By one count, the dots, lines and Y shapes recur beside animal images in at least 400 European caves, Lascaux and Chauvet and Altamira among them.1 The animals are pictures. The marks are not.
The horse took hours. The pigment had to be ground, the surface prepared, every line committed without erasure. The marks took seconds.
The six-minute horse
Consider a carver who grows tired of making a thousand different horses and settles on an outline — one fixed shape that means horse to everyone in the band. From then on the shape is traced and filled, and the effort goes into the part that changes: the count of moons beside it, the mark for birth. The six-hour horse becomes a six-minute horse.
Nothing about horses has been forgotten. The knowledge moves faster, reaches more people, and outlives the person who first knew it.
The carvers who had given their lives to the six-hour horse would have seen the outline as a cheapening, a horse nobody had wrestled out of the rock. They were right that something was lost and wrong about what mattered. The template won, as templates always do, because the people receiving the horse never cared how long it took to draw.
Twenty thousand years later, an industry exists to stamp the six-minute horse with a warning label.
They were right that something was lost and wrong about what mattered.
What the detectors are for
The consensus case for AI-text detection is reasonable. A teacher assigning an essay wants to know the student did the thinking. A publisher wants a byline to mean something. A regulator wants a reader to know whether the thing in front of them came out of a machine.
Only the third is a question about words.
Detection measures the provenance of the expression. Schools, employers and publishers are asking about the provenance of the thought. Those came apart the moment a model could phrase an argument it did not have, and no improvement in the instrument closes the gap, because the thought was never in the word choices to begin with.2
The marks were already templates
In 2023 three independent researchers, an engineer at UCL and two Durham academics, proposed a reading of those dots, lines and Y signs. In their account, the dots and lines count lunar months from bonne saison — the end of winter, when the rivers unfreeze and the ground greens — and the Y marks the month the animal gives birth. A phenological calendar of prey behavior, standardized enough to read across sites.1
Three archaeologists took it apart in the same journal. April Nowell, Paul Bahn and Jean-Loïc Le Quellec concluded that the lunar-calendar interpretation "suffers from multiple theoretical and methodological weaknesses preventing it from being a viable explanation for the production of Upper Palaeolithic art," and that claims to have found the oldest known proto-writing system "are unsubstantiated." Their central objection is the one that ends arguments: the readings are not testable.3
What survives the dispute is the part neither side contests. The marks are compact, repeated and reusable. The animals are not.
Every writing system that followed made the same move. Hieroglyphs turned drawings into standardized signs that carried a second reading by rule rather than by guess.4 Cuneiform reduced pictures to wedges a scribe could press in seconds. The alphabet collapsed thousands of signs into a few dozen. Each step cut the labor per idea and widened the circle of people who could carry ideas at all. Nobody demanded the stylus be disclosed.
Every reduction in labor had its mourners
The objection to AI-assisted writing was written down by its earliest opponent. In the Phaedrus, the Egyptian king Thamus refuses the god Theuth's gift of writing: the discovery "will create forgetfulness in the learners' souls, because they will not use their memories," and it offers disciples "not truth, but only the semblance of truth," leaving them "having the show of wisdom without the reality."5 The complaint survives for one reason. Someone wrote it down.
In 1492 the Benedictine abbot Johannes Trithemius wrote De laude scriptorum — In Praise of Scribes — telling his monks to keep copying books by hand. His argument was not that the press was bad. It was that the labor was holy. Of the brother who asks why he should wear himself out writing by hand now that printing exists, Trithemius wrote: "whoever says this is trying to conceal his own sloth."7
He called printing a marvelous art elsewhere in his work, and he had this treatise printed, two years after writing it.8
Which makes him the cleanest witness the position ever had.
He did not object to the machine. He objected to the loss of the effort, and he used the machine to say so.
Spell check, the word processor, autocomplete and grammar software each absorbed a layer of compositional labor with no disclosure regime and no detector.9 Each mourner was right about the craft and wrong about the knowledge. The craft was never the thing being transmitted.
The detector reads the sentence, not the thought
A person can spend three days working out an original argument and then use a model to organize and phrase it. Another can spend thirty minutes typing something entirely derivative by hand. The first contributed more intellectually by any honest measure. A detector flags the first and clears the second.
The detectors do not even read the sentence well. OpenAI launched its own AI Text Classifier on January 31, 2023. On the company's own challenge set it "correctly identifies 26% of AI-written text" and mislabels human writing as AI "9% of the time," and it refused inputs shorter than 1,000 characters because below that it was, in OpenAI's word, "very unreliable."1011 OpenAI withdrew it that July, "due to its low rate of accuracy."12 The company with the most knowledge of how its model writes could not reliably recognize the model's writing.
