Openai To Watermark Chatgpt And Codex Text In The Eu
OpenAI is rolling out invisible text watermarks for eligible ChatGPT and Codex users in the EU, while offering optional watermarking to API customers worldwide.
OpenAI is preparing to add invisible, machine-readable watermarks to eligible ChatGPT and Codex text outputs in the European Union. The rollout is intended to meet EU AI transparency requirements, according to Engadget and The Verge. OpenAI is not making watermarking the global default at launch, but API customers worldwide can opt in for select models.
Table of Contents
Where the watermark will appear
The Verge reports that the EU rollout will reach eligible ChatGPT and Codex users across all plans over the coming weeks. Engadget describes the change as applying to text and code generated in the EU, while The Verge characterizes it as watermarking text output from ChatGPT and Codex. Neither report says the feature will be enabled by default for users outside the EU.
According to The Verge, API customers around the world can opt in to watermarked text output for select models starting now. OpenAI is also working with cloud partners to make the option available for outputs accessed through their services in the coming weeks. Engadget notes that OpenAI has not specified which models will offer the option to customers outside the EU.
What textGrain can and cannot detect
OpenAI calls its proprietary system textGrain. As Engadget explains, it places a statistical signal in a model’s word choices that a detector can look for later. OpenAI says its approach matched or exceeded the performance of other methods, including SynthID for text, according to both reports. The Verge also says OpenAI presented benchmark scores showing similar performance for watermarked and unwatermarked text.
Detection remains uncertain. Engadget reports that OpenAI puts the success rate for shorter text at around 80 percent and says mathematical content and edited text can also be harder to assess. The Verge emphasizes OpenAI’s warning that a watermark does not establish whether a passage is accurate, who owns it, how much a person contributed or whether a person wrote it.
Why detector access is limited
Engadget links the rollout to Article 50 of the EU AI Act, which requires providers of generative AI systems to make text outputs identifiable in a machine-readable form. It reports that existing providers have until December 2 to comply. Both outlets note that Anthropic has also introduced text watermarking in response to the EU rules.
Approved researchers and expert organizations can apply for access to OpenAI’s detector, The Verge reports, but it will not be publicly available at launch because it could miss watermarks or produce false positives. The detector reports whether it finds an OpenAI watermark without identifying a user or exposing prompts or conversations. Engadget says OpenAI also plans to make the watermarking technology open source.
Sources
This story was compiled by AI from the reports below. Read the originals for the full details.