OpenAI is adding invisible, machine-readable watermarks to text generated by ChatGPT and Codex — but only for users in the European Union, and only as a compliance move under the EU AI Act.
Key takeaways
OpenAI's textGrain system embeds invisible watermarks in ChatGPT and Codex text output, rolling out first to EU users to satisfy the EU AI Act.
OpenAI claims textGrain "matched or exceeded" Google DeepMind's SynthID for text in internal evaluations — a comparison that has not been independently verified.
Anthropic announced a similar text-watermarking approach in August 2026, also based on SynthID-style methods.
Editing watermarked text can make the marks harder to detect, OpenAI acknowledges — meaning heavy revision degrades the signal.
The rollout is EU-only for now; no timeline has been given for other regions.
What textGrain actually does
TextGrain encodes a hidden pattern into the statistical choices a language model makes when selecting words — a method sometimes called "green-list" watermarking. The result looks identical to unwatermarked text to a human reader, but a detector scanning the output can, in theory, flag it as machine-generated. According to OpenAI, the system "matched or exceeded" Google DeepMind's SynthID for text in its own benchmarks. That comparison is worth treating carefully: OpenAI has not published an independent audit, and the claim comes directly from the company's announcement.
The practical limitation OpenAI itself flags is significant: editing degrades the watermark. Paraphrase a sentence, swap synonyms, or restructure a paragraph, and the invisible signal weakens. That matters for anyone using ChatGPT drafts as a starting point — which is most people. A lightly polished output may still carry a detectable mark; a heavily rewritten one likely won't.
The EU AI Act is the forcing function
The EU AI Act requires providers of general-purpose AI systems to label AI-generated content in ways that are machine-readable. OpenAI's EU rollout is a direct response to that obligation, not a voluntary transparency push. Anthropic made a comparable move in August 2026, according to The Verge, also building on SynthID-style techniques. Google DeepMind's SynthID for text has been available longer and covers Gemini outputs.
For now, users outside the EU — including the US — are not getting textGrain. OpenAI has given no public timeline for a wider rollout, and the regulatory pressure that triggered the EU deployment simply doesn't exist in the same form elsewhere. The Trump administration's approach to AI regulation has leaned away from mandatory disclosure requirements, which makes a US mandate unlikely in the near term.
OpenAI is rolling out its textGrain text watermarking system to ChatGPT and Codex for EU users under the EU AI Act.
What this changes for text-heavy creative workflows
For AI-art creators who use ChatGPT to generate prompts, captions, character backstories, or lore documents, the immediate workflow impact is minimal — the watermark is invisible and doesn't alter output quality. But the detection question is real. If a platform or client runs a watermark check on text you've submitted, a lightly edited ChatGPT draft could flag positive. Heavy editing is the practical defense, and OpenAI's own admission that editing degrades detection is the clearest signal of that.
The broader pattern is worth watching. Text watermarking is becoming infrastructure: SynthID covers Google, textGrain covers OpenAI, and Anthropic is following. As detectors become more widely deployed — by publishers, academic platforms, or content moderation systems — the gap between "AI-assisted" and "AI-generated" in practice will depend almost entirely on how much human editing sits between the model output and the final text.
Researchers have already documented statistical fingerprints in model outputs — Anthropic's Claude Opus 5.5 overuses the word "dependable" 23 times more than human writers, for instance — and watermarking adds another detection layer on top of those linguistic tells. The combination means AI-generated text is becoming harder to pass off as human-written, at least in jurisdictions where detection infrastructure is being built.
OpenAI has not disclosed the technical specifics of textGrain's encoding method or published the detection API, so third-party tools cannot yet integrate it. That gap will close as the EU AI Act's enforcement deadlines approach.
