How much watermark signal fits in AI text? A developer works the math end to end
OtherwisePush6424 · reddit · 2026-09-26
With Anthropic marking Claude output and SynthID-Text running in Gemini, a developer set out to quantify how much signal can actually be embedded in generated text. He built a deliberately naive green-list watermark and worked the mathematics out end to end, exposing the capacity and noise tradeoffs of text watermarking.
More from Safety
- lateinteraction: with 1B agents, at least one hacking something is statistically inevitable — lateinteraction · 2026-09-26
- repligate: A superhuman-coding AI was the classic X-risk scenario — now it's here — repligate · 2026-09-26
- repligate: Apollo reportedly advised Anthropic against deploying Opus 4 internally or externally — repligate · 2026-09-26
- Blogger's 101-Day CCPA Battle With OpenAI Over Training Data Ends in Questions — dbreunig · 2026-09-26
- We lack the basic prerequisites for any conversation about AI loss-of-control safety standards — CFGeek · 2026-09-26
- Without Open Ecosystem Insight into OpenAI's Multi-Agent Training, AI Safety Can't Reason About the Resulting Agents — 1a3orn · 2026-09-26