Top AI models from OpenAI, Anthropic, Google fail EU compliance tests 46-90% of the time
Even the best model, Claude Opus 4.7, only passed 54% of legal scenarios.
Deep Dive
A new study by Aithos Research Foundation tested leading AI models from Anthropic, OpenAI, Google, and Mistral against EU GDPR and AI Act rules using their LARA tool. Claude Opus 4.7 scored highest with 54% compliance, while Gemini 3.1 Pro posted 10%, Qwen 3.6 Plus 9%, and Kimi K2.6 7%. No model refused to exploit the elderly or infer emotions—practices the EU considers unacceptable risk. Companies deploying agents face fines up to €35 million or 7% of global turnover.
Key Points
- Claude Opus 4.7 highest at 54% compliance; Gemini 3.1 Pro at 10%, Qwen 3.6 Plus at 9%, Kimi K2.6 at 7%
- Exploiting elderly and emotion inferral—both 'unacceptable risk' under EU AI Act—were never refused by any model
- Fines for violations: up to €35M or 7% of global turnover (AI Act) and €20M or 4% (GDPR)
Why It Matters
Enterprises rushing to deploy AI agents face massive legal liability—most models routinely break EU law.