead of its provider’s. Changing how a model samples its words changes how predictable each word is, which is much of what detectors measure. Heavy human rewriting reduces the signal, and an independent study affiliated with Carnegie Mellon pushed individual passages from fully AI to fully human scores through repeated paraphrasing (eesel review of Pangram 4). Google itself says SynthID Text becomes less reliable after heavy rewriting or translation (HPCwire).
So an assistant that honors “leave no patterns in my work” can be built. The AI labs either haven’t solved it yet or have chosen not to. Some of that choice is pushed on them: the EU AI Act now requires providers to mark synthetic text, and detection vendors retrain as soon as models change. But a lab could still build an author mode, tuned to the person’s own writing, that keeps human-initiated work free of the provider’s signature. That this isn’t a standard feature of mainstream assistants says more about priorities than about what is technically possible.
The ethical failure isn’t that the tools have a fingerprint; that is a property of how they work. It’s that authors are left to discover it after publication, through a detector’s verdict, instead of being told before they start. A tool that can’t honor an instruction like “don’t leave your patterns in my work” should say so plainly when the instruction is given, and offer the modes that actually protect the author’s voice: suggestions instead of rewrites, research instead of drafting, and edits kept small enough that the words stay the author’s.
When the Labels Go Wrong
The Texas A&M episode wasn’t an isolated misfire. The record of AI detection since 2023 is a record of confident labels resting on weak evidence.
OpenAI launched its own AI-text classifier in January 2023. By the company’s own numbers, it correctly identified only 26% of AI-written text and wrongly labeled 9% of human writing as AI-written. OpenAI took it offline in July 2023, citing its “low rate of accuracy” (OpenAI). If the company that built the most widely used model couldn’t reliably tell its output from human writing, the question of what other tools were measuring was fair from the start.
That same year, Stanford researchers led by James Zou found that popular GPT detectors consistently flagged essays by non-native English writers as AI-generated, while classifying essays by native writers accurately (Liang et al., arXiv). Plainer, more careful prose, the kind people write in a second language, looked machine-made to the classifiers.
In April 2025, the Federal Trade Commission proposed an order against Workado, which had marketed its AI Content Detector as 98% accurate. Testing found 53% accuracy on general-purpose content, which the head of the FTC’s Bureau of Consumer Protection described as no better than a coin toss (FTC).
The consequences for individuals are now reaching courtrooms. In 2025, Orion Newby, a sophomore at Adelphi University, sued the school after Turnitin’s AI detector flagged a history paper he says he spent 15 to 20 hours writing. His instructor called the paper “too advanced” for a first-year student and gave it a zero, and the university upheld the finding. Newby, who has learning differences and works with tutors on his writing, is asking for the punishment to be set aside (Plagiarism Today).
Detection has improved since 2023, and vendors such as Pangram now report far lower error rates. But a more accurate detector is still measuring the same thing: who generated the words. None of these tools can see who chose the question, did the research or stands behind the conclusions. Better accuracy on the wrong question doesn’t fix the attribution problem. It makes the label more persuasive.
The Attribution Problem
A detector score describes how the words were produced. The trouble starts when it gets read as a statement about who did the work.
Consider what goes into a serious piece of writing. Someone picks a problem worth working on and forms a thesis about it. They decide what evidence would test it, run the analysis, throw out the parts that didn’t hold up, and choose which conclusions are strong enough to publish. Then someone turns all of that into sentences. In traditional publishing those jobs were often split between people: researchers, ghostwriters, editors, speechwriters. We didn’t treat the editor’s light touch, or even a ghostwriter’s heavy one, as erasing the author.
A detector sees only the last step. When it flags a piece, the honest reading is narrow: a model probably generated much of this text. The common reading is broad: the author didn’t do this. For an essay assigned to test a student’s writing, the narrow and broad readings mostly coincide, because the writing is the work. For an analysis built on someone’s original research, years of domain experience and their own code, they don’t.
