← The Polylog AI Intelligence Brief
Morning Edition · Sunday, September 6, 2026Published at 2:32 AM EDT · New York
Anthropic Details a Text Watermark That Biases Claude's Word Choices Without Changing Output Quality
The method adds no hidden characters and no extra tokens, and a detected mark means text passed through Claude, not that Claude wrote it.
Anthropic published an explanation of how Claude's text watermark works after developers asked what it does to outputs. The mechanism is a keyed bias over token selection. Among the highest-scoring candidate tokens at each step, a key influ…
Continue the AI Intelligence Brief
Track frontier labs, chips, export controls, model releases, regulation, and AI infrastructure.
- 5 AI intelligence signals a day
- Frontier labs, compute, and chips
- Model releases and AI infrastructure
- Source-grounded analysis with confidence labels
The Global Intelligence Brief stays free.
Part of a tracked trend
Provenance Becomes Training Infrastructure
As machine-generated text takes over a growing share of the open web, watermarking and provenance tracking move from optional safety features into required components of training pipelines, and clean human data becomes a priced, contested input.
Townsquare
Start a discussion in Townsquare.
More from this edition
- OpenAI Opens GPT-6 Astra to Business Subscribers Days After Rating It Critical for Cyber Capability
- Nvidia Closes $12.93 Billion Purchase of Hugging Face, Buying the Distribution Layer for Open Models
- Anthropic's Claude Fable 5.1 More Than Doubles Its Predecessor on an Agentic Science Benchmark
- Claude Agents Produced a Machine-Checked Proof of Fermat's Last Theorem in Eleven Days
- Anthropic Previews a Hardware Standard Letting AI Agents Drive Lab and Factory Equipment
- Meta Prices Its Model Programming Interface Below Rivals as Muse Spark Reaches Version 1.3
- Bessent Says AI Companies Have Done a Horrendous Job Explaining Data Centers to Americans
Comments
0No comments yet.