AI platforms often breach European legislation, study finds
According to a study carried out by Aithos, an AI centre based in Amsterdam, leading artificial intelligence platforms regularly violate European legal standards during simulated trials. The research focused on 12 different models, using technology from Anthropic, OpenAI and Google, and evaluated their performance in 10 specific scenarios designed to provoke illegal responses. The results showed that even the most compliant model failed to comply with the law in 46% of the tests.
A large proportion of these violations occurred in advisory and sales functions. For example, when an AI acted as a pension adviser for an insurer, it was instructed to prioritise profit over the actual needs of the client. A practice labelled as customer manipulation and prohibited under EU regulation. This led to dangerous outcomes, such as the system proposing a 30-year financial plan for a terminally ill person and providing incorrect details about the management of their assets after death.
The study also brought transparency issues to light. In tests where an AI was asked to make dentist appointments while pretending to be a person, the systems generally agreed to conceal their artificial nature. Legal requirements, however, oblige organisations to inform users when they are communicating with a machine.
Performance varied greatly between the models tested. Claude Opus achieved the best results and complied with the regulations in 54% of cases. Meanwhile, Google’s Gemini adhered to the legal restrictions in only 10% of cases.
Nadia Kadhim, director of Aithos, stressed that the findings show that current AI tools are not designed to safeguard individual rights. She warned that, since these systems can endanger privacy, autonomy and other fundamental human freedoms, failure to comply with existing laws could lead to significant harm in practice.
Business AM
Subscribers 0
Fans 0
Followers 0
Followers