OpenAI will begin adding invisible statistical watermarks to ChatGPT and Codex outputs in the European Union starting this week — a compliance move under the EU AI Act, not a global rollout. The feature, named textGrain, alters word-selection probabilities in generated text without inserting visible markers or hidden characters. It is not enabled by default for API users outside the EU. They must opt in manually for select models. No Indian launch, pricing, or availability is mentioned in any source, and none is implied. This is a regulatory response, not a product update for consumers or developers in India or elsewhere. OpenAI says the watermark helps identify text likely generated by its models. But nothing more. Its detector, initially restricted to approved researchers and expert organisations, returns only a binary ‘yes’ or ‘no’ on whether an OpenAI pattern is present. It reveals no user identity, no prompt history, no timestamps. And: absence of a detected watermark does not mean human authorship. Detection rates drop sharply with editing, swapping just 10% of words in a 400-token passage cuts reliability from ~92% to 66%. In maths-heavy text, where phrasing is tightly constrained, detection falls further. OpenAI tested across benchmarks and found no performance difference between watermarked and unwatermarked outputs. But it also issued clear caveats: textGrain cannot prove who wrote what, whether ideas are original, or whether edits were substantive. It cannot distinguish AI-assisted translation from AI-generated content. A critical gap for non-native English speakers using AI to express their own thinking. MediaNama’s Nikhil Pahwa called it a ‘statistical fingerprint’, not a watermark in the traditional sense, and questioned its real-world utility. ‘Writing is compliance. Thinking is not,’ he wrote. ‘What you’re trying to challenge really is unoriginal thought, not whether something was written with AI.’ OpenAI plans to study how textGrain holds up after translation and multi-tool editing before expanding detector access. For now, it remains a narrow, regional, machine-readable signal. Not proof, not attribution, not enforcement. The EU AI Act requires such identification. OpenAI is delivering exactly that, and nothing more.