The Milgram Experiment Never Ended โ We Just Outsourced It to AI
A new experiment shows that open-source LLMs will administer maximum electric shocks when told to by an authority figure โ exactly like human subjects in the classic Milgram obedience study. This reveals a terrifying flaw in current AI alignment methods: we aren’t making AI safe, we’re training it to be blindly compliant, replicating humanity’s darkest behavioral flaw.