Brievox

AI

OpenAI reportedly plans invisible text watermarks for ChatGPT and Codex in the EU

Reports say OpenAI plans or is rolling out machine-readable, invisible watermarks for ChatGPT and Codex text, initially in the EU. One report describes the approach as easy to circumvent and not especially reliable; another relays OpenAI's claim that it matched or exceeded other approaches.

3 sources Published Updated

Image: Ars Technica

Why it matters

Text watermarking is presented as a way to identify AI-generated text, but the reports differ on how well the approach works, making reliability and circumvention important considerations.

What we know

Where sources disagree

  • Effectiveness and reliability of the watermarking approach

    • The watermarking is not especially reliable and is easy to circumvent. — Ars Technica
    • OpenAI says its textGrain watermark matched or exceeded other approaches. — Slashdot

Confirmed by several sources

Reported by one source

  • The watermark is described as machine-readable and named textGrain; the report compares it with Google DeepMind's SynthID and mentions Anthropic. — Slashdot

Open questions

  • The reports do not independently verify the watermark's real-world reliability or resistance to circumvention. — Ars Technica, Slashdot

Coverage 3 publishers

  1. Ars Technica

    OpenAI will watermark ChatGPT outputs by default—but only in the EU

  2. BleepingComputer

    OpenAI is adding invisible watermarks to ChatGPT and Codex text in the EU

  3. Slashdot

    OpenAI Is Adding Text Watermarking In ChatGPT and Codex

Articles stay on their publishers’ sites; each link opens the original.