Study: LLM watermarking shifts refusals and tool calls

Anthropic will embed invisible watermarks based on Google DeepMind's SynthID-Text in future Claude models. A study finds watermarking alters token sampling, causing 'sampling drift' that changes model refusals and agent tool calls, with effects varying by model and key. The findings matter for EU AI Act Article 50(2) compliance and AI safety.

• User Says Anthropic Wrongly Tagged Them for Training Competitor Model2026-10-31
• Anthropic study: harmless-looking data can train models to sabotage oversight2026-10-31
• Anthropic model finds Linux kernel bug enabling remote root access2026-10-30
• HN users question sudden Opus 5.5 quality drop2026-10-30
• Johns Hopkins cancer mapping tech picked for Anthropic AI for Science2026-10-29