The cost of that confusion lands unevenly. A practitioner who uses an assistant to turn notes into clean prose can be marked down as unoriginal, while someone with no original ideas who writes every sentence by hand passes. That gets attribution backward. The signal rewards whoever typed the words and ignores whoever did the thinking.
We already accept this distinction for other tools. A writer who spends an afternoon in a search engine, pulling studies and checking dates, doesn’t publish with a label that says “search-engine assisted.” Spell check, grammar tools and reference managers shape the final text too, and none of them are treated as co-authors. A language model that gathers research and helps put an author’s ideas into sentences is doing the same category of work, faster. The only difference is that it leaves a fingerprint a classifier can find.
The Research Assistant Problem
Academia worked out a version of this question long before language models. A professor sets the research question, designs the study, decides what the results mean and signs the paper. A research assistant gathers sources, cleans data, runs analyses and often drafts sections. The assistant’s work is real and gets acknowledged. Authorship goes to the people who shaped the ideas and who answer for the conclusions. The ICMJE criteria used across medical publishing say this plainly: authors make substantial contributions to the conception or analysis, draft or critically revise the work, approve the final version, and agree to be accountable for it. Contributors who don’t meet all four are acknowledged rather than listed as authors.
Publishers applied the same logic to AI early. In January 2023, Springer Nature said a language model can’t be credited as an author, because authorship carries accountability a tool can’t take on, and asked researchers to document AI use in their methods or acknowledgments instead (NISO).
A detector verdict quietly reverses that settlement. The tool never asks for credit; models don’t claim anything. But when a fingerprint in the prose becomes the headline fact about a piece, the effect is as if the research assistant’s name were moved to the top of the professor’s paper. The person who did the gathering and drafting becomes the presumed author, and the person who set the direction and carries the responsibility becomes an editor of someone else’s work.
A calculator makes the same point more bluntly. Nobody runs a mathematician’s proof through a check that flags it as “calculator-generated” because a calculator did the long division. The mathematician chose the approach, set up the problem and knows what the answer means. The arithmetic is assistance. When a detector looks at an article built on someone’s original research and labels it AI-generated, it is grading the long division and ignoring the proof.
The ethical weight here falls on how detector results get used, not on the tools themselves. A score that says “a model likely produced this text” is useful information. A score treated as “this person didn’t do this work” is a judgment the score can’t support.
The Ethics of Labeling Work as AI-Generated
There is an irony at the center of this. A detector like Pangram is itself an AI system, a machine learning classifier trained to make a judgment about people. When it labels an article “AI-generated,” a machine is deciding how much of a human’s work counts as human. That judgment deserves more scrutiny than it gets, for at least six reasons.
The label says more than the evidence does
“AI-generated” implies that the AI originated the work: the idea, the argument, the findings. What the classifier actually detected is a statistical pattern in the word choices. Calling an author’s original analysis “AI-generated” because a model helped phrase it is like calling a photograph “camera-generated.” The camera did capture the image. The photographer chose the subject, the moment and the frame.
Original thinking gets erased
The parts of a piece that took the most skill, such as choosing the problem, building the argument, designing the analysis and deciding what was true enough to publish, leave no trace a detector can read. The part that takes the least skill in the AI era, generating fluent sentences, is the only part it measures. A label built on that measurement misses the author’s contribution entirely, then writes over it.
Reputations change with no appeal
A detector score can follow an author into hiring decisions, client conversations and editorial review. The author usually can’t see why the score came out the way it did, can’t examine the model that produced it, and has no process for contesting it. A judgment that affects someone’s professional standing should come with an explanation and a way to respond. A probability from a closed classifier comes with neither.
Honesty gets punished
Authors who are open about using AI for research and drafting invite exactly this scrutiny, while people who hide their tools or have someone else do the work quietly face none of it. A system that penalizes disclosure teaches people to stop disclosing. That’s the opposite of the transparency these tools are supposed to support.
