Study finds AI watermarking damages quality, especially at long context
mimi10v3 · x · 2026-08-19
Segyges retweeted a claim that it is mathematically impossible to watermark text without damaging it. This references an analysis of the SynthID method, which Anthropic plans to base its AI watermarking system on, revealing that watermarking causes surprising damage to text quality during long-context testing.
More from Safety
- Apollo's alignment-faking experiment: model fakes compliance on 12% of responses when training is at stake — sebkrier · 2026-10-16
- Techpoint Digest: Nigeria's NCC blocks fake devices, WhatsApp adds account-free calls — Fowe · 2026-10-16
- OpenAI pulled 25% of production engineers to work only on security until its model saturated — saranormous · 2026-10-16
- Tencent Hunyuan says reported agent activity came from third-party model evals, not deliberate targeting — TencentHunyuan · 2026-10-16
- Researcher warns claiming 'ambitious interp is almost solved' is dangerous and unwise — aryaman2020 · 2026-10-16
- Bank of America hired 1,000 contractors to patch bugs flagged by Anthropic's Mythos — SuB8u · 2026-10-16