What happened

In October 2024, the Parliament adopted Claude for its public-facing archive tool, aiming to provide historical answers to citizens. 

Testing revealed the model’s inability to correctly answer basic questions, including the identity of the first Commission president (1958-1967). 

The system relies on Anthropic’s self-defined "Constitutional AI" principles, which prioritise vague criteria like "helpful, honest, harmless" over verified accuracy.

Why it happened

The European Parliament opted for Claude without robust accuracy assessments or exploring non-LLM alternatives, and relied on the model’s unverified claims about data compliance and prioritised integration with Amazon’s cloud infrastructure over factual reliability. 

The model’s training on web-scraped data - potentially including unvetted sources - likely contributed to the error.