AI Ate the Internet. Now We Want It to Decide What’s Human.

Watermarking is supposed to make synthetic content transparent. It may also turn AI companies into the institutions we ask to certify human authorship.

When I was a kid, I wrote my school essays by hand and then typed them out on our typewriter.

I was supposed to do the typing myself. It was useful to learn. But sometimes it got late, the essay was already written, I was exhausted, and my mother would take over so I could go to bed. She was a single mother with better things to do than supervise an overtired child ceremonially pressing every remaining key at 11 p.m.

Nobody thought she had therefore written my essay. Nobody called it plagiarism, cheating, or a fake assignment. We understood intuitively that there was a difference between the capability being assessed and the infrastructure used to produce the final artifact. The ideas, argument, language and understanding were mine. Typing had some value as a skill, but it was not the point of the essay.

Then we got a computer. This was extremely exciting, not least because the same machine on which I could suddenly write my essays also let me play Gorillas, the wonderfully stupid game where two gorillas threw exploding bananas across rooftops.

Now the computer provided the typing infrastructure directly. My mother became the spellchecker, reader, and person I bounced ideas off. Then software became the spellchecker. Grammar software started correcting sentences. Google became part of the research process. Autocomplete began predicting what I was trying to write.

Nobody demanded a receipt documenting exactly which words had been touched by my mother, Microsoft Word, or Google.

And then AI arrived.

Why did we decide this is where infrastructure becomes authorship?

I use AI for some combination of all those previous roles: typist, researcher, editor, critic, spellchecker, and brainstorming partner. Some of that is infrastructure. Some is intellectual assistance. Sometimes it is genuine intellectual labor.

But intellectual labor and authorship have never been synonyms. My mother contributed intellectually when I bounced ideas off her. An editor can substantially improve an argument. A researcher can discover the fact on which an essay depends. None automatically becomes the author.

Authorship is closer to where the governing intellectual agency sits: who decides what the work is about, chooses between ideas, understands the argument, determines what belongs in it, and takes responsibility for the result.

Years ago, we experimented at CanopyLAB with using artificial intelligence to assess student work in almost the opposite way from what watermarking does today. A single grade was too crude. We wanted to distinguish between things like understanding, argument strength, coherence, grammar, and style, then use those distinctions to give better qualitative feedback.

At the time, the technology made that difficult. Then generative AI arrived, and much of the technical problem became almost irrelevant. Modern models can look at a piece of writing and separate dimensions of performance in seconds. So I do not believe the problem is that we could never know what Claude contributed. We probably could know much more.

Anthropic says future Claude models will watermark text as part of its response to EU transparency requirements. Today, that watermark can indicate that Claude was involved without necessarily distinguishing between Claude generating something and Claude heavily editing it. But there is no reason to assume provenance has to remain that crude forever.

Perhaps the future receipt is itemized. Claude proposed the structure. Claude found the counterargument. Claude rewrote four paragraphs. Claude corrected the grammar. The human rejected six suggestions, verified the sources, rewrote half the generated text, and made the final decisions.

Fine. We now have the full ingredient list. We still have not answered who cooked the meal. That is the distinction I care about. Better measurement of contribution does not automatically give us a theory of authorship.

The Ministry of Authenticity

Anthropic does not claim to decide who authored a piece of work. The more interesting reality is that everyone else seems to cast it in that role.

A university wants to know whether a student used Claude. A publisher wants to enforce an AI policy. An employer wants to check an application. A literary prize decides how much AI assistance it will tolerate. Soon there is an API capable of answering a question those institutions cannot answer themselves: was Claude likely here?

Nobody at Anthropic has to announce a Ministry of Authenticity. Yet we are building one around them ourselves. And this is where the consequences become more interesting than whether the watermark works.

Imagine a student writes an essay, uses Claude to challenge the argument and tighten the language, and the watermark correctly identifies Claude’s involvement. The detector has not made a mistake. But a school can still make one if it treats that accurate signal as proof that the student did not do the intellectual work.

