OpenAI to Watermark ChatGPT Text in EU — India's Rules Go Further

OpenAI has built a way to quietly tag everything ChatGPT writes. For now, it's only switching that tag on for one part of the planet — and it isn't India.

What's actually changing

On October 5, OpenAI said it would start adding invisible watermarks to ChatGPT and Codex text for users in the European Union, under a system it calls textGrain. The idea is simple even if the engineering isn't: instead of stamping a visible tag on AI-written text, textGrain nudges the model's word choices in a subtle statistical pattern. A human reading the text won't notice anything odd, but a detector that knows what to look for can flag it as machine-written.

The rollout is tied directly to the EU's AI Act, the bloc's main law for regulating artificial intelligence, which requires generative AI providers to make AI-generated text identifiable in a machine-readable way. Systems already on the market have until December 2 to fall in line, which is why OpenAI is moving now rather than later.

Why only Europe gets it by default

Here's the catch: textGrain is landing as an EU-only default. Everywhere else, including the API that developers worldwide build on, the watermark stays off unless someone deliberately turns it on. OpenAI has been upfront that the technology isn't bulletproof.

"Text generated with OpenAI tools may be too short, edited, or translated for detection to work reliably."

That's a real limitation, not corporate hedging. Paraphrase a watermarked paragraph, translate it, or trim it down to a tweet-length snippet, and the statistical fingerprint can wash out. The detector itself isn't public either — OpenAI says it's limiting access to vetted researchers and institutions, presumably to stop bad actors from reverse-engineering ways around it. For context, Anthropic's Claude already applies text watermarking to supported models globally, not just where a specific law forces its hand, which makes OpenAI's EU-only approach look more like minimum-compliance than genuine commitment to the idea.

• textGrain shifts word-choice patterns to create a detectable signature, without changing how the text reads
• It's on by default for ChatGPT and Codex in the EU only, rolling out over the coming weeks
• The API keeps it off by default everywhere, including for EU developers, unless explicitly enabled
• Short text, heavy edits, and translation all degrade or erase the watermark

Where this leaves Indian users

This is where it gets interesting for readers here. India isn't waiting on OpenAI to decide whether Indian ChatGPT users deserve the same treatment — it already has its own rules on the books, and in some ways they go further than what the EU is asking for. MeitY (the Ministry of Electronics and Information Technology) amended the IT Rules earlier this year to bring what it calls "synthetically generated information" — its formal term for AI-generated or altered audio, video, and images, deepfakes included — under mandatory labelling requirements. Platforms that let people create or spread such content have to attach a prominent, visible label plus permanent provenance metadata that can't simply be stripped out, and unlawful flagged content now has to come down within three hours instead of the previous thirty-six. We covered the earlier version of this push in detail when the labelling and tracking framework was first proposed.

Notice the gap: India's rules are built around visible labels people can actually see, aimed squarely at images, audio and video where deepfakes do the most damage — ahead of a country that runs more WhatsApp forwards and election-season misinformation than almost anywhere else. OpenAI's textGrain, by contrast, is an invisible, text-only watermark that Indian ChatGPT users won't get by default at all, because it's scoped to the EU. For Indian developers building on OpenAI's API, nothing changes procedurally, but it's a reminder that "AI-generated content" compliance in India increasingly means satisfying MeitY's provenance rules yourself, rather than relying on a watermark baked in by the model provider.

The bigger picture

Watermarking AI text was always going to arrive unevenly, law by law, rather than as one global standard, and this is the first clear proof of that. A regulation-shaped patchwork means the same chatbot behaves differently depending on which passport your IP address holds, which is a strange place for a supposedly universal tool to land. Indian policymakers clearly aren't willing to wait for Silicon Valley to extend the courtesy on its own timeline — and given how often AI-written text and deepfakes already show up in Indian political and financial scams, that impatience looks increasingly justified.