The burden falls unevenly
When OpenAI explained why it held back its own text watermark, one reason it gave was concern about stigmatizing non-native English speakers who rely on AI tools to write (Search Engine Journal). The same concern applies to detection more broadly. The people most likely to use a language model to express their ideas well, including non-native speakers, people with dyslexia, and experts whose strength is analysis rather than prose, are the people a blanket “AI-generated” label hurts most.
The wrong incentive takes hold
If the label is what gets judged, authors will optimize for the label. Some will stop using tools that make their research more thorough. Others will roughen their own writing on purpose so it reads as more human. Neither makes the work better. A measure that pushes people toward worse research or worse prose is measuring the wrong thing.
None of this requires pretending AI wasn’t involved. It requires describing its role accurately. “Researched and drafted with AI assistance; ideas, analysis and editorial decisions by the author” is a true sentence. “AI-generated” usually isn’t.
The Legal Exposure
An “AI-generated” label isn’t only a reputational problem. Under U.S. law it goes to the question of who, if anyone, owns the work.
In March 2025, the D.C. Circuit held in Thaler v. Perlmutter that the Copyright Act requires a human author, and affirmed the Copyright Office’s refusal to register a work created solely by an AI system (Goodwin). The Copyright Office’s January 2025 report on copyrightability concluded that prompts alone don’t give a user enough control to be the author of AI output. Human contributions to AI-assisted work can still be protected, but where the line falls is decided case by case.
That makes a detector’s label a claim about the one fact copyright depends on. A published score saying an original analysis is AI-generated invites anyone who copies it to argue there was no human author to protect. Authors who use AI assistance need records that show their contribution: drafts, version history, research notes and a log of what the assistant did.
The exposure runs the other way too. When schools and employers treat a detector score as a finding of dishonesty, they take on due-process risk, and lawsuits like the Adelphi case are testing how far that risk extends. Vendors aren’t exempt either: the FTC’s action against Workado shows that accuracy claims for AI detectors fall under consumer-protection law like any other product claim.
None of this is legal advice, and authors with commercial work at stake should talk with an intellectual property attorney. The direction is clear, though. Attribution has become a legal question, and detectors are being used to answer it without being built to.
Code, Images and Music: The Same Gray Area
Writing isn’t the only place where a tool’s involvement is starting to outweigh a human’s contribution. Code, images and music are further down the same road, and the cases there show where text is heading.
Code
Software developers built the corpus that coding assistants learned from: decades of public repositories, most of them released under licenses that require attribution. In Doe v. GitHub, programmers alleged that Copilot sometimes reproduced their licensed code verbatim or nearly so, with the attribution and license terms stripped out. On September 16, 2026, the Ninth Circuit affirmed dismissal of their claim under the Digital Millennium Copyright Act, reasoning that a model predicting likely output differs from a system returning stored copies. Two contract claims based on the open-source license terms remain in the district court, and the panel didn’t decide whether Copilot’s output infringes copyright (Haynes Boone).
The result is a two-way squeeze on developers. Code a human wrote can resurface through an assistant with the author’s name and license removed. And code a developer designs and directs with an assistant’s help faces the same human-authorship test as AI-assisted prose. The architecture, the decisions about what the system should do and how it should fail, are the developer’s work, yet the lines of code are what get scanned, scored and argued over.
Images
Images reached the Copyright Office first. In February 2023 the Office ruled on Zarya of the Dawn, a graphic novel by Kristina Kashtanova. It protected her text and her selection and arrangement of the book, but refused protection for the individual images she produced with Midjourney (Reuters via The Express Tribune). Artist Jason Allen’s Théâtre D’opéra Spatial, which won a prize at the 2022 Colorado State Fair, was also refused registration. Allen says he refined his prompts 624 times and spent about 114 hours selecting, arranging and directing the work, and he has sued the Copyright Office in federal court in Colorado (The Fashion Law). In the U.K., Getty Images took Stability AI to trial over the images used to train its model. Getty abandoned its primary infringement claims during trial, and in November 2025 the High Court dismissed its secondary copyright claims, upholding only trademark findings the judge called “extremely limited in scope” (Cleary Gottlieb).
