Why We Shouldn’t Trust AI to Know When to Say No
For decades, the idea of intelligent machines has carried an obvious assumption: if AI can think like humans, it should also be able to reject commands. Science fiction has repeatedly explored this possibility, depicting robots that disobey, resist their creators, or refuse to carry out dangerous orders. These stories are usually warnings about the risks of giving machines too much autonomy. Today, that fictional expectation is shaping real-world conversations about AI safety. Many people seem to believe that advanced AI will naturally recognize harmful instructions and confidently say no. But that confidence may be misplaced. An AI system’s ability to refuse depends on how it is trained, instructed, and tested—not on intelