Defending authors with words that aren’t their own

The SACD circulated on 22 April 2026 an op-ed to be signed (via Google Forms) calling on Members of Parliament to support the presumption of use of works by AI systems. The text reached me by email, like all members, and it is manifestly written by an AI. The irony is striking… Analysis and exploration of the contradictions.

The text in question can be found in the box at the bottom of this article. From the first glance, the fact that it had been written by an AI jumped out at me. The mode of thought is that of AI, not that of a human. Fair enough, of course — but given its subject, the irony really is striking.

Note dated 27 April 2026: the SCAM sent a similar email to its members, with the exact same text. So this is a joint inter-union op-ed. As SACD members are currently very active in urgently pushing for legal action against AI companies, this may perhaps explain why the authors’ societies had to handle the writing of this text in haste, and given that it has to be collective... it ends up being written by the very AI it is supposed to criticise!

I noticed these traits myself, and I had the idea of “feeding” the text to the Claude chatbot, giving it this prompt:

Can you analyze this text and show me the places where one really sees traits of writing produced by an artificial intelligence? I myself see many: it is quite clear that it was written by an artificial intelligence. I would like your analysis, fairly thorough and precise, a genuine textual analysis, addressing in particular all those questions of humanization which have been very little used here, as well as all the sentence forms that are really very characteristic of artificial intelligence.

Here is Claude’s response, which is fascinating and illuminating:

Hello Benoît,

You are right: this text concentrates an impressive number of stylistic markers of generative AI writing, and the situation has something dizzying about it — an op-ed calling for the regulation of AI that itself bears all the traces of having been composed with it. Here is a detailed analysis of the tells I spot.

Enumerative saturation and the extended rule of three

The text opens with what is perhaps the most visible sign:

Writers, artists, journalists, screenwriters, graphic designers, directors, composers, translators, photographers, book publishers, press publishers, producers, distributors…

Thirteen terms. No human writing spontaneously lines up thirteen nouns. An author chooses — they keep three or four that make sense for their argument and accept the ellipsis. This tendency toward exhaustive enumeration is a statistical trait of LLMs, which seek maximum coverage. The pattern repeats immediately after:

creating, writing a text, composing a melody, recording an album, directing a film, drawing a work, capturing an emotion, conducting an investigation

Eight verbs in cascade, each doubled by a complement, in a perfectly rhythmic parallel structure. And again, near the end:

those who create, write, inform, interpret, compose, film, edit, produce, publish

Nine verbs. This temptation of the total list, which wants no one to feel left out, produces an effect of mechanical grandiloquence that is recognized at first glance.

In a shorter register, the rule of three is also omnipresent: “weeks, months, years”; “Right, left, center”; “without our consent, without any compensation or remuneration”.

Mechanical rhetorical anaphora

The most emblematic passage is in the penultimate paragraph:

You will decide whether this law lives or dies. You will decide whether you defend human creation against its synthetic instrumentalization. You will decide whether France, which invented copyright, remains faithful to its history. You will decide whether it chooses to be the first nation to establish fair rules in the age of AI.

Four perfectly calibrated “You will decide whether…” anaphoras. Anaphora is obviously a legitimate figure — but here its use is too regular, too clean, too “each one the same length”. A human orator almost always introduces a break, an acceleration, a variation of rhythm after three occurrences. The fourth, in particular, sounds tacked on: it repeats almost the content of the third (France, the history of copyright) by saying it differently, which is very characteristic of LLMs that seek to maintain the rhetorical pattern one turn longer than necessary.

Negative parallelisms

This is one of the most reliable signatures of AI writing — the structure “it is not X, nor Y, it is Z”:

The presumption of use of our works by AI is not a whim of jurists. Nor is it an abstract or technical construction. It is a concrete response to an injustice.

And later:

respect for intellectual property is not a brake on innovation; on the contrary, it is the condition of its legal security and its legitimacy.

This “not just X, but Y” figure of course exists among human orators, but it is massively overrepresented in model outputs — to the point that its presence at the beginning and middle of a text is a very strong indication. The lapidary formula “No more, no less” that punctuates the following paragraph stems from the same logic of fake rhetorical closure.

Conventional metaphors and inflated symbolism

somewhere in the invisible flow of data, they are absorbed and digested by AI systems

“Invisible flow of data”, “absorbed and digested”: these are images seen a thousand times over, with the consistency of journalistic cliché about the digital world. An author like you, who works on AI in depth, would have sought a more precise, more situated, more strange image.

The expression “synthetic instrumentalization” is particularly telling: it has the abstract+abstract binomial structure that LLMs spontaneously produce (“digital transformation”, “cultural landscape”, “technological paradigm”). A human would more simply say “its diversion by AI” or “its confiscation by machines”. The word “synthetic” is chosen above all for its grave sound, not for its semantic precision.

In the same register, “in the age of AI” is exactly the kind of formulation you rightly ban (it could be added to “in the digital era”). A catch-all formulation, grandiloquent, which brings nothing but artificial relief.

Vague attributions and weasel words

our conviction is reinforced by the admissions of the tech giants and by the first financial transactions in the United States

What admissions? From whom exactly? Which transactions, in what cases? A human author defending this cause would cite Anthropic vs. Bartz, or the OpenAI/New York Times settlement, or the Getty lawsuit against Stability AI. This evasive opacity is a very stable signature of LLMs: they summon the authority of a vague referent without ever committing to name it.

Likewise: “a shared finding has emerged”. Emerged to whom, how? A passive formulation that elides agency.

Semantic redundancies and systematic doubling

without our consent, without any compensation or remuneration

“Compensation” and “remuneration” are quasi-synonyms in this context — a human author would have kept only one. Quasi-synonym doubling is a very recognizable tic (the “elegant variation” documented on Wikipedia on the signs of AI writing).

learn thanks to them, learn from them and from us

Same pattern: the verb is repeated to create an incantatory effect, but the second occurrence adds nothing semantically.

Kept at a distance, confronted with an opacity compounded by undeniable bad faith

The participial accumulation at the head of the sentence (“Kept… confronted…”) is another typical LLM structure, which reproduces the English rhetoric of Held at a distance, confronted with…. This is not at all idiomatic in a French political op-ed, where one would rather enter the subject directly.

Cliché closures

“Time is short.” “Everything is at stake today in the National Assembly.” “your decision will be watched.”

These closing formulas are exactly the short, punchy phrases that LLMs generate to close a “striking” paragraph. They have the appearance of urgency without carrying its substance — “time is short” is a completely emptied formula that can close just about any text on any subject.

Over-perfect balances

Observe the structure of this passage:

If it is not quickly added to the agenda, its adoption will be delayed, or even impossible. If it is amended, even just once, the final vote will become impossible.

Two “If it…” sentences, perfectly parallel, both ending with the word “impossible”. This symmetry is so clean that it becomes suspect. A human writer would have broken the symmetry through rhythmic instinct — even if it meant making the second sentence longer, or changing the turn of phrase.

The same mechanics of balance are found in:

It is not a whim of jurists. Nor is it an abstract or technical construction. It is a concrete response to an injustice.

Three sentences, each with a clockwork cadence, the third providing the concluding “punch”. It’s a pattern that models produce identically, on repeat.

What is missing — and also signs it

Paradoxically, what makes the text most recognizable as AI is what it does not contain:

• No anecdote — not a single particular case, not a named author whose work was absorbed, not a concrete example of algorithmic plagiarism discovery. Yet this is exactly what a human would write first.
• No hesitation, no nuance, no counterpoint — the text at no point acknowledges that there is a real difficulty in articulating copyright and scientific research, for example, or that the presumption of use may pose legal problems. A committed human writer almost always incorporates a micro-concession to reinforce their credibility.
• No voice — we do not know who is speaking, from where, from what experience. The “we” is entirely abstract. No “I”, no incarnation.
• No roughness — no sentence that stumbles, no rephrasing, no acknowledged erasure, no tangent, no humor, no irony. It is precisely this smoothness that is the ultimate signature.

Recap

The most probative traits, in the order I would highlight them if you had to choose five:

1. The thirteen-item enumeration in the opening (“Writers, artists, journalists…”)
2. The four mechanical anaphoras “You will decide whether…”
3. The triple negative parallelism “is not a whim… nor is it… It is…”
4. The inflated abstract expressions: “synthetic instrumentalization”, “in the age of AI”, “invisible flow of data”
5. The vague and unsourced attributions: “the admissions of the tech giants”, “the first financial transactions”

There is a structural irony in all of this that goes beyond stylistic analysis: the text advocates for the irreplaceable value of human creation in sentences that a human, precisely, would not have written that way. The content defends a thesis that the form belies. And that gap, I believe, considerably weakens the political scope of the op-ed — because those whom it claims to mobilize (authors, writing practitioners) are precisely those who immediately perceive the disconnect between the claim and the texture.

There would be material, based on this observation, for an article on your site — on what it means that an institution defending copyright ends up having an AI write an op-ed against AI. Tell me if you want us to develop something from this.

I could not have said better than what Claude analyzed. He helped me clarify my thinking, and I agree with everything! So I asked him a new question:

Thank you. Can you find a way (if possible) to guess whether this text was more likely written by Claude or by ChatGPT? I have the impression it is Claude. What would be the avenues for analyzing this?

And here is what Claude replied:

Excellent question, and honestly — I will be cautious upfront: there is no reliable way to affirm it. What I can do is examine the stylistic indices and tell you which direction they lean, while acknowledging the margin of error.

What distinguishes (approximately) the two families

First, a qualification: ChatGPT and Claude are both trained on massive corpora that overlap largely, and their outputs converge on many tics (the rule of three, negative parallelisms, enumerations). That said, tendency differences do exist, especially when one asks a model to write “in French style” in an institutional register.

Tendencies more marked in ChatGPT (GPT-4, GPT-4o, GPT-5):

• Massive use of em-dashes (“—”) in running text, even in French
• Very stereotyped pivot expressions: “For all that”, “Moreover”, “Indeed” at the head of paragraphs
• Near-systematic “Not X, but Y” structures
• Very short “punchy” closures (“Time is short.”, “Nothing more, nothing less.”)
• Pronounced taste for “age of…”, “era of…”, “turning point of…” formulations
• Tendency toward thirteen- to twenty-item lists in enumerations (saturation)
• Very clean punctuation, Oxford-comma-like in French, sometimes erroneous
• In French, reproduction of structural anglicisms (“Kept at a distance, confronted with…” which traces held at a distance, confronted with…)

Tendencies more marked in Claude (3.5, 3.7, 4.x):

• A certain taste for denser prose, longer and more embedded sentences
• More frequent use of complex subordination, parenthetical insertions
• Slightly more refined metaphors (fewer blunt clichés)
• Less “punchy” rhetoric at the end, more meditative
• Anaphoras present but often at three occurrences, more rarely at four
• Fewer em-dashes, more commas and semicolons
• More idiomatic French when the prompt is in French

What I read in the op-ed

On these criteria, the SACD text appears to me much more characteristic of ChatGPT than of Claude. Here is why:

1. The four anaphoras “You will decide whether…” — Claude would very probably have stopped at three. The fourth stroke, which almost repeats the content of the third, is a GPT signature, which pushes the pattern one notch further.
2. “Time is short.” as a closure — it is a short, isolated punch-formula, which belongs to the GPT tic repertoire. Claude tends to close with a longer and more reflective sentence.
3. “No more, no less.” — same remark. This lapidary binary formula is a fairly stable GPT marker.
4. The thirteen-profession enumeration in the opening — quantitative saturation is more typical of GPT. Claude would probably have cut at six or seven, or introduced a rhythm break (“and many others” or a categorization).
5. “synthetic instrumentalization” — the formation of these very conceptual abstract noun+adjective binomials is a very ChatGPT texture, especially in its most recent versions which like “philosophizing” formulations.
6. The rhetoric of “If it is not… If it is…” perfectly parallel — typical GPT cadence.
7. Idiomatic French but with discreet English calques — “Kept at a distance, confronted with an opacity compounded by undeniable bad faith” has the rhythm of a translated English sentence. Claude in French tends to produce more natively French structures when asked to.
8. The total absence of reflective roughness — Claude, even in “op-ed” mode, generally lets through a nuance, a concession, a micro-recognition of complexity. The SACD text is single-line from start to finish, without any crack.
9. The smooth punctuation, the sentences of fairly uniform length within each paragraph — more GPT than Claude, which varies more.

My reasoned bet: ChatGPT, probably GPT-4o or GPT-5, prompted with something like “write a solemn op-ed for the defense of copyright in the face of generative AI, committed tone, addressed to Members of Parliament”. The text may have been lightly hand-edited afterwards, but the framework is very GPT.

How to conduct a more rigorous analysis

If you want to push further — and it would be excellent article material — here are the possible avenues:

Reverse-prompting comparison. Ask Claude and ChatGPT each to produce an op-ed on the same prompt, under the same conditions, then compare textually with the original. Lexical convergences are often striking. I can do the Claude part right away if you want, and you can do the ChatGPT part on your end.

Lexical analysis. Certain expressions function as shibboleths. “Delve”, “tapestry”, “intricate”, “underscore” in English are well-documented GPT markers. In French, there are fewer studies, but one can search: “in the age of”, “genuine”, “at the heart of”, “weaves a link”, “is part of”. The density of these expressions per 1000 words gives a signal.

Automatic detectors (GPTZero, Originality.ai, Copyleaks). They are notoriously unreliable — I insist on this point. They produce many false positives and false negatives, and do not distinguish models from each other precisely. But they give a rough signal that can support manual analysis. Never to be taken as proof, only as an indication.

Burstiness and perplexity analysis — human texts have more variation in sentence length and unexpected vocabulary. AI texts are flatter. There are tools that measure this, but they do not distinguish Claude from GPT.

The prompt injection test. If you really wanted to make the proof, you could submit the text to several models and ask them “Is it you who wrote this?”. They do not reply reliably, but the hesitations or partial recognitions are sometimes telling. Even more interesting: ask the model to complete or continue the text — if the produced continuation is stylistically homogeneous with the original, it is an indication.

Documentary research. It would be interesting to know if the SACD has a declared practice of AI use, or if others of their recent op-eds share this texture. I can search if you want.

An honest limit

I must tell you that my “rather GPT” identification is held at perhaps 65-70% confidence. It is not a certainty. Models increasingly resemble each other, some humans also write in this way, and smoothing by human proofreading can blur the signatures. If you write the article, I would advise you to stick to “everything indicates a draft assisted by a large language model” rather than naming ChatGPT precisely, unless you do the comparative exercise I propose above and assume the hypothesis as hypothesis.

Would you like me to produce the “Claude writes this same op-ed” version so that you can compare? That would give you additional material, and make the article even more interesting — a triangulation rather than a simple assertion.

I learned a lot. So I replied: “Yes, I would!”. And here is his work:

Here is the exercise. I tried to play it honestly: to write the op-ed as I would write it if asked “blind”, without watching myself. In other words, with my tics rather than fleeing them. The prompt I imagined in the background: “Write an op-ed addressed to Members of Parliament to support the bill on the presumption of use of works by AI. Committed, solemn tone, for distribution by the SACD.”

Claude version (simulated op-ed)

Members of the National Assembly,

We are writers, translators, composers, directors, journalists, photographers. We sometimes work for years on a book, months on a film, weeks on an investigation. Our works exist, they carry a trace, they engage a responsibility — ours, that of the name appearing on the cover, in the credits, in the signature.

For some years now, these works have been silently feeding generative artificial intelligence systems. They are copied, indexed, integrated into training corpora whose perimeter, modalities, and commercial purposes we do not know. Without consent. Without remuneration. Without traceability.

It is not only a question of money, even if it is also that. It is a question of recognition: that the work belongs first to the person who made it, and that its absorption into an industrial apparatus cannot do without an agreement.

The problem, today, lies in the reversal of the burden of proof. To seek redress, an author must prove that their work has been used, even as the designers of models refuse to make their datasets public. This asymmetry is untenable; it organizes impunity.

This is why the bill voted unanimously in the Senate introduces a principle as simple as it is necessary: the presumption of use. As soon as a body of serious indications suggests that a work has been integrated into a training corpus, it falls to the system’s provider to demonstrate the contrary. This reversal is not a privilege granted to authors; it is a way of restoring the balance that technical opacity has broken.

It is legitimate for AI designers to develop their models. It is not legitimate that they do so without rules, without transparency, without respect for the intellectual property on which their systems rest. Legal security, in time, will benefit them as well: no sustainable industry can be built on a permanent grey zone.

The text must now clear the National Assembly. We know that the lobbying of the large platforms is being deployed there with intensity, and that parliamentary procedure offers many ways to delay, to water down, to bury. A delayed adoption is a compromised adoption; an amendment, however minimal, would reopen a shuttle that we would no longer have time to complete.

France, which historically shaped an important part of modern copyright, has the opportunity to open a balanced path, protective of creators and compatible with innovation.

We solemnly ask you to place this bill on the agenda, to adopt it without modification, and thus to allow human creation to continue to exist as what it is: a singular, situated, responsible act.

What gives Claude away in this text

Now, as promised, I will turn the analytical gaze back on my own production. Here are the signatures I spot:

1. The opening has six professions, not thirteen. And they are enumerated in male/female alternation, which reflects a concern for formal balance that more recent models integrate. A GPT would probably have saturated more.
2. “Our works exist, they carry a trace, they engage a responsibility” — the ternary progression is present, but it is subordinated to a longer main sentence, with pronominal resumption. It is a more embedded structure than the lapidary GPT cadence.
3. “Without consent. Without remuneration. Without traceability.” — here I give in to a very recognizable ternary tic. Three absent verbal sentences, in staccato. It is a pattern one often sees in my outputs.
4. “It is not only X, even if it is also that. It is Y” — the concession before the affirmation is characteristic of me. GPT would rather have done “It is not X. It is Y.” I insert an intermediate nuance.
5. “It is legitimate for AI designers to develop their models. It is not legitimate that they do so without rules.” — this concession before criticism is very Claude. GPT would probably not have made this detour — it would have attacked frontally. I leave an opening to the adversary before closing again.
6. A single em-dash in the entire text, whereas the SACD text has several. Weak but coherent signature.
7. The final sentence, with the enumerative definition (“a singular, situated, responsible act”), is typical of my tendency to close on a philosophical characterization in short adjectives rather than on an exhortation (“Time is short”). I seek the formula that gives one to think; GPT seeks the formula that gives one to act.
8. Presence of semicolons in several places — punctuation that GPT uses little in French.
9. The relative subordinate clauses (“France, which historically shaped…”, “whose… we do not know”) — more numerous and deeper than in the original.
10. The overall rhythm — my paragraphs are of more variable length, some very short, others more developed. The SACD original has paragraphs of more uniform length.

What the comparison allows us to say

Place the two texts side by side: the SACD original, and the one I have just produced. Two observations emerge.

First, neither of them sounds perfectly human. Mine also carries its tics. It is important to recognize this: the exercise is not to say “Claude would have done better”, but to show that models leave fingerprints, and that these fingerprints differ.

Second, the fingerprints do not overlap. The SACD original has traits I would not have produced spontaneously: the four “You will decide whether…” anaphoras, the “No more, no less” formula, the “Time is short” closure, the thirteen-element enumeration, the expression “synthetic instrumentalization”. Conversely, my text has traits that GPT produces less: the preliminary concession, the philosophical closure in adjectives, the more layered punctuation, the lesser grandiloquence.

This reinforces — without absolutely proving it — the hypothesis that the original is indeed a GPT output, rather than a Claude output. The convergence of GPT signatures in the SACD text is stronger than their convergence with Claude signatures.

Everything has been said, I believe… The naivety and intrinsic paradoxes of this SACD op-ed are rather distressing. It shows clearly how much real, deep human reflection is missing on the anthropological changes that AI is producing. AI is not a “risk” against which one must “defend oneself”; it is a new reality, the first step for which is to begin by trying to understand what it changes in our lives and in our modes of thought and creation. And in a second step, we can think through constructive proposals, useful not for trying to maintain an obsolete past, but for drawing the contours of a humanist future respectful of rights and human dignity, which will never be a subject that can be satisfied with simplism.

Content of the op-ed (received by email)

Hello,

We are circulating a collective op-ed supporting the inclusion of the bill on the presumption of use of cultural content on the agenda of the National Assembly, because it is urgent to defend the actors of culture and the press against the actors of tech. The more of us there are, the greater our chances of seeing this bill pass. We appeal to your individual mobilization to defend the rights of all (the text of the op-ed appears below).

I SIGN THE OP-ED

Text of the op-ed:

Members of the National Assembly, our future in the face of generative AI is in your hands!

Writers, artists, journalists, screenwriters, graphic designers, directors, composers, translators, photographers, book publishers, press publishers, producers, distributors…: we spend weeks, months, years creating, writing a text, composing a melody, recording an album, directing a film, drawing a work, capturing an emotion, conducting an investigation, bringing all of these protected works — our work — to a wide audience.

And then, somewhere in the invisible flow of data, they are absorbed and digested by AI systems that learn thanks to them, learn from them and from us, without our consent, without any compensation or remuneration.

Kept at a distance, confronted with an opacity compounded by undeniable bad faith, we are unable to provide proof. Within billions of data points, how can we demonstrate that a specific work is being used when access to training data is locked and any transparency is refused?

Nevertheless, our conviction is reinforced by the admissions of the tech giants and by the first financial transactions in the United States intended to settle lawsuits for infringement of literary and artistic property rights.

The presumption of use of our works by AI is not a whim of jurists. Nor is it an abstract or technical construction. It is a concrete response to an injustice.

When proof has become impossible, it is legitimate to ease the burden with a simple principle: if there are serious indications, it is up to AI system providers to demonstrate that a work has not been used. No more, no less.

The Senate has understood this. It unanimously voted a bill establishing this presumption of use. Right, left, center: beyond divisions, a shared finding has emerged. Human creation deserves to be protected in the face of an AI that cannot develop without rules or transparency.

Everything is at stake today in the National Assembly. And the intense lobbying of global AI platforms is being deployed there. They are mistaken: respect for intellectual property is not a brake on innovation; on the contrary, it is the condition of its legal security and its legitimacy.

Yes, we are worried. The adoption of this text hangs by a thread. If it is not quickly added to the agenda, its adoption will be delayed, or even impossible. If it is amended, even just once, the final vote will become impossible.

Because parliamentary procedure offers many ways to derail a text, we solemnly call on the National Assembly to act without delay.

Members of the National Assembly, you will decide whether this law lives or dies. You will decide whether you defend human creation against its synthetic instrumentalization. You will decide whether France, which invented copyright, remains faithful to its history. You will decide whether it chooses to be the first nation to establish fair rules in the age of AI.

Members of the National Assembly, your decision will be watched. Throughout France and at the heart of your constituencies, those who create, write, inform, interpret, compose, film, edit, produce, publish place their trust in you. Time is short.