The mark tells you a tool was used. It can't tell you whether anyone was thinking.
In August, Anthropic confirmed that text from its Claude models now carries an invisible "watermark," a subtle bias in word choice that a detector can read in aggregate. It was built to satisfy the EU AI Act, and Anthropic applied it everywhere rather than only where the law required. A lot of people are furious.
I'm not. Go ahead and mark it.
When I take a photograph, the camera sets the exposure. Mark that too if you like. The camera handles the light. I still decide what to point it at, and what I point it at is the photograph.
Do I use AI to write and communicate? Absolutely. The text is free. The ideas are not.
My father was dean of admissions at Swarthmore, then Stanford, then Princeton. He was chairman of the College Board. He was well known and respected in his field, and people loved to hear him speak. When he died, his house was so full of books that you walked through them on pathways, and the books were read and marked up, not shelved for show.
He never wrote one.
Maybe his mind wasn't built for organizing everything he was thinking into long, structured prose. Maybe he was a little afraid of it. Writing something down for other people takes a long time and carries weight, and it opens you to criticism in a way talking doesn't. I don't know which it was. I think I inherited it either way.
He was not the obvious candidate for any of those jobs. His parents hadn't gone to college. He was a self-described high school troublemaker who went into the military, came out, worked nights at the post office, and finished college in three years. His high school principal had written to the local college, or so the family story goes, telling them under no circumstances to consider admitting him (although they did).
So I think he spent his career looking past the record, because his own record had been wrong about him. He was among the first to put creative prompts on an application. One of the early Stanford questions I remember was some version of: if you were stranded on a desert island, what books would you bring and why? He seemed to be seeking a window the transcript couldn't provide. I learned a lot from that outside-in view of the machine, and I can't think about any of this without him in the room.
Working with an AI writing partner has dramatically increased my ability to communicate things I care about and, honestly, things I think are worth communicating. If a watermark shows up on this essay, it isn't marking fraud. It's marking help.
Here's what the watermark is. Claude is nudged, at the level of individual word choices, toward patterns a detector can recognize across a long enough passage. No identity is attached; it can't be traced to a person or a conversation. Its strength tracks how much text the model generated, not how much thinking you did: a heavy revision of your own argument can mark clearly, while a thin prompt that produces something you never thought about marks the same. A free tool that strips it appeared on GitHub within a day of the announcement, and the familiar arms race began. OpenAI built the same thing in 2024 and shelved it, partly because it was easy to evade and partly because it fell hardest on people writing in a second language.
There are good reasons for the rule that produced it. The harms the AI Act has in view are real: synthetic media in elections, impersonation, fraud, fabricated imagery of people who never consented. The law even names the category it cares about most, deepfakes and AI-generated text published to inform the public on matters of public interest. In those cases provenance is exactly the right question, and I want it answered as much as the next guy.
This essay is not that category, and neither is a huge amount of writing the mark will actually land on.
The strongest defense of the mark is that it's harmless: non-identifying, low-stakes, and everyone will be doing it within a year anyway. I grant all of it.
It's harmless because it's empty. It answers one question, did a machine touch this text, and that question has no relationship to the one anyone actually cares about, which is whether there's thinking in it. A mark calibrated to the volume of generated text carries no information about that difference. It's provenance theater: a real signal about an irrelevant question.
Schools got here first. Detection software, oral defenses, a return to blue books. The stated goal is academic integrity. What the apparatus protects is the sorting signal, the grade and the credential, because a polished essay used to be costly enough to stand in for competence and now it isn't.
And the apparatus cannot tell a cheater from a non-native speaker removing friction, or from a person who thinks better than they write, or from someone with processing differences using the tool as scaffolding. Maybe my father. Certainly me.
Ask a room of teachers to describe a time they truly learned something, and the answers converge: someone took an interest in me, someone trusted me, someone pushed me past what I thought I could do, I had a question that was actually mine. Nobody describes a proctor. We know what learning requires. We built something else, and now we're defending it.
None of this is new, or particular to AI. It's what tools do. Combustion, the telephone, television, the computer. Each one extended what a person could do, and each one cost something, to the people who used it and to the people who had built a life on the older way. That's not a complaint about technology. It's the description of it.
The Amish are worth a look here, not because they refuse, but because they don't. They ask, before adopting anything, what it will do to the community, and they let the answer govern. A phone in a shed at the end of the lane, not in the house. The remarkable thing isn't the restraint. It's that they ask first. The rest of us ask afterward, usually in the form of a panic, about technologies we otherwise experience as weather. Adoption is inevitable not just because tools are irresistible, but because that kind of deliberateness is rare.
And every tool creates a new population of people who can now do the thing. That population is larger than the old one and made of different people; the ones who are good with the tool were not the ones who were good at the output before it. That's a power shift, and the leverage is the whole reason tools transform anything. The cost is real. The iceman goes out of business when refrigeration arrives. The Knowledge, years of London streets in a cabbie's head, devalues the day the phone can route a stranger.
Writing is messier than ice. It bundled production with thinking, so some real cognitive skill devalues alongside the mechanical part. There are writers whose ideas come from wrestling with their sentences, and we don't yet know how much of that we're trading away.
But notice where the detection impulse comes from. Not from readers who can't tell good ideas from bad. From the displaced: people whose expertise was, in part, the expensive production of fluent text, and who correctly sense that the expense is gone. The watermark restores a mark of distinction. I did it the hard way. That's a human thing to want. It's also not an epistemic claim. It tells you who paid the old price, not whether the ideas are any good.
Writing itself was the first tool to extend thinking directly, and the first recorded reaction to it is the watermark reaction. In the Phaedrus, Socrates has the king Thamus tell the inventor of letters that writing will produce forgetfulness, that people will seem to know things they don't, that they'll have the appearance of wisdom without the reality.
He was right. Oral memory degraded. The bards were displaced. And the verdict went the other way anyway. Nobody watermarks a book to prove a pen was involved.
AI is the first tool since writing to extend composition itself, rather than storage or calculation, and that is new, and I think it's what frightens people. But I think writing was the tool my father had and couldn't use. This one might have reached him.
The current response was inevitable. Performance artifacts were expensive enough to serve as proxies for competence. The cost collapsed. Any individual or institution whose legitimacy rests on a proxy will try to restore its scarcity, because the alternative is admitting the proxy never measured what it claimed. A regulation lands, the labs comply, the strip tools appear within twenty-four hours. The loop has begun.
We are reluctant to govern the tool, so we're governing the mark it leaves.
LLM writing will be absorbed, the way calculators and spell-check and cameras were, and not all of it will be good. Some people will outsource the ideas along with the text. Some fluency will be hollow. No watermark catches that, because the failure is in the thinking, not the provenance. What would help is structural and slower: make unsupported claims expensive, reward verification, keep the reader's own judgment in the loop.
The photographic image is what it is regardless of the automatic settings. Mark it. Then look at the picture and tell me whether it's any good.
