OpenAI has announced a new text watermarking framework called textGrain to comply with European Union AI Act requirements for identifying machine-generated text. The system introduces an invisible statistical signal directly into generated text outputs, enabling a specialized detector to assess whether a passage originated from OpenAI's models.
Rollout Across ChatGPT, Codex, and API
Starting on October 5, 2026, global API customers can choose to opt in to text watermarking for select models. For consumer-facing products, OpenAI plans to deploy invisible watermarks for eligible ChatGPT and Codex text outputs in the European Union in the coming weeks.
Access to the corresponding detection tool will not be public immediately. OpenAI has opened an application process for approved researchers and expert organizations to test and evaluate the detector's efficacy.
Statistical Word Choice and Benchmark Impact
The textGrain mechanism works by applying subtle statistical adjustments to a model's word selection. According to OpenAI, internal evaluations showed that watermarked outputs scored slightly higher on select benchmarks compared to unwatermarked text, though the company observed no meaningful difference in overall performance quality.
OpenAI also expressed plans to release the underlying textGrain technology as open source in the future, allowing external entities to build on the method.
Detection Accuracy and Limitations
OpenAI highlighted significant constraints in detection capability, noting that the tool can produce both false positives and false negatives. Reliability drops on short passages and domain-specific content such as mathematics, where constraints on vocabulary leave minimal flexibility for word choice.
Post-generation editing severely weakens the watermark. In OpenAI's tests on 400-token passages, substituting 10 percent of the words with synonyms reduced detection accuracy from roughly 92 percent to 66 percent. Replacing 25 percent of the words dropped detection accuracy down to 17 percent.
OpenAI emphasized that the mark does not measure human contribution, verify factual accuracy, establish ownership, or identify individual users.
AI-generated editorial illustration; not a documentary photograph.
Source: OpenAI details new text watermarking system for ChatGPT, Codex, and the API
SIGNAL TITANS
Leave a Reply