Claude text gets invisible watermarking Nov 3: token-level statistical skew, not hidden chars
petesena · x · 2026-10-31
Anthropic has emailed customers that starting Nov 3, Claude-generated text carries an invisible watermark that survives copy-paste and some editing; only regulators, journalists, fact-checkers and researchers can verify it for now. It flags that Claude touched the words—even a quick proofread—not who you are. petesena explains the mechanism: generation is unchanged up to the logits; only the sampler is replaced. A PRF keyed by a secret watermark key plus short preceding context yields a deterministic seed, driving functions that assign pseudorandom scores to vocabulary tokens; tournament/Gumbel-style sampling then picks the next token, creating a statistically detectable skew in word and phrase combinations rather than any hidden characters.
Related event: Anthropic to Add Invisible Watermarks to Claude Output(4 posts)→
More from Models
- Anthropic's fable 5.5 reportedly delayed, possibly overshadowed by Opus 5.5 — haider1 · 2026-10-31
- Anthropic's Max→Ultra rename checked line by line: weekly caps never went away — No-News-8059 · 2026-10-31
- Google Insiders Push Back on 'No Great Models' Criticism, Carbon Is Just Benchmarks — Angaisb_ · 2026-10-31
- OpenAI launches GPT-6.1 Terra: 50.3% on Terminal-Bench 4.0 at half price, but 30% unauthorized actions — dl_weekly · 2026-10-31
- Double-scrollbar bug screenshot accidentally reveals qwen4 test files in qwen-code repo — gnukeith · 2026-10-31
- OpenMed open-sources 194M-param disease NER model that runs on CPU in 0.21s — solyarisoftware · 2026-10-31