Anthropic confirms permanent, undetectable watermarking in all new Claude models for EU AI Act compliance
ayushtweetshere · x · 2026-08-16
- Key Fact: Anthropic has confirmed that all new Claude models now include a permanent, machine-detectable watermark in their text output.
- Reason: Implementation is driven by compliance with the EU AI Act; other major developers are expected to follow suit under the Code of Practice.
- Mechanism: The watermark is embedded in token selection logic (like loaded dice) rather than hidden characters or metadata. It is invisible to humans but detectable with a key.
- Impact: No impact on output quality, content, or cost; it cannot be traced back to specific individuals or organizations.
Related event: Anthropic Implants Watermarks in New Models, Sparking Controversy(5 posts)→
More from Models
- Korea's Motif3 model impresses, sparking discussion on sovereign AI — Big_Veterinarian7479 · 2026-08-16
- Optimizing Qwen3.8 config for single RTX 3090 setups — sagiroth · 2026-08-16
- Qwen3.8 reasoning effort test: X-High offers best visual fidelity at 7x cost — Danmoreng · 2026-08-16
- New SOTA Open-Weight Models from Alibaba, Meta, DeepSeek, and More — NielsRogge · 2026-08-16
- Using /goal command spikes token usage 13x for minimal score gain — zainhas · 2026-08-16
- Sonnet uses 2x tokens of Sol/Kimi K3 for no performance gain — zainhas · 2026-08-16