OpenAI will watermark ChatGPT and Codex text in the EU to meet AI Act rules

What's happening

OpenAI will embed an invisible watermark in text produced by ChatGPT and Codex for users in the European Union, according to TechCrunch. The change, disclosed in a company blog post on Monday, is intended to align OpenAI with the transparency provisions of the EU AI Act, which took effect on August 2 and require AI-generated content to be marked in a way other systems can detect.

The rollout will happen over the coming weeks and applies to eligible ChatGPT and Codex users on all plans, but only within the EU. Developers using OpenAI's API can enable watermarking for select models anywhere in the world starting immediately, though it is switched off by default. TechCrunch notes that OpenAI is not making text watermarking a global default at launch.

How the watermark works

The watermark is not a visible symbol or a metadata tag. According to the company, it works by subtly steering the model's word choices as text is generated, producing a statistical pattern that human readers cannot perceive but a detector can identify. Because the signal lives in the wording itself, it survives copy-and-paste and travels with the text. OpenAI says the watermark does not identify individual users and that testing showed no meaningful change in model performance when it is enabled.

Alongside the announcement, OpenAI published a technical report on the method, called textGrain, written with researchers from the University of Pennsylvania and Yale. The report describes using a secret key to bias next-word predictions — small nudges that, accumulated across a passage, let a detector distinguish machine-written text using only the text and the key.

Fragile under editing

The watermark is not tamper-proof. In OpenAI's own testing, replacing roughly 10% of words with synonyms caused detection rates to fall from about 92% to 66%. The company also acknowledged that short passages, math answers and translated text are harder to detect reliably.

Those limitations explain why initial access to the detector is limited to approved researchers and expert organizations, which OpenAI says can help evaluate its reliability and responsible uses. The company also cautioned that a missing watermark does not prove human authorship: the text may be too short, too heavily edited, or produced by another provider's model. Watermarks, OpenAI noted, can indicate that one of its systems generated or contributed to a passage, but not how much human judgment, editing or creativity went into it.

Industry context

The move follows Anthropic's announcement two months earlier that it would watermark text generated by Claude, a measure Anthropic is applying worldwide. That decision drew backlash from some users who argued they supplied the instructions, context and decisions while the model was simply the instrument.

The Wall Street Journal reported in 2024 that OpenAI had built a text watermarking system previously but held off on releasing it, partly over concerns that users would switch to rivals that did not watermark. TechCrunch also points out that Anthropic, Google, Meta, Microsoft and OpenAI are among the companies that have committed to following the EU's code of practice on AI-generated content.

Why it matters

This is one of the first large-scale deployments of text watermarking by a major model provider, and it serves as a practical test of the AI Act's transparency requirements. The limitations OpenAI itself reported — fragility under synonym substitution and weak detection on short, mathematical or translated text — show that provenance through watermarking is far from settled science, and the decision to restrict detector access signals the company's own caution about false accusations and misuse. For developers outside the EU, the off-by-default API flag means watermarking remains a choice rather than an obligation for now, but the direction is clear: content provenance is becoming a regulatory expectation, and the debate over who truly authors AI-assisted text is likely to grow as these markers spread.