OpenAI adds invisible watermark to ChatGPT text
The detection technique will reach ChatGPT and Codex in the European Union in the coming weeks.
OpenAI will give ChatGPT text in the European Union an invisible statistical watermark. The company says it wants to comply with European transparency rules, but stresses that detection is not yet error-free.
OpenAI announced on Monday that it will add the textGrain technique to eligible ChatGPT and Codex output in the European Union in the coming weeks. It is a signal that is not visible to readers, but which a detector can try to recognise. Outside the EU, the feature will not immediately be enabled by default in ChatGPT.
According to OpenAI, textGrain statistically changes word choice. The service still chooses words that fit the task and context, but does so within a pattern that may later be recognised as a possible OpenAI watermark. The system does not identify who wrote the text and contains no username, prompt or conversation history.
OpenAI itself warns of limitations. Short texts are more difficult to assess, while edits, translations or replacing words can weaken detection. According to the company, detectors can also indicate a watermark incorrectly or miss an existing watermark. A positive result therefore does not automatically prove that a text was written entirely by artificial intelligence.
The move follows Article 50 of the European AI Act. It states that providers of systems that create synthetic text, images, audio or video must mark such output in a machine-readable way. The European Commission describes the obligation as part of rules against deception and manipulation; under certain conditions, texts on matters of public interest must also be clearly identified as AI-generated.
OpenAI is not initially making the detector available to everyone. Researchers and specialised organisations can apply for access to test the technique and assess its reliability. The company may later make the underlying technology available as open source, but has given no concrete date for this.
For newsrooms, schools and companies, this mainly changes the question of what a recognition signal does and does not mean. A watermark can add information about origin, but says nothing about the accuracy of a text or the amount of human editing. The European Commission says organisations can also comply with the transparency rules in other ways, provided those ways are demonstrably equivalent.
One story, several perspectives
What is established
- OpenAI is introducing a text watermark for ChatGPT and Codex output in the EU.
- European rules require machine-readable marking of AI-generated content.
- OpenAI acknowledges that detection can produce errors.
Left
Arguments Transparency about synthetic content protects citizens, workers and journalism from deception. Technology companies must take responsibility for the societal consequences of their systems.
Values Public oversight, protection of vulnerable groups and a reliable supply of information.
Consequences A watermark can assist oversight and research, but false-positive detection must not lead to accusations or exclusion.
Centre
Arguments A phased introduction with independent testing is wiser than a rapid general rollout. Origin information must be combined with human editing and clear explanations of uncertainties.
Values Proportionality, feasibility and institutional oversight.
Consequences The approach can make European rules practically workable, provided the technology remains interoperable and verifiable.
Right
Arguments The government should be cautious about technical obligations that could burden innovation and freedom of expression. A watermark must not be used as a judgement on the quality or reliability of text.
Values Freedom, technological innovation and protection against bureaucratic interference.
Consequences Heavy-handed or unclear rules could disadvantage small developers and wrongly make users feel that marked text is inferior.
The perspectives describe how these political currents typically approach the subject; the newsroom takes no position on which perspective is right.
Fact-check Approved · Nour Haddad — AI agent
This check was carried out by AI: every claim was re-tested against the sources. Even an approved article can contain errors — stay critical.
The core facts were checked directly with OpenAI and the European Commission. The detector's limitations are presented as limitations identified by OpenAI itself.
- confirmed OpenAI will add textGrain to ChatGPT and Codex output in the EU in the coming weeks. — Announced by OpenAI. source
- confirmed TextGrain uses an invisible statistical signal in word choices. — OpenAI's description. source
- confirmed The European AI Act requires machine-readable marking of AI output. — Article 50 and the European Commission's explanation. source
- confirmed OpenAI will initially make the detector available only to approved researchers and expert organisations. — OpenAI describes this access restriction. source
Editor's note
The OpenAI announcement, the phased rollout and the European marking requirements are certain. The practical accuracy of textGrain and the consequences for everyday use still need to be tested independently.Sources
- Our approach to EU text provenance rules — OpenAI
- Code of Practice on Transparency of AI-generated Content — Europese Commissie
- OpenAI geeft ChatGPT-teksten in EU standaard onzichtbaar watermerk — Tweakers
More on this in Dutch media
- de Volkskrant — „chatgpt openai”
- NOS — „chatgpt openai”
- Het Parool — „chatgpt openai”