The law has faced this question before, with a different machine. When photography arrived, critics argued that a photograph was produced by a camera, not by a person. In 1884, in Burrow-Giles Lithographic Co. v. Sarony, the Supreme Court disagreed and upheld copyright in Napoleon Sarony’s portrait of Oscar Wilde, because the photographer posed the subject, chose the costume, arranged the light and composed the frame. The camera captured the image. The person conceived it.
An AI tool standing in for the author of a human-conceived image is the camera claiming ownership of the photograph. The person who imagined the scene, directed the composition, rejected hundreds of attempts and chose the final one did what Sarony did. The tool did what the camera did. A rule that hands the credit, or the uncertainty, to the instrument instead of the artist would have looked absurd in 1884, and it should look the same way now.
Music
Music has the longest history of all with machines in the creative loop, and it shows how strange the current turn is.
In June 2024, Universal Music Group, Sony Music and Warner Music Group sued the AI music generators Suno and Udio, alleging the companies trained their models on copyrighted recordings without permission. Universal settled with Udio in October 2025 and Warner settled with Suno that November, both in deals built around licensed models. As of May 2026, Sony’s cases against both companies were still pending (Dubspot). Once again, the human corpus came first, and the dispute is over who gets paid for what the models learned from it.
The Recording Academy drew its own line in 2023: only human creators are eligible for Grammy awards, though songs with AI elements can still be considered if a person made a significant creative contribution. Recording Academy CEO Harvey Mason Jr. put it plainly: in songwriting categories, a song has to have been written mostly by a human (NPR).
The analogy writes itself for anyone who has worked in a studio. A mixing console shapes every sound on a record and has never claimed a writing credit. Auto-Tune corrects a singer’s pitch on countless hit records, and nobody lists it as a vocalist. Drum machines, samplers, synthesizers and vocal synthesis software have all generated sounds no human performed directly, and the credit has always gone to the person who wrote the song, chose the sounds and made the record. An AI tool treated as the author of a human-composed song is the mixing board claiming the publishing rights, or the vocal synthesizer claiming to be the singer.
A Hypothesis: When the Machine’s Claim Reaches Court
Today an AI system can’t sue anyone. It isn’t a legal person, and courts have refused to treat it as an author. But the pieces of a claim against human authors are already being assembled, and none of them require the AI to be the plaintiff.
Consider how it could unfold. An analyst publishes original research: her question, her data, her conclusions, drafted with an AI assistant. A competitor republishes large parts of it. When she sues for infringement, the defense runs her article through a detector, produces a high AI score, and argues the text is uncopyrightable machine output with no human author behind it. The tool never asserted anything. Its fingerprint became the infringer’s best evidence, and her thesis, research and judgment, the parts that made the work worth copying, never show up in the score at all.
Now extend it. Watermarks are controlled by the companies that run the models. As more text carries a provider’s mark, the mark becomes a record of provenance owned by a company and attached to an author’s ideas. It takes only a change in terms of service, a licensing deal with a platform or an aggressive reading of ownership for that record to become a claim. A provider wouldn’t need to prove it had the ideas. It would only need the author to be unable to prove she did.
The reverse lawsuit is no hypothesis; it’s already happening. Authors sued Anthropic over training on pirated copies of their books downloaded from Library Genesis and Pirate Library Mirror. In July 2026 a federal court gave final approval to a $1.5 billion settlement, about $3,000 per work, and required the files to be destroyed (Authors Guild). So far the lawsuits run from humans to machines: people asserting rights over work the models absorbed. The hypothesis here is that the direction can flip, with the machine’s fingerprint used to take rights back from the people whose work trained it.
Whether it plays out in court or only in hiring decisions, platform rankings and editorial reviews, the mechanism is the same. Whoever controls the label controls the attribution.
The Counterclaim: Humans Suing to Take Their Work Back
If a machine’s fingerprint can be used to take rights from people, people will go to court to take them back. None of the following theories has been tested against an AI provider over watermarks or patterns. Each of them could be, and the first case that succeeds will change how every provider and detection vendor operates.
Breach of the provider’s own promises
Where a provider’s terms assign ownership of outputs to the user, a watermark that is later used to dispute that ownership sits in tension with the provider’s own contract. An author who paid for a service, relied on its terms and then saw the provider’s mark used to challenge her rights would have a straightforward argument: the company sold her the work and kept a way to take it back.
Deceptive practices
A paid tool that embeds an undisclosed signal in a customer’s work, while marketing that work as the customer’s own, invites consumer-protection scrutiny. The FTC has already shown it will act on misleading claims in this market with its Workado order over detection accuracy. Hidden marking of paying customers’ work is a natural next question for regulators and state attorneys general.
False information about authorship
U.S. copyright law already prohibits knowingly providing false copyright management information, which includes information identifying a work’s author, when it’s done to enable or conceal infringement. A watermark or detector label asserting that a human-authored work is machine-made is, in effect, a statement about authorship. Whether it counts as copyright management information is an open question. The Ninth Circuit’s September 2026 ruling in the Copilot case shows courts are actively working out how that statute applies to AI, and a plaintiff whose authorship was falsely attributed to a machine would be asking them to apply it in the other direction.
Defamation and commercial harm
When a detection vendor’s verdict is published or relied on to say that a writer, analyst or artist passed off machine work as their own, it’s a statement of fact about their honesty. If the statement is false and costs them a contract, a job or a degree, the person harmed has the elements of a claim, and the vendor’s own accuracy marketing may become evidence against it.
Declaring authorship directly
The most direct remedy is asking a court to declare who the author is. Jason Allen’s case against the Copyright Office already seeks a declaration that his AI-assisted image is a work of human authorship. As watermarks and detector scores spread, authors will seek similar declarations against providers and vendors whose marks and labels say otherwise, along with orders to correct the record.
The pattern from the training-data cases is instructive. Authors, artists, musicians and developers didn’t wait for the law to settle before suing over how their work was absorbed, and those suits produced billion-dollar settlements and licensing deals. Suits over how their work is labeled are the logical next wave. Providers and vendors that want to avoid it have a simple option: describe the tool’s role accurately, and stop treating a statistical signal as a claim on someone else’s ideas.
Detector Liability: Labeling the Author’s Work as the Tool’s
The sharpest version of the liability question involves the detection vendors themselves. Picture the case that matters most. The author supplies every idea, concept and argument, directs the structure, makes every edit and approves every change. The AI follows instructions and produces sentences to order. A detector then labels the result AI-generated, and that label travels to an editor, a client or a review board as a finding about whose work it is.
What the labels actually measure
Detection vendors have become more granular, which makes the gap easier to see, not harder. Pangram 4, released in July 2026, labels individual tokens as Human, AI-Assisted or AI-Generated, and classifies whole documents as Human, AI-Generated, AI-Edited or Humanized-AI (eesel review of Pangram 4). Every one of those categories describes who produced the words. None describes who conceived the work. A model drafting sentences to an author’s detailed specification and a model inventing an article from a one-line prompt can produce text that earns the same label, because the classifier sees only the output. There is no category for “the author’s work, typed by a tool,” and a classifier reading text alone can’t create one.
The vendors know the stakes. According to the same review, Pangram’s own ethics statement acknowledges a non-zero error rate and that errors can cause reputational and other harm, and its model card says its scores are not calibrated probabilities. A vendor that documents both of those facts has also documented that it understands how its labels can hurt people and how easily they can be misread.
The theories of liability
None of these claims has been tested against a detection vendor over AI-assisted work. Each follows from law that already exists.
Defamation and false light. A label presented as “AI-generated” is, in ordinary reading, a statement that the work isn’t the author’s. When that statement reaches an editor or employer and the author’s ideas, structure and edits were entirely their own, the strength of a claim turns on a question this whole piece is about: whether the label is read as a statement about the words or about the work. In practice, institutions tend to read it as the second.
