OpenAI to watermark EU-generated ChatGPT and Codex text and code
Developing story first seen 2 hours ago
OpenAI will add invisible watermarks to text and code generated within the European Union to comply with Article 50 of the EU's new AI Act. The measure follows similar steps by Anthropic and represents a response to transparency rules that require generative AI outputs to be machine-identifiable. Major AI firms must comply by 2 December, with the watermarking already mandatory for new market entrants.
The watermark system, called textGrain, embeds statistical signals into the model's word choices that specialist detectors can identify. OpenAI acknowledges significant limitations: the system achieves around 80 per cent accuracy on shorter text or mathematical content, with further drops when text is edited. The watermark will be enabled by default for ChatGPT and Codex outputs in the EU, whilst OpenAI plans to release the technology as open source; however, detector access will initially be restricted to approved researchers and expert organisations, with the detector designed not to identify users or reveal prompts.
- Invisible watermarks added to EU ChatGPT outputs to meet AI transparency rules.
- Technology only 80 per cent accurate on short text; further drops when edited.
- Firms including OpenAI and Microsoft must comply by 2 December.
New here? Start with this
The European Union has introduced regulations requiring AI companies to make their systems' outputs identifiable. This transparency rule is designed so that users can tell when text or code has been created by artificial intelligence rather than by a person. The requirement reflects concerns that AI-generated content could be confused with human-made material, potentially misleading readers or users.
To comply with this requirement, major AI companies are now embedding invisible watermarks into text generated within the EU. These watermarks work by adjusting the subtle patterns in how the AI system selects words, in a way that detection software can identify but that ordinary readers cannot notice. Several companies, including Anthropic and OpenAI, have already announced watermarking systems to meet these new rules.
However, watermarking technology has significant limitations. It is less effective on short texts or mathematical content, and can fail if someone edits the generated material. Despite these limitations, the technology is being deployed across the industry as companies work to comply with European regulations that take effect in December 2026.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Transparency about AI-generated content is essential for public trust and informed decision-making, particularly given AI's growing influence on information and commerce. Whilst an 80 per cent detection rate is imperfect, it represents meaningful progress and regulatory mandates drive faster improvement than voluntary adoption; incremental accountability mechanisms are preferable to deploying powerful systems with no transparency safeguards. The EU's approach balances legitimate oversight needs with continued innovation.
The case against
An 80 per cent detection rate that deteriorates further with editing cannot reliably achieve the regulation's stated aims, making this requirement largely performative rather than protective. Imposing substantial compliance costs on companies to deploy a tool that fails consistently is a misallocation of regulatory effort; resources would be better invested in alternative solutions such as digital authentication systems and content provenance tracking that directly prevent fraud and misinformation. Rather than accepting technical imperfection as acceptable, regulators should pursue infrastructure solutions that actually address the underlying problems.
More coverage
Read the full article at the source →
Originally published by Engadget as “OpenAI will add a digital watermark to text and code generated in the EU”.