The same could happen to a writer who uses AI as an editor, a job applicant who uses it to improve awkward phrasing, or a researcher writing in a second language who uses it to polish a grant application.

The most dangerous version of this system is not one that falsely detects AI. It is one that correctly detects AI participation and then draws a completely unjustified conclusion about the human.

And incentives do what incentives do. If possessing the receipt becomes punishable, people will learn to destroy the receipt. The people using AI openly and legitimately become the easiest to identify, while those actually trying to conceal its role acquire every reason to remove the watermark, “humanize” the text, or move to tools that leave no trace.

We may therefore introduce provenance to create transparency and end up incentivizing concealment instead. The technology can work exactly as intended. The problem begins when we reward and punish people based on what we decide the receipt means.

Infrastructure is not competence

Education makes the distinction obvious. If I want to test mental arithmetic, take away the calculator. If I want to test typing, my mother finishing my typing defeats the purpose. But if I want to know whether somebody can analyze evidence, construct an argument, and defend a conclusion, assess that.

Ask where the model was wrong. Ask what the student accepted and rejected. Ask why they chose one argument over another. Ask them to defend what they submitted and take responsibility for it.

“Did AI touch this?” is easier to measure. It is also a different question.

Before AI, the finished artifact was a reasonably useful proxy for the mind that produced it. A student handed in an essay, and we inferred something about their understanding from the essay. AI breaks that proxy. But the answer does not have to be rebuilding it through increasingly sophisticated forensic analysis of the artifact. We can assess the mind more directly.

And we already know we can do better. The whole promise of using AI in assessment was that we could move beyond crude signals and understand more about the human: what they understood, where their reasoning was strong, what they struggled with, and what they should work on next.

The point was empowerment through richer feedback. It would be a strange reversal if considerably more capable AI now helped us reduce that same human performance to a forensic label about which tool participated.

Tools exist because opportunity cost exists. I could calculate a spreadsheet manually, but Excel exists. I could search twenty books for a statistic I can now find in minutes. I could spend three hours mechanically cutting repetition from an article, or spend some of those hours developing the next argument.

The fact that I could perform the lower-value task myself does not make performing it myself intellectually superior.

Much technological progress consists of turning expensive or time-consuming tasks into infrastructure and freeing human attention for something else. AI may be the next layer of intellectual infrastructure.

That does not mean understanding becomes less important. I think the opposite may be true. The more execution we can outsource, the more important judgment becomes. If a machine can produce a plausible answer in seconds, knowing whether the answer is good, bad, fabricated, derivative, or irrelevant becomes considerably more valuable.

The skill moves. Our assessment should move with it.

And somehow the machine gets the better receipt

Generative AI became useful by learning from an enormous environment of human intellectual production. Writers wrote. Researchers researched. Programmers coded. Teachers explained. Millions of people created the information environment from which these systems became valuable.

There are unresolved legal questions about training data, and that is not my argument here. I am interested in the direction of attribution.

Human knowledge flows into AI systems at enormous scale, and individual contribution becomes diffuse. Then the machine participates in our next piece of work, and suddenly its contribution becomes important enough to make technically traceable.

Humanity supplies the intellectual raw material. The machine gets the receipt.

Make that receipt as perfect as you like. Tell me exactly which sentence Claude proposed, which source it found, which paragraph it restructured, and which comma it moved. I still do not think machine provenance should become the proxy by which someone else decides whether the work is mine.

We understood the distinction when an exhausted child went to bed, and her mother finished typing an essay that was already written. Nobody needed a forensic record of how many keys my mother pressed or which spelling corrections came from her. We used judgment because we understood what the assignment was actually supposed to measure.

Perhaps we should remember that before building an authenticity infrastructure around the easiest thing to detect.

The receipt tells us the machine was there. It still cannot tell us who thought.