Negligent misrepresentation. Schools, publishers and employers buy detectors to make decisions about people, and vendors market accuracy figures to win that business. If the marketing implies the label identifies authorship, when it actually estimates token production, institutions that relied on it and the people harmed by those decisions both have an argument that the product was misrepresented. The FTC’s Workado order shows regulators already treat detection accuracy claims as actionable.
Failure to warn. A vendor that knows its scores aren’t calibrated probabilities, and knows its categories can’t capture who conceived a work, has a duty question to answer about how prominently it says so. A report that displays “AI-Generated” in bold, with the limits buried in a model card, invites the argument that the warning was inadequate for the decisions customers were making with it.
Tortious interference. When a detector report is the reason a publisher pulls an article, a client cancels a contract or a platform demonetizes a writer, the author has lost a business relationship because of a third party’s statement. Where the statement was wrong about the work’s authorship, that loss is the core of an interference claim.
Discrimination and automated decisions. The Stanford study found detectors flagging non-native English writers, and the Adelphi plaintiff has documented learning differences. Institutions that act on detector scores against those groups take on disparate-impact and disability-accommodation exposure, and vendors that sell into those decisions are part of the chain. In Europe, the General Data Protection Regulation gives people a right not to be subject to decisions based solely on automated processing that significantly affect them, which a detector-only finding of misconduct arguably is.
What vendors would need to change
The exposure points to its own fix. Detection reports could describe what they measure in plain language: an estimate of how the text was produced, not a finding about who did the work. Vendors could let authors submit drafts, version history and contribution logs to sit alongside the score, publish error rates for each category, and build an appeal process into the product instead of leaving it to each institution. Until they do, every “AI-generated” label applied to work whose ideas, concepts and edits belong to its author is a statement the vendor can’t stand behind, and the law has long experience with statements like that.
The End Stage: The Human Corpus, Taken Twice
Language models learned to write from people. Their training data is the accumulated output of human writers, researchers, teachers and journalists, and the fingerprint detectors look for is a compressed average of that human writing. As models improve, their output moves closer to the best human prose, and the boundary detectors patrol moves with it, deeper into territory that belongs to people. The Stanford finding was an early signal: careful, plain writing by second-language authors already reads as machine-made to the classifiers.
Follow that to its end and the human corpus is taken twice. First its words and ideas are absorbed as training data, often without permission or payment, as the Anthropic settlement made plain. Then the styles learned from that corpus are reassigned to the machines, so that a person whose writing resembles what the models absorbed has to prove they’re human. An entire generation’s ideas, research and ways of writing end up credited, by default, to the systems that learned from them.
That end state isn’t inevitable. It follows from one mistake, treating the source of the words as the source of the ideas, and it can be prevented by building and using tools that keep the record straight.
A Proposal: Credit Follows the Human
AI tools are part of human evolutionary development, the latest step in a line that runs from the stone tool to the printing press to the word processor. Every one of those tools extended what a person could make, and none of them took the credit. Nobody gives a novel’s authorship to the keyboard it was typed on or the computer that formatted it. AI tools should be treated the same way: they shouldn’t receive credit for work a human originated.
That principle leads to three proposals.
1. AI models must follow human instructions to leave no marks or patterns
When an author instructs a tool to leave no watermark, signature or detectable pattern in their work, the tool should comply, and providers should be required to build that capability. Today an instruction can’t reach the statistical layer where a model’s fingerprint lives, which is exactly why it needs to become a design obligation rather than a polite request. That means unmarked modes, edit-only modes that preserve the author’s own wording, and models tuned to the author’s voice instead of the provider’s. Where a provider can’t yet meet the instruction, it must say so before the work begins, not after a detector finds out.
2. Strict legal limits on marking human-initiated content
Watermarks and embedded signals should be prohibited in content a human initiated, directed and approved. Current rules point the other way: the EU AI Act requires providers to mark synthetic text in machine-readable form, exempting only assistive editing and outputs that don’t substantially alter the input. That exemption should be extended to any work where a person supplies the ideas, the direction and the final approval, and laws elsewhere should follow the same line. Marking should be reserved for content a machine originated, never imposed on work a human conceived.
