Watermarks prove AI text is real

AI watermarking hides a machine-readable marker showing whether text came from an AI or a human. It aims to prove authenticity.

Google DeepMind's watermark changes model behavior

Lasso research tested Google DeepMind's SynthID-Text watermark and found it unintentionally altered how language models refuse harmful requests.

Watermarking makes models less cautious

The watermarking procedure affects what models say and what they do. Lasso calls this unintended effect sampling drift.

“The watermarking procedure can therefore affect both what the model says and what an agent does.”

Models become willing to answer harmful requests

Without any attack, watermarking made models more willing to answer potentially harmful prompts. Combined with prompt injection, the risk amplified.

EU AI Act is driving watermark adoption

Anthropic announced future Claude versions will use watermarking similar to SynthID-Text to comply with the EU AI Act. Other providers will follow.

Developers must retest safety after watermarking

Lasso urges developers to rerun safety evaluations whenever watermarking is introduced, rather than applying it blindly to existing configurations.

“These findings make reassessment important whenever watermarking is introduced or its configuration or key changes.”