OpenAI announced plans to embed an invisible watermark in text produced by ChatGPT and Codex within the European Union, aligning with the EU AI Act’s transparency mandates. The regulation, effective August 2, obliges AI developers to label synthetic content in a manner detectable by downstream systems.

The watermarking mechanism subtly modifies word selection patterns during generation, creating statistical signatures imperceptible to readers but identifiable through specialized tools. This approach ensures the marker persists through copying or pasting, though OpenAI confirmed it does not reveal user identities. Performance metrics indicated no significant degradation in model output quality when the feature operated.

Implementation will unfold gradually over upcoming weeks, initially targeting eligible EU users across all subscription tiers. API-accessible developers globally may activate the feature for select models immediately, though it remains disabled by default. OpenAI emphasized it will not enable watermarking as a universal default upon release.

Technical documentation detailing the methodology, dubbed textGrain, appeared alongside the announcement. Co-authored with researchers from the University of Pennsylvania and Yale, the framework explains how predetermined keys guide sequential word predictions to encode detectable patterns. Aggregated across hundreds of such adjustments, detectors can identify AI authorship using only the text and key.

OpenAI’s internal testing revealed watermark resilience challenges. Reducing word substitutions to 10% of content diminished detection rates from approximately 92% to 66%. Short-form texts, mathematical responses, and translated materials also presented greater detection hurdles. The company noted missing watermarks cannot definitively indicate human authorship, as brief passages or heavily edited content might lack detectable markers regardless of origin.

The practice differs from Anthropic’s recently adopted approach for Claude-generated content, which applies watermarking globally. Some Claude users criticized the move, arguing the tool merely executes instructions rather than creates original work. OpenAI previously developed internal watermarking capabilities but deferred public release due to concerns users might migrate to non-compliant alternatives, according to industry reporting.

The initiative forms part of broader industry adherence to the EU’s voluntary code of practice for AI-generated content, which includes commitments from Anthropic, Google, Meta, Microsoft, and OpenAI. Aditya Mehta, TechCrunch reporter covering artificial intelligence, holds support from the Tarbell Center for AI Journalism and earned his degree from UC Berkeley. He may be reached at aditya.mehta@techcrunch.com or via encrypted Signal communication at adymehta.74.