There are two questions that can be asked of work made with AI, and since August 2026 they have mostly been asked by different institutions.

The first is the way the platforms do it. Medium requires that "any story incorporating AI assistance be clearly labeled as such." Deezer runs every incoming track through a detector and tags the ones it judges fully machine-made. Spotify puts a badge on artist profiles it believes are AI personas and keeps them out of its recommendations by default. They draw the line in different places, from any assistance to fully generated tracks to invented artists, but each line describes how the work was produced. The question underneath is the same: was the tool there, and how much of it?

The second is the way the European Union's AI Act does it. Article 50, which applies from 2 August 2026, says that whoever uses an AI system to generate or manipulate text "published with the purpose of informing the public on matters of public interest" must disclose that the text was artificially generated. Then it adds an exception. The duty does not apply where the content "has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content." The law has its own marking rules as well, but here it asks a different question: who answers for it?

The first question can be answered by software at the scale of tens of thousands of tracks a day. The second cannot be answered by software at all. That difference, between what can be detected and what has to be held by a person, is the subject of this piece. It is also, I should say at once, my own situation. I write The Manifest with AI. I have done so since I started publishing in September 2025, and this piece was researched and drafted with Claude, a model made by Anthropic. What I do, and what a label cannot see, is set out after the studies, as one case of what they describe.

A Hundred Poems and 2,330 Readers

The cleanest evidence on how the label works comes from poetry, because poems are short enough to rate and familiar enough that people think they can tell.

Brian Porter and Edouard Machery, philosophers at the University of Pittsburgh, chose ten well-known poets writing in English, from Chaucer to the contemporary poet Dorothea Lasky, and took five poems by each. They asked ChatGPT 3.5 to write five poems in the style of each poet, which gave them a hundred poems, half human and half machine. The study appeared in Scientific Reports in November 2024.

In the first experiment, 1,634 people tried to sort them. They were right 46.6 percent of the time, below what guessing would give, and they were more likely to take a machine poem for a human one than to recognize a poem a person had written.

In the second, 696 people rated the poems on fourteen qualities. Across all of them, whatever they had been told about authorship, the AI poems came out ahead on thirteen. The widest gap was rhythm. Originality was the only quality where the difference was too small to count. The authors' explanation for both results is the same: the machine poems were simpler and more direct, and readers took the difficulty of the human poems, the work of an Eliot, for incoherence.

The researchers also told some readers that the poem in front of them was by a person and others that it was by AI. Readers told it was AI gave lower ratings for overall quality than readers told it was human, regardless of who had actually written it. The text did not change; the belief about its origin did, and the rating moved with the belief.

It is worth being exact about what that shows. The study does not show that readers were wrong to mark down machine work. It shows that, for the same text, the information about where it came from carried weight of its own.

Eight Clips Marked "Composer: AI"

Music adds something the poetry study cannot. A reader can decide a poem is worse; a listener's experience while the music is playing is harder to argue with.

Sarah Wu, now at Stanford, and Kevin Holmes at Reed College tested that experience in two experiments with 399 American participants, published this year in Cognitive Research: Principles and Implications. Their measure was not liking. It was what they call narrative listening: whether instrumental music makes a listener imagine a story, and how absorbed they become in it.

In the second experiment, listeners heard eight clips. Four were human compositions, by Beethoven, Ketèlbey, Mozart and Ravel; four were made by an AI composition program called AIVA. Each clip carried a label on screen, "Composer: Human" or "Composer: AI," assigned regardless of who had actually written it. Under the human label, listeners reported imagining a story 57 percent of the time. Under the AI label, 50 percent. Their engagement with those stories, on a six-point scale, dropped from 3.42 to 3.23. The drop in engagement is small in statistical terms. It held whether or not the piece was really by a machine, with two exceptions at the level of single clips: the real Ravel, and on engagement the real Beethoven.

The authors' account of why is the part I find most useful. Listeners attributed less communicative intention to music they believed was machine-made, and intention was a strong predictor of how much they listened for a story; the authors suggest this may be what drives the label effect. A story seems to need someone behind it who meant something, and a label that suggests otherwise may take some of the meaning away before the music starts, even when a person chose, steered or edited what the machine produced.

Generation, Editing and 786 Raters

Neither of those studies tests the case most people now face, which is not pure machine output but a person working with a model. That is the question a label like "AI-assisted" claims to answer, and it is the question that decides whether a label is informative or merely accurate.

Zhuoyan Li and Ming Yin at Purdue, with Chen Liang and Jing Peng at the University of Connecticut, tested it directly in a study presented at the EMNLP conference in November 2024. Writers produced short argumentative essays and short creative stories in one of three ways: on their own, with AI editing what they had written, or with AI generating the text. That gave 407 pieces of 200 to 250 words. Then 786 other people rated them, some told how each piece was made and some not.

Disclosing that AI had generated the text lowered quality ratings for essays and for stories. Disclosing that AI had only edited the text lowered ratings for stories and left essays untouched. Disclosure also made judgments of the same piece more scattered, as though readers were no longer sure what they were evaluating. And the readers most confident in their own writing marked down the most.

The ranking effect is the striking number. Among the essays in the top tenth by rating, AI-generated pieces made up 63.6 percent for the group of readers left uninformed about method, and 24.5 percent for a different group of readers who knew. The essays were the same; only the information given to each group differed.

Read with the poetry and music studies, the picture sharpens. Readers do not react to one thing called AI. They distinguish between a machine that writes and a machine that helps, and they forgive help in some kinds of work and not in others. Any label that cannot make that distinction is throwing away information readers have shown they use.

I don't know where a piece like this one falls on that line. It argues, which is where the study found help forgiven, and it also tells a personal story, which is closer to where it was not.

Thirteen Months, 376 Pieces

What "a machine that helps" means in practice is easier to show than to define, so here is one case, the one I know from the inside. When I began, the method was simple and the results were uneven: a subject, a model, a draft, a publish button. Some early pieces were thin, some near-copies of their Medium versions. On 9 June 2026 every one of the 293 posts then live was saved as a snapshot, and the early ones were rebuilt to the standard the later ones had reached.

What changed between the first month and the thirteenth is less the model than the set of rules that now stand between a draft and the publish button, and every one of those rules exists because something went wrong first.

In May 2026 a fixed pre-publication check went in: how a piece handles time, what systems an effect travels through, who profits, what is missing from the official account. A month later it turned out to have run after one major piece went live, so it became a gate rather than an audit. In the same weeks the tag behind the badge that marks claims as documented or speculative stopped being applied, and about twenty-six articles lost it before anyone noticed. No piece can now be published without it.

On 10 June a different kind of check went in, the one I would point to if asked what separates this from typing a prompt. At a few hundred publications, the deepest risk is no longer a wrong number. It is that the archive's own ideas start to feel verified simply because they are familiar. So every major piece now goes to a second, independent agent that has not seen it being written, with web access and a single instruction: try to break it, by checking every load-bearing number, date and name against outside sources. The first run, on eight pieces, found one real error and six precision problems, none of them fatal to a thesis and all of them caught before a reader found them.

In August, readers pointed out that one widely read piece had a predictable rhythm. They were right: rules meant to make sentences memorable, applied to every section, produced the same kind of sentence too often. The fix was a ceiling, counted by a script before anything goes out, on the short contrastive lines models produce so easily, and a rule that the body says at least once what I am not sure of.

In September came a gate before any writing starts, described in the next section. In October the whole system moved into a single database that holds every piece from idea to publication and every gate it passed or did not, with a note on why. A gate may be skipped; it may not go unrecorded. At the time of writing 376 pieces are live on this site.

This piece went through the same machinery. Its first central claim, that resistance to AI work comes from the label rather than the quality, was broader than the evidence, and outside reviews pushed it back to what the studies support. The independent check found that I had the range of poets in a study wrong and had attributed a finding to the wrong group of readers. Both are fixed in the sections above. Two decisions went against what the machinery proposed. The first opening the model drafted led with readers' criticism of an earlier piece; I cut it, because a disclosure should read as a statement of method rather than a defence. And both outside reviews preferred a different title, built on a formula that has performed well for this archive. I kept this one, because it says what the evidence supports and no more.

None of that is visible in a label that says "AI-assisted." The studies show that readers act on what a label tells them. This is some of what it leaves out.

What the Model Asks Before a Word Is Written

The rules above are the history. What they add up to is a model of how a piece gets made.

Every subject starts with a fixed sequence of questions, asked before any drafting. What is the public story, stated neutrally? What system underneath decides what is physically, economically or legally possible? And where is the discontinuity: two facts that are each documented and rarely placed side by side? From there the subject is pushed down a chain, from the event to the mechanism, the infrastructure, the constraint and the consequence, and at each step there is a test. Is the angle actually new, or a known frame in new words? Does the hidden variable decide the outcome, or only contribute to it? Is the named bottleneck the constraint, or one of several? A subject that fails is sent one layer deeper or dropped, and the drop is recorded too, so the archive remembers what it chose not to write.

Inside the text, every load-bearing claim is sorted three ways: what the source says, what I infer from it, and what I cannot yet establish. Claims carry levels, from documented fact to marked hypothesis to open speculation, and the reader sees the difference in the wording. Every piece on a contested subject has to state the strongest objection to itself. Every piece has to name the evidence that would make me reverse it, and the longer ones end with an evidence map. When a piece makes a dated forecast, the archive goes back later and scores it in public.

The choices in that model are mine. AI helped write some of the rules down and applies all of them, quickly and at scale, to its own drafts as hard as to anything else. What I built is the set of questions, the order they come in, the thresholds a piece has to clear, and the decision about what happens when it does not. That is the human contribution in this archive. It is also the one thing a label on the finished text has no way to record: "AI-assisted" describes a piece that went through all of this and a piece that went through none of it in exactly the same two words.

Two Rulebooks, August 2026

So far the evidence is about individual readers and one writer's method. The question that makes this more than psychology is what happens when platforms build the label into the systems that decide what readers ever see.

Deezer reported on 21 July that on peak days in June, fully AI-generated tracks were more than half of everything delivered to it: nearly 90,000 a day. It says its detector catches them with 99.8 percent accuracy, and that it tags them, removes them from algorithmic recommendations and keeps them out of editorial playlists. In the same release it put their share of all streams at between 1 and 3 percent, and said that up to 85 percent of the streams on those tracks in 2025 were fraudulent, generated by bots to draw money from the royalty pool.

Those figures contain three separate problems that end up under one tag. Fraud takes money from a shared pot. Volume clutters search and recommendation with tracks nobody requested. Competition, AI music taking listening from human musicians, is the third, and Deezer's figure of 1 to 3 percent of streams suggests that direct displacement is still limited. It does not show how much listening would otherwise have gone to people, or what the flood does to search and discovery. A tag that marks a track as AI-made says nothing about which of the three problems it belongs to.

There is an economic chain under all of this, and it explains why the label matters now rather than five years ago. On my reading: when production gets cheap, supply grows faster than attention; when attention is the scarce thing, selection becomes the valuable thing; and whoever sets the categories selection runs on shapes the market. The figures above do not prove every link, but they show the first two plainly: half the uploads, a few percent of the listening.

Spotify's approach runs through the same channels. Since April 2024 a track must reach 1,000 streams in a year, from a minimum number of unique listeners, to earn recording royalties at all, and Spotify says tracks with a high probability of being AI-generated are ten times more likely to fall short. On 20 October it said it had removed more than 50 million AI "slop and spam" tracks from its recommendations this year. Profiles it judges to be AI personas carry a badge and are left out of recommendations by default unless a listener follows them.

Medium does not detect; it asks. Its rule requires a label on any story that incorporates AI assistance. Writing that is AI-generated may not go behind the paywall, disclosed or not. Stories that use AI assistance without saying so are limited to the author's own network.

Set that against Article 50. The AI Act's provider rule, that systems generating synthetic content must mark it in a machine-readable way, does not apply to systems that perform "an assistive function for standard editing" or do not substantially alter what they are given; for generative systems already on the market before August, the Digital Omnibus, published in the Official Journal on 24 July 2026, gave providers until 2 December 2026 to start marking. Its deployer rule for text on matters of public interest lifts the disclosure duty where a human has reviewed the text and someone holds editorial responsibility. The Commission's guidelines of 20 July 2026, which interpret the rule rather than add to it, make that exception demanding: the review must be a substantive examination of the content, including fact-checking, a cursory or automated check does not count, significant AI changes after sign-off cancel it, and the person who holds editorial responsibility must be publicly identifiable. The Commission also published a voluntary Code of Practice on marking and labelling AI-generated content on 10 June 2026, with one section for providers and one for deployers. The law also marks provenance, so the contrast is one of emphasis rather than kind. But for text on matters of public interest, the editorial-review exception recognizes a dimension of responsibility that most simple platform labels leave out.

It would be a mistake to treat either as a measure of quality. A publisher can hold editorial responsibility and still do poor editorial work, and a coarse platform label can still tell a listener something true about how a track was made. The platform records provenance; the law organizes disclosure and accountability. Whether the human judgment behind a piece was any good remains a separate question that neither answers.

Neither line is wrong. Platforms fighting fraud at the scale of 90,000 tracks a day need something that runs automatically, and the presence of a tool can be checked automatically. Responsibility cannot. On my reading, that is the plain reason the platform labels are coarse: the measurement that scales is the one that gets built. A label makes a flood of work legible to a system that cannot read it, and what the simplification drops is whatever cannot be counted. But the consequence is worth stating carefully. A classification of how something was produced now sits upstream of decisions about distribution and payment, before anyone has read or heard the work.

Whether platform labels move judgment in the way the experimental labels did is a separate question, still untested as far as I can find. That link is my hypothesis. What is documented is that labels of this kind move judgment in controlled settings, that the platform labels exist, and that they feed into who gets recommended and paid. If studies of real platform labels show they change nothing about how people judge or play the work, the psychological half of this concern is wrong; what the labels do to distribution and pay would still stand.

The Runner, the Car and the Training Data

The strongest objection to everything above goes further than the claim that AI work is bad: a lower rating after the label may be correct.

When a reader admires a poem, part of what they admire is that a person found those words. A runner who covers a hundred meters in ten seconds has done something extraordinary. A car that does the same has done nothing worth remarking. The clock shows the same number for both, and the number means something different in each case. If a reader learns that a machine was involved, the achievement they thought they were admiring has changed, and adjusting the rating is an honest update rather than a bias.

I accept that, and I think it makes the problem sharper. If readers value the human part of a work, they have a stake in knowing how large it is. A poem produced in seconds from a prompt and an essay built over weeks, with a model doing the reading and drafting and a person doing the choosing, checking and answering for it, are different performances. The label most platforms use treats them as one.

The second objection is about consent, and it stands apart from quality entirely. The models were trained on work by people who were not asked. The major record companies sued Suno and Udio, the two best-known music generators, in June 2024. Universal settled with Udio in October 2025 and Warner with Suno in November 2025; both settlements came with licensing deals for new models trained on licensed music. Universal and Sony were still in court with Suno as of the most recent filings I could find. The model I write with has the same history. On 20 July 2026 a federal court in California gave final approval to a $1.5 billion settlement between Anthropic, which makes Claude, and authors who said pirated copies of their books had been used to train it: about $3,000 per work, for roughly 482,000 works. An earlier ruling in the case had held that using the books to train the model was fair use, and so was scanning print books Anthropic had bought, but that downloading pirated copies to build a permanent central library was not. I write with a tool whose maker paid for that, and I am not going to pretend the question is somebody else's.

Put the objections next to the studies and the public argument becomes easier to read. People are answering six different questions as if they were one: whether the work is good, who made it, how it was made, whether the training was lawful, who answers for the result, and who decides how widely it travels. Someone can admire an AI-assisted song and object to how its model was trained. Someone can accept AI entirely and still want streaming fraud stopped. A single tag is being asked to carry all six questions, and it can only answer one of them.

Where Judgment Pays: 758 Consultants

My own starting belief was that hybrid work, with someone who knows a field steering a model, can be better than what most people produce alone. The research supports a narrower claim, and the narrower claim is the more interesting one.

Anil Doshi and Oliver Hauser asked 300 people to write very short stories, some with ideas from a language model and some without, and had 600 others judge them; the study appeared in Science Advances in July 2024. Writers with access to up to five AI ideas produced stories judged 8.1 percent more novel and 9 percent more useful. The least creative writers gained the most, enough to close the gap with the most creative. But the stories written with a single AI idea were 10.7 percent more similar to one another than those written without. Each story improved while the collection narrowed.

Fabrizio Dell'Acqua and colleagues from Harvard Business School, Wharton, MIT and other schools, working with the Boston Consulting Group, put the same question to 758 consultants doing realistic work, in a working paper from September 2023. On tasks inside what the model could handle, consultants using GPT-4 produced work rated more than 40 percent higher in quality. On one task deliberately placed just outside that boundary, they reached the right answer about 60 to 70 percent of the time, against 84.5 percent for consultants working without it. The tool made good work better where it was competent and made people worse where it was not, because they trusted it past the point where it deserved trust.

Together those results locate the human contribution: in knowing where the model is reliable, in catching it where it fails, and in refusing the drift toward sameness. Consultants are not essayists, and neither study tests a method like mine, so I read them as describing a general mechanism rather than validating my own rules. Still, the rules were built against exactly these risks. The independent check is there for the task outside the boundary, the rhythm ceiling for the narrowing collection, and the gate before writing so that a subject with no mechanism is dropped instead of being made fluent. A control process proves that checking happens, not that the checking is always enough.

There is a version of the same risk that no ceiling catches. If every subject goes through the same method, readers may learn the shape of the argument before the finding arrives, and a method can start producing its own conclusions. The test I try to hold to is whether a piece would have ended differently if the evidence had come out differently. This one did change on the way: its central claim is narrower now than when the research began.

1982

Disputes over what a new tool does to the person using it are not new, and the closest precedent is a fight inside a musicians' union.

In January 1982 Barry Manilow toured Britain with synthesizers playing the parts that orchestral musicians would once have been hired for. That spring Britain's Musicians' Union passed a motion, later backed by a resolution of its executive, to ban synthesizers, drum machines and other electronic devices "capable of recreating the sounds of conventional musical instruments." It was barely enforced. Synthesizer players, many of them members, formed a union of their own. In 1997 the Musicians' Union admitted DJs, musicians whose instrument is other people's recordings, and the ban went with that decision.

The union was defending real jobs, and it was not foolish to do so. The instructive part is who got to decide which technical acts still counted as musical work. In 1982 it was a union motion; by 1997 the same union had moved its definition far enough to admit people who play records. The comparison has a limit that matters: a synthesizer makes no musical decisions, and a generative model makes a great many. That is why the question of what the human contributed is harder now than it was then, and why a label that records only the presence of the tool answers less of it.

What a Label Would Have to Measure

It is possible to record more. Spotify's AI credits let artists state where AI was used, by role or instrument, and Spotify says it receives tens of thousands of them a day. That is the beginning of a record of the human part rather than only the machine's. It depends on makers telling the truth and on someone reading the credits, and it is the exception. Most labels in use stop at the coarser question.

For text on matters of public interest, Article 50 points at the finer one. It does not ask how much of a text a machine wrote. It asks whether a person reviewed it, checked it, and stands publicly behind its publication. That test covers text published to inform the public; songs and other audio fall under a separate duty with no editorial exception. My reading, which the studies suggest but do not prove, is that this is close to what readers care about when they mark down the label: less the tool than whether anyone stands behind the work.

A detector can be very good at the first question. No detector can answer the second. You can detect the use of AI in a text. You cannot detect responsibility.

So the platforms and the law are building two different instruments at the same moment, while the meaning of authorship is itself unsettled. Is the human part the writing of each sentence, the idea, the choice among fifty versions, the check of every figure, the decision to publish, or the willingness to answer for what was published? I don't think there is one answer for every kind of work. I do think the platforms are choosing one in practice, because their labels already decide what is seen.

For this archive the human part is the questions, the method and the choices, and what holds them together is the last item on that list. The method and the sources are on the page, and the mistakes, when someone finds them, are mine.

Evidence Map

Facts, interpretations, forecasts, and disconfirming signals.

Core claim. Platforms increasingly classify creative work by whether AI was used and tie that classification to distribution and pay, while the evidence shows that readers' judgments respond to what they are told about authorship, about how much AI did, and about the intention behind the work, which a tool-based classification records only coarsely. That readers also weigh who is responsible for the result is the author's reading, not something the studies test.

Evidence level. Facts: high. The experiments (Porter and Machery; Wu and Holmes; Li and colleagues; Doshi and Hauser; Dell'Acqua and colleagues), the platform rules (Deezer, Spotify, Medium) and the text of Article 50 are documented. Interpretation: medium. That platform labels shift judgment the way experimental labels do is the author's hypothesis. Forecast: speculative. That the label will become less informative about quality as hybrid work spreads, while its consequences grow.

What would confirm this. Platform data or field studies showing that AI-labelled work by verified human makers is played, recommended or read less than comparable unlabelled work; more platforms adopting tool-based labels without a category for human editorial responsibility.

What would disprove this. The claim has three parts, and each fails separately. That labels change how creative work is judged: robust studies finding no meaningful effect for relevant work and audiences. That platform classifications affect visibility and pay: evidence that the classification has no material effect on recommendations or payouts in practice. That simple AI labels under-represent the human contribution: a widely used classification that reliably records the nature and extent of human involvement, as Spotify's AI credits begin to.

Watchlist. Through October 2027: how regulators apply the Commission's July 2026 Article 50 guidelines, which require substantive human review including fact-checking and a publicly identifiable person holding editorial responsibility; the outcome of the remaining music cases against Suno; Deezer's next figures on AI share of uploads against share of streams; any change to Medium's AI policy.

Sources. Brian Porter and Edouard Machery, AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably (Scientific Reports, 2024); Sarah Wu and Kevin Holmes, Is there a "mind" behind the music? (Cognitive Research: Principles and Implications, 2026); Zhuoyan Li and colleagues, How Does the Disclosure of AI Assistance Affect the Perceptions of Writing? (EMNLP 2024); the text of Article 50 of the EU AI Act; Deezer's July 2026 report on AI uploads; and Medium's AI content policy.

Jerry van der Laan writes The Manifest Archive, Forensic Narrative Intelligence Writing on the systems beneath power, money, and history. He traces the structures beneath them.

The questions, the editorial judgments and the responsibility for every published claim are mine. AI assists with the research, the drafting and the source checks.