How many of you have used ChatGPT to write something this week?
I'm sure most of you have, even if it was only to fix the grammar on an email or make a message to your manager sound a little less blunt. (I do it too.)
Now think about what happens if every one of those could be traced back to the AI that helped write it.
The student who ran her assignment through ChatGPT the night before submission, and the employee whose client email went through Claude before it went out, would both be carrying an invisible mark that their college or company could check for.
That is already starting to happen in Europe, and it's what I want to write about today.
On 5 October, OpenAI announced it would start watermarking eligible ChatGPT and Codex text in the European Union over the coming weeks, because the EU AI Act requires AI-generated text to be identifiable by a machine, while outside Europe it is only an opt-in setting for businesses using its API.
Anthropic had already gone further in August, putting an invisible watermark into the text newer Claude models write in every country, India included.
Ask ChatGPT the same question from Delhi and the answer carries no mark.
I'm more interested in why so many people are this upset about being marked at all.
My opinion: I'd rather India were marked too.
For those who haven't come across this before, the watermark doesn't change your spelling or hide characters in the text.
It works while the model is writing, because the model picks each next word, or piece of a word, from a list of likely options, and OpenAI's method, which it calls textGrain, uses a secret key to steer those picks into a pattern its detector can find.
Think of a die that's been weighted a little, where two rolls look completely normal and 400 rolls show the lean.
That is why length matters so much.
The mark carries nothing about who you are, only a signal that an OpenAI model wrote or processed part of the text, and for now only approved researchers and expert organisations can check for it.
In 2024, the Wall Street Journal reported that OpenAI had a watermark its own documents called 99.9% effective when enough text was generated, and that it had been ready to release for about a year.
One reason it stayed on the shelf was a survey in which nearly 30% of users said they would use ChatGPT less if it carried a watermark and a rival didn't.
In August, that's roughly what happened to Anthropic, when within two weeks some people paying at least $100 a month said they had cancelled Claude (Anthropic says it hasn't seen a rise in cancellations).
Some of them used Claude to proofread or translate their own writing, and worried that writing could now carry a mark too.
I think a lot of the anger is plain embarrassment about being seen to need help.
The same research found the penalty was weaker when the person doing the judging uses AI a lot themselves, and in a hiring test those frequent users even favoured candidates who used AI.
By that logic India should be the easiest place in the world to be open about it, because 92% of knowledge workers here said they used AI at work in 2024, against 75% globally.
It hasn't worked out that way, and most of us will happily list ChatGPT as a skill on our CV and still remove any sign of it from the email we send our manager.
I get some version of this question from students almost every week, whether running an assignment through ChatGPT "just for the grammar" still counts as their own work, and most of them ask it after class, as if the question itself were an admission.
A consultant who asks Claude to rework three clumsy sentences may get a mark in the words Claude chose, though Anthropic says a handful of edits might be too few to register, and that the mark can't tell "Claude wrote this" from "Claude heavily edited this".
Its help pages put it this way: a detected mark is "a signal that content was processed by Claude, but is not fully conclusive."
OpenAI also worried back in 2024 that a mark could stigmatise AI as a writing tool for non-native English speakers.
India's new IT rules, in force since February, require labels on realistic AI-made audio, images and video, deepfakes included, while plain AI-written text was left out, apart from the usual rules on unlawful content.
So most Indian colleges and offices that want to check still only have detectors that guess from style.
If English is your second language, you lean on common words and simple sentences because that's the English you're sure of, and the detector reads that careful English as machine-made.
A watermark looks for a pattern tied to a secret key and ignores writing style, which I think makes it a much fairer test for anyone writing in a second language, even though it can still miss marked text and, rarely, flag text that was never marked.
And where the line falls is a choice each company made, because Anthropic marks Claude everywhere, saying it doesn't yet have a durable way to limit the mark by region, while OpenAI marks ChatGPT only in Europe and leaves businesses elsewhere to opt in through its API.
What worries me is the student in Indore or Coimbatore who writes her own essay in careful English, gets flagged by a detector, and has nothing to point to.
A watermark wouldn't prove she wrote it, and OpenAI says plainly that a missing mark doesn't prove a human wrote the text, but it would stop a guess about her grammar from being treated as evidence.
Some of the people cancelling worry that even a mark on work they only edited could raise questions about who wrote it, and I understand that, though I'd rather we stopped treating help as something to hide.
So tell me: if your company or college could check every document for a mark tomorrow, would you want it switched on?
Reply and tell me. I read every one.
Nandini
PS: In Europe, a document with no mark at all may soon raise its own question, whether the person wrote it alone or scrubbed it first. I can't tell yet which one people will assume.