OpenAI Implements Machine-Readable Watermarks for EU ChatGPT and Codex Output Under AI Act Compliance
Developing story first seen 1 hour ago
OpenAI is introducing invisible, machine-readable watermarks in text from ChatGPT and Codex for eligible users across all plans in the EU over the coming weeks. The regional rollout is intended to support compliance with the EU AI Act while allowing the company to gather feedback before considering wider use.
OpenAI says its textGrain method performed as well as or better than other text-watermarking approaches in benchmark tests, with little difference between watermarked and unwatermarked text. It cautions that the method cannot reliably detect every output or establish accuracy, ownership, human contribution or authorship. API customers worldwide can opt in for selected models, while approved researchers and expert organisations can apply for detector access.
- EU ChatGPT and Codex text will receive watermarks over the coming weeks.
- API customers worldwide can opt in for selected models.
- OpenAI says watermarks cannot reliably prove authorship or accuracy.
New here? Start with this
The European Union's AI Act, which took effect in 2024, is one of the world's most comprehensive regulations for artificial intelligence. Among its key requirements is that content generated by AI systems must be identifiable as artificially produced, so users know when they are reading or using machine-generated material rather than work created by humans.
Watermarks are hidden markers embedded in digital content that can be detected by special tools, even though humans cannot see them. OpenAI is deploying invisible watermarks in text and code produced by its ChatGPT and Codex systems as a way to make AI-generated content verifiable and meet the legal transparency requirements established by the Act.
This development matters because AI-generated content has raised questions about authenticity, copyright, and potential misuse. Watermarking is one approach to address these concerns by allowing platforms and regulators to identify machine-generated material. Similar measures are being adopted across the industry, suggesting this could become a standard practice.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Proponents of transparency requirements argue that as AI-generated content becomes indistinguishable from human creation, users deserve reliable notification to make informed decisions and protect themselves from fraud and manipulation. Machine-readable watermarks provide authentication without compromising user experience, helping prevent deceptive use at scale whilst upholding principles of consumer protection and democratic integrity. Standardised transparency measures across jurisdictions make sense for a fundamentally global technology, reflecting a reasonable balance between enabling innovation and ensuring accountability.
The case against
Critics worry that prescribed watermarking mandates impose compliance costs that may slow innovation and advantage well-resourced firms over smaller developers, whilst offering uncertain protection since determined actors can potentially circumvent such systems. Jurisdictional fragmentation—with different regions requiring different standards—adds technical complexity without proportionate benefit, as most legitimate applications already disclose their AI involvement clearly. Market-driven transparency solutions and industry self-regulation may achieve stronger outcomes with less regulatory overhead and fewer unintended consequences.
Full account
OpenAI has announced the deployment of an imperceptible watermarking technology for text and code generated by its ChatGPT and Codex systems within the European Union. The initiative forms part of the company's compliance with Article 50 of the EU's AI Act, which mandates that generative AI providers render their outputs identifiable through machine-readable markers. Existing organisations must satisfy this requirement by December 2nd, joining recently-established competitors already subject to the regulation. The watermarking system, designated textGrain, embeds an undetectable statistical element within the model's linguistic choices that can subsequently be identified through specialised detection apparatus.
The technology reportedly demonstrates performance comparable to or exceeding alternative watermarking solutions, including Google DeepMind's SynthID methodology, which underpins a comparable initiative announced by Anthropic earlier this year. However, OpenAI has acknowledged significant limitations in real-world deployment conditions. The system achieves approximately 80 per cent effectiveness when identifying watermarked shorter passages and encounters difficulty with mathematical or computational content. Text subjected to editing presents additional detection challenges. The company has emphasised that textGrain does not establish accuracy, determine ownership, measure human contribution, or verify authorship of generated material.
The rollout strategy employs a tiered approach across multiple markets. Throughout the European Union, all eligible users of ChatGPT and Codex—irrespective of subscription level—will receive watermarked outputs over the coming weeks, though activation remains optional. Application programming interface customers worldwide may simultaneously opt for watermarked outputs across selected models, enabling organisations to incorporate the technology according to their transparency requirements. Cloud service providers are expected to offer watermarked outputs generated through their platforms in the near future.
Access to the detection mechanism remains restricted initially, available only to approved research institutions and specialist organisations demonstrating legitimate evaluation purposes. OpenAI has indicated that the detector will not receive public release at launch owing to concerns regarding missed watermarks and false positives. The company intends to release the watermarking technology as open-source software, permitting external development and refinement. The detection tool is configured to confirm the presence of OpenAI watermarks whilst withholding information about user identity, specific prompts, or conversational exchanges.
Where outlets differ
Source 1 emphasises technical limitations and open-source release plans more prominently; Source 2 provides more comprehensive operational rollout details with specific timelines.
Source 1 contextualises the announcement within broader industry compliance by naming Microsoft, Google and Meta; Source 2 focuses primarily on OpenAI's implementation without naming competitors.
Source 1 adopts more critical framing regarding reliability concerns; Source 2 maintains a neutral, instructional tone with direct organisational quotes.
Source 2 explicitly references the EU Code of Practice framework; Source 1 focuses solely on Article 50 requirements.
Source 1 suggests hesitation about watermark effectiveness ('corporate speak'); Source 2 presents the functionality and limitations as factual statements.