AI watermarks and the college essay

Until this year, the only way to tell whether AI wrote an essay was a reader's sense that it sounded like a machine. In 2026 the three biggest AI companies started marking their text at the source, and one of them published how its mark works. None of this is a reason to panic, and the details are not what most people assume. But it changes the picture for every student who will write a college essay from here on, and most of all for this year's juniors, who apply in fall 2027.

What changed this year

The European Union's AI Act now requires AI companies to make generated text identifiable by machine, with obligations that began in August 2026. The companies responded differently:

• Anthropic (Claude) announced in August that its new Claude models watermark their text, with older models to follow, and that it applies the watermark worldwide rather than only in Europe because "we don't yet have a durable way to scope it by region."¹
• Google (Gemini) says text from its Gemini app and website carries its SynthID watermark.²
• OpenAI (ChatGPT) announced on October 5, 2026 that it will watermark ChatGPT text in the European Union over the following weeks, and is not making it a default elsewhere yet.³

How it actually works

Most talk about catching AI writing has focused on giveaways: hidden characters, the em-dash, phrases like "it's not X, it's Y." The watermarks work nothing like that.

Anthropic's explanation is the most detailed. Whenever Anthropic's Claude chooses between near-equivalent words, a secret key and the words before it decide the choice, instead of chance. A reader cannot see the difference. Someone holding the key can check whether a passage keeps landing on the choices the key predicts, and estimate how likely it is that the model wrote it. Anthropic says plainly that "there are no hidden characters," nothing is added to the text, and the watermark carries no information about who used it. Its method is adapted from Google's SynthID, and OpenAI describes its own as the same general idea: a statistical pattern in word choice.

So the old tricks do not apply. There is no character to delete and no phrase to avoid, because the mark is spread across hundreds of ordinary word choices.

What it means for an essay

Rewording it lightly does not reliably remove it. Anthropic says "light editing probably won't remove the watermark completely," while a complete rewrite will. OpenAI's own tests found the mark fading as more words changed: swapping a quarter of the words for synonyms cut detection from about 92% to about 17%.³ A student would have to rewrite most of it, at which point they may as well have written it.

"Polishing" with AI is where the line sits, and the watermark tracks it. Anthropic notes the watermark cannot tell text its model wrote from text its model heavily edited, but that light proofreading of a person's own writing leaves very little mark, because there are few word choices for the model to make. In practice: a student's own draft with a few spelling fixes looks like the student's. A draft handed to AI to "make it better" comes back as the AI's sentences, and carries its mark.

Very few people can check for it today. None of the three companies offers an open text checker. Anthropic's detection tool is in private preview for eligible groups that include educational organizations; OpenAI's is limited to approved researchers; Google's public checker covers images, video and audio, not text. We have not found any college that says it checks essays for watermarks. Anthropic and OpenAI both describe these limits as a starting point, so access is likely to widen.

A missing watermark proves nothing either way. OpenAI says outright that its absence "does not prove human authorship". Other AI tools exist, and many do not mark anything. A watermark is one piece of evidence, not a verdict.

What to expect in 2027

The students who feel this most are today's high school juniors, who write their essays in the summer and fall of 2027. By then, a few things are already set:

• Marking will be the norm, not the exception. The EU's deadline for AI tools that were already on the market to start marking their text is December 2, 2026.⁴ Anthropic already marks worldwide. OpenAI's marking is EU-only for now, so whether ChatGPT text written in the US carries a mark is the open question to watch.
• More people will be able to check. Anthropic says it plans to widen access to its detector over time, and educational organizations are already among those eligible. OpenAI calls its current access "initial".

And a few things we expect, though no one has announced them:

• More colleges will publish their own AI rules, spelling out what is allowed in the essay (brainstorming, grammar help) and what is not. Read them for every school on the list, because they will not all match.
• More weight on the parts of an application that are hard to outsource: interviews, graded schoolwork, writing done under supervision, and recommendations from teachers who watched the student work. Our essay guide covers how that shift is already starting.
• The essay itself will not go away. What changes is how much a reader trusts it on its own. An essay that matches everything else in the file is still the best way for a student to be heard.

The rule that matters

The Common App's fraud policy lists, among the forms of fraud, "intentionally misrepresenting as one's own original work... the substantive content or output of an artificial intelligence platform, technology, or algorithm."⁵ If it finds fraud, it may suspend or terminate the student's account and tell the colleges on their list. Applicants certify that their application is their own work when they submit it.

The word that does the work there is "substantive." Spelling and grammar help is not what the policy is about; sentences the student did not write are. Individual colleges draw their own lines on brainstorming and feedback, so anything beyond a spellcheck is worth checking against each school's admissions page.

Protect a student who did the work

The bigger risk for most honest students is being wrongly suspected, not caught. Older AI detectors, which guess from writing style rather than reading a watermark, have been known to flag real student writing as AI. A watermark check is a different and more specific test, but the best protection is the same for both:

• Write in a document that keeps version history, such as Google Docs, from the first rough notes onward.
• Keep the messy drafts. False starts, rewrites and cut paragraphs are the most convincing evidence that the writing is the writer's.
• Keep every sentence the student's own. That rule from the essay guide applies to AI and adults alike: a rough sentence in the student's voice beats a polished one in anyone else's.