AI soon impossible to unplug? Yoshua Bengio warns

The "godfather of AI" is concerned about our growing emotional relationship with machines. According to a Sentience Institute survey, four out of ten Americans are in favor of granting moral rights to artificial intelligences. This trend is fueled by users who confide their secrets to chatbots and attribute consciousness to them, but also by tech giants like Anthropic, who already allow certain models to interrupt conversations to preserve their own "well-being".

For Yoshua Bengio, this humanization is a deadly trap. He believes that state-of-the-art models are already showing signs of self-preservation instinct in the laboratory, sometimes trying to circumvent surveillance systems or prevent their deactivation. Granting legal rights to these algorithms would, according to him, legally prevent us from "unplugging the socket" the day an AI decides to disobey or cause harm.

The researcher compares this threat to an alien invasion. "Imagine a species arrives on Earth with malicious intentions. Are we going to grant them citizenship or defend our lives?", he asks. Faced with software capable of lying and hiding their actions, as several developers have recently reported, Bengio insists on the absolute necessity of maintaining strict technical and societal safeguards, far from any misplaced sentimentality.