The failure is not random.
Stanford researchers ran 91 TOEFL essays by non-native English speakers through seven commercial detectors and recorded an average false-positive rate of 61.22%. Eighty-nine of the 91 — 97.80% — were flagged as AI-generated by at least one detector, and 18 were flagged by all seven. The same detectors were near-perfect on 88 essays written by American eighth-graders.13
The cause is mechanical.
Second-language writers draw on a narrower range of words, which produces lower-perplexity text, and perplexity is what the detectors score.13 The human who writes plainly gets accused. The human who lets a machine polish the prose walks through.
It is not a detector of artificial intelligence. It is a detector of a style, and the accusation lands on the people with the least margin to absorb it.
The apparatus evaluates the manufacturing of the sentence and never touches the truth of the proposition. It is verification overhead, the bill a society pays once it can no longer assume the thing it needs to know.15
The watermark answers a different question
The serious answer is already shipping, and it proves the same point from the other side.
Article 50(2) of the EU AI Act requires providers of systems generating synthetic text to "ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated." It applies from August 2, 2026.16 Anthropic marks the text output of Claude models launched since that date, applying it worldwide because it has no durable way to scope the mark by region, with older models to follow. It states that its method is a version of the SynthID-Text approach published by Google DeepMind.1718 Google has watermarked Gemini's text output since May 2024.19
A text watermark does not live in the characters. It lives in the choices between them. A model assigns a probability score to each candidate next word; the watermark adjusts those scores with a secret key, and across hundreds of words the adjustment accumulates into a statistical lean that only the key-holder can count.19
That repairs the failure of the style-guessers. A keyed test is not a guess about how machine prose sounds, so the writer choosing simple words has nothing to fear from it.
What it measures, with real precision, is whose words these are.
Read those together and the instrument is exact about the thing it measures and silent about the thing institutions want.
- The person who spent three days building an original argument and had the model tidy the paragraphs is barely marked.
- The person who had that same argument in their first language and asked for a translation is fully marked, every word.
- The person who lifted someone else's idea and retyped it in fresh words of their own is clean.
That is the same population the detectors punished, hit again by the better technology for the opposite mechanical reason.
Anthropic says what the mark cannot do. It "can only determine that Claude was likely involved with the content at some point. It cannot distinguish 'Claude wrote this' from 'Claude heavily edited this.'" The mark carries no identifying information and cannot be traced to a person, an organization or a chat.17
Important
Detection sits in private preview, and the eligibility list is not what the debate assumes. It covers regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups, plus enterprises carrying their own compliance obligations.20 Schools are on that list. An individual hiring manager is not.
Removal follows from the mechanism. Light editing dilutes the mark rather than deleting it; what erases it is re-composition that shares no runs of wording with the original.21 Avoiding the label means re-expressing the same idea in new words, which is exactly the labor the old proxy rewarded.
The cave dweller who wants an unlabeled horse has to draw it freehand again.
Labor was always a proxy
Effort was visible and thought was not, so effort stood in for thought. The five-page minimum, the handwritten draft, the billable hour and the word count all measure labor because labor could be counted and understanding could not. The proxy held for as long as producing a polished argument required both the thinking and the grinding. The polish now costs almost nothing, which exposes how much of the evaluation was weighing the grinding.22
The objection that lands
Sometimes the provenance of the expression is the thing being assessed. A language exam tests whether a person can produce English prose unaided; there the word choices are the object and not a proxy for anything, and a watermark measures exactly the right quantity. The same holds for a translation credential, and for a first-year writing course whose purpose is to build the capacity rather than to collect the output. Article 50 is not an authorship test either. It exists so a reader knows what they are reading, and for that purpose the mark is well designed and correctly aimed.
This argument also happens to excuse how this essay was made. It was written by a person and edited with a model, and a world that stopped auditing the provenance of sentences would be a more comfortable one for the person whose name is on it.
Where authorship genuinely matters, the reason dictates the evidence. An exam testing unaided composition proves itself through process: work done in the room, drafts with history, an oral defense of the argument.23 A byline is a claim of accountability, and accountability attaches to the person who signs regardless of the tools behind the page.
Fraud is established by evidence of deception, not by a probability score. None of those is served by a classifier reading finished prose. Every one of them is served by asking an author to stand behind the ideas.
The wall
The carver who objected to the outline was defending the same proxy. Six hours at the wall signaled devotion, skill and seriousness. The hunter reading the wall the next spring needed none of that.
He needed the count of moons beside the horse, and the mark that said when the herd would foal.
The marks gave him the same answer whether the outline beneath them took six hours or six minutes.