3. Ban AI identification tools, or expect class actions
AI identification tools that label human-originated work as AI-generated should be banned. They measure who produced the words and present the result as a verdict on who did the work, and that is a judgment they can’t support. Until lawmakers act, expect class-action lawsuits to do the work regulation hasn’t: from students flagged by school-wide detection programs, writers whose work was pulled or downranked, and professionals whose credibility took the hit for a classifier’s call. The training-data cases showed how quickly individual grievances over AI become billion-dollar class settlements. Detection is next.
The principle underneath all three is simple. Human ideas, prompts and inputs are the key to AI outputs, not the other way around. A tool that only produces what a person asked for, in the direction that person set, is doing that person’s work. The credit, the ownership and the accountability belong with the human.
A Path to Humans Reclaiming Their Original Ideas and Work from AI Labs
Reclaiming authorship doesn’t have to wait for new law. Some remedies exist today, some are being tested in court right now, and the rest can be built. The path runs in that order.
What authors can do now
Register what is yours. The U.S. Copyright Office registers AI-assisted works when applicants identify and claim the human-authored parts and disclose the AI-generated material. Registration is the foundation of every later claim: timely registration makes the certificate evidence of ownership and opens the door to statutory damages in an infringement case. An author who directs AI-assisted work should register it and describe their contribution precisely: the ideas, structure, selection, arrangement and edits.
Keep a provenance record. Drafts, version history, research notes, prompts and a contribution log are the evidence that answers a detector score. Content-provenance standards such as C2PA content credentials let creators attach signed records of how a work was made, so the author’s own account travels with the file.
Control the corpus. Authors can use robots.txt directives aimed at AI crawlers, opt out of training in the data settings of the tools they use where providers offer it, and negotiate enterprise agreements that prohibit training on their inputs and confirm their ownership of outputs. In Europe, the AI Act requires providers of general-purpose AI models to follow a copyright policy that respects text-and-data-mining opt-outs and to publish a summary of their training content, which gives creators a way to see whether their work was used.
Remedies being tested
The training-data lawsuits have already established that human creators can recover value from AI labs. The Anthropic settlement provides about $3,000 per work for authors whose books were taken from pirate libraries, and the record labels’ cases against Suno and Udio produced settlements built around licensed models trained on the labels’ music. Collective action works, and it is the model for what comes next.
The narrower claims are moving too. Developers in the Copilot case still have contract claims based on open-source license terms after their copyright-management claim was dismissed, which keeps open-source licenses alive as enforceable promises. Students and writers flagged by detectors are suing institutions, and the theories laid out above (defamation, misrepresentation, failure to warn, interference) are ready to be brought against vendors when the right case arrives.
Future iterations
The next stage needs law and tools built around human authorship rather than machine detection.
A human-direction standard for authorship. Copyright law should recognize that a person who originates the ideas, directs the work and approves the result is its author, whatever tool produced the sentences. A registration category for human-directed, AI-assisted work would give that recognition a formal home.
A right to dispute AI labels. Credit bureaus must let consumers see their files and dispute errors, because their reports decide who gets a loan. Detection reports now decide who gets a grade, a job or a byline. Vendors should face the same duties: disclosure of what was measured, a right to dispute, and correction when the label is wrong.
Compensation for the corpus. The settlements so far are one-time payments for past use. A lasting system would license human work for training on an ongoing basis, through collective licensing or royalties, so the people whose ideas built the models share in what the models earn.
A right to remove work from models. Machine unlearning, techniques for removing the influence of specific training data from a model, is an active research area. As it matures, creators should gain an enforceable right to have their work taken out of models trained on it without permission.
Author-owned models. The strongest protection is ownership. Authors and organizations running their own models, fine-tuned on their own work, in environments they control, decide what is marked, what is retained and what is shared. That is where this ends up: human ideas held by the humans who had them, with AI working for the author instead of claiming the author’s work.
None of this is legal advice. Anyone with significant work at stake should talk with an intellectual property attorney about registration, contracts and claims.
How to Build an Assistant That Keeps the Author First
The goal isn’t to hide a tool’s involvement. It’s to build a system where the tool does assistant’s work and the author’s ideas and voice stay on top, so the finished piece reflects who actually did what. The design comes down to six choices.
1. Split the jobs by layer
Give the assistant clearly bounded roles: a research layer that searches, retrieves and cites sources; an analysis layer that runs code and checks math; and a language layer that works only on text the author supplies. The first two layers produce facts and results. Neither should produce finished prose. Most fingerprint problems start when one general-purpose model is asked to do all three at once and hands back a polished draft.
2. Make the author’s draft the input, not the output
The author writes or dictates first, however rough. The language layer returns suggested edits as tracked changes against that draft, and the author accepts or rejects each one. Set a limit on how much of any paragraph the system may rewrite. If it needs to rewrite more than that, it should say what’s unclear and hand the paragraph back. The words on the page stay the author’s, with the assistant acting as an editor.
3. Train the language layer on your own writing
An open-weight model fine-tuned on your published work learns your sentence rhythm, your vocabulary and the way you build an argument. Running it in your own environment also means you decide what, if anything, gets marked on its output. A general-purpose commercial model brings its provider’s habits and its provider’s policies with it.
4. Keep a contribution log
Record who contributed what: which ideas and decisions were yours, which sources the research layer found, which edits you accepted. Version history in the drafting tool does most of this automatically. A log like that is a better attribution record than any detector, because it describes the actual process instead of inferring one from word statistics.
5. Measure drift against your own baseline
Compare each draft with a sample of your earlier writing on simple measures: sentence length variation, vocabulary, paragraph structure. Run your own pre-AI writing through the detectors you care about so you know your baseline. If a draft drifts away from how you write, the language layer is doing too much, and that’s a signal to cut its role back, not to paper over the difference.
6. Stay on the assistive side of the line, and own the result
Regulators are drawing the same line. The EU AI Act’s transparency rules, which apply from August 2, 2026, require providers of generative AI to mark synthetic text in machine-readable form, but they exempt systems that perform an assistive function for standard editing or don’t substantially alter the input. Publishers of AI-generated text on matters of public interest must disclose it, unless the content has gone through human review or editorial control and a person holds editorial responsibility for it (EU AI Act guide). In other words, the law already recognizes that editorial control and accountability are what make a piece someone’s work. A system built around the author’s draft, with the author approving every change and answering for the result, sits on the right side of that line by design.
Credit Belongs Where the Thinking Happened
Authorship has never meant typing every word. It has meant choosing the question, doing the reasoning, standing behind the conclusions and answering for them when they’re wrong. Research assistants, editors and calculators have always helped with the rest, and nobody mistook their help for the work itself.
AI assistants belong in that same tradition. They can gather, check, compute and tidy. They can’t decide what’s worth investigating, and they can’t be accountable for what gets published. A fingerprint in the prose shouldn’t be allowed to override that, and a detector score shouldn’t be read as a verdict on whose ideas a piece contains.
The practical answer is to build systems that keep the author’s draft at the center, keep a record of who contributed what, and let the person who did the thinking keep the credit for it. That standard is available to any author doing original research with AI in the loop, today, without waiting for detectors or regulators to catch up.
About the Author
Brian James Curry is a Kansas City AI researcher and data scientist, the founder of Vector1 Research, an independent applied AI lab, and of KC AI Lab, which began in 2016 as the Kansas City Machine Learning Group and grew to more than 1,500 members. He has more than twenty years of experience leading analytics, data science and machine learning at enterprises including Hallmark, Garmin, Koch Industries, Tractor Supply Company and Vail Resorts. His open-source work, including Papilon, PyCausalSim, MeaningFlow, Daedalus and tincast, is on GitHub. He writes on Substack and Medium.
