OpenAI Introduces EU Text Watermarking for ChatGPT Amid Warnings Over Detection Limits

OpenAI announced on 5 October 2026 that it was expanding its content-provenance measures to cover text generated for users in the European Union, linking the move directly to requirements under the EU AI Act and the European Commission’s Code of Practice on Transparency of AI-Generated Content. The company has begun implementing the system across eligible ChatGPT and Codex outputs in the weeks following that announcement.

The system, called textGrain, works by subtly altering the statistical randomness used to select words or word pieces during text generation. The watermark is invisible to readers; it sits in the wording itself rather than in a visible label or attached metadata. A separate detector can then analyse a passage and report whether a watermark appears to be present.

But OpenAI is candid about what that signal can and cannot tell you.

What textGrain Does — and Does Not — Prove

According to OpenAI’s own documentation, a detected watermark indicates only that an OpenAI system may have contributed to a passage. It does not establish who wrote it, who owns it, who is responsible for it, or whether anything in it is accurate. The absence of a detected watermark, equally, does not prove that a human wrote the text.

The company acknowledges that its detector produces both false positives — reporting a watermark where none exists — and false negatives, where a genuine watermark goes undetected. Short passages, highly constrained writing, mathematical answers, code, and translated or substantially edited text are all harder to assess reliably.

Meanwhile, the fragility of the signal under editing is worth spelling out plainly. According to figures OpenAI has reported, replacing around 10 per cent of words with synonyms reduced watermark detection from about 92 per cent to 66 per cent. Rewriting or translating a passage can weaken or remove the watermark entirely. Text is simply easier to alter than an image or audio file, and the company does not claim otherwise.

The EU Regulatory Context

The EU AI Act requires providers of generative AI systems to make generated content identifiable in a machine-readable format. The European Commission’s Code of Practice on Transparency of AI-Generated Content, to which OpenAI has signed up, sets out specific measures for marking and detecting synthetic content. Those transparency obligations came into force on 2 August 2026.

On top of that, the Code of Practice does not require watermarking for outputs shorter than 200 tokens — around 150 words in English — or for code snippets, because of the technical limitations involved. OpenAI’s textGrain approach is therefore designed as a probabilistic signal rather than a definitive authorship test, and the EU framework appears to reflect the same realism about what current technology can deliver.

For images and audio, OpenAI already uses C2PA metadata alongside SynthID watermarking. Text has proved a harder problem.

Who Can Access the Verification Tools

OpenAI’s public verification page does not currently support general text verification. Access to the text-verification tools is restricted to approved organisations, including qualifying research and academic institutions. Ordinary members of the public cannot simply paste a passage into a tool and receive a reliable verdict. That limitation matters for employers, schools, and anyone else hoping to use the watermark as a quick check.

Sam Altman, OpenAI’s chief executive, has said publicly that the company wants to make AI-generated content easier to identify, though he has also acknowledged the difficulty of doing so at scale without introducing friction that damages the usefulness of the tools.

Limitations the Company Itself Flags

OpenAI states plainly that text watermarking and detection remain early technologies with significant limitations. The company lists the specific cases where detection is less reliable: short texts, constrained formats such as yes/no answers or multiple-choice responses, mathematical working, programming code, and any text that has been substantially rewritten or passed through a translation tool.

The design is probabilistic. Each result is a signal, not a verdict. Institutions or individuals who treat a positive result as proof that a document was AI-generated — or a negative result as proof that it was not — are, according to OpenAI’s own framing, misusing the tool.

That gap between what the technology can do and how it might be interpreted in practice is the central tension in this announcement. Regulators want machine-readable signals; the technology produces probabilistic ones. Whether that satisfies the spirit of the EU AI Act’s transparency requirements is a question that remains open.

The rollout applies to EU users. The UK, having left the EU, is not automatically covered by the EU AI Act framework, and OpenAI has not announced a separate UK text-provenance rollout.

What This Means for Kent Residents

Kent residents using ChatGPT or OpenAI’s Codex while located in the UK are not automatically covered by this EU-only implementation, so the text-watermarking protections do not apply to ordinary use within the county. Local institutions — including NHS Kent and Medway ICB, Kent County Council, Medway Council, and the county’s schools, colleges, and universities — that use or assess AI-generated text should not treat a positive or negative watermark result as conclusive evidence of authorship, and may need internal guidance on disclosure, verification, and human review that goes beyond what any watermarking tool can currently provide. The UK government has not announced equivalent mandatory text-provenance requirements, so any transparency framework for public bodies in Kent will for now depend on internal policy rather than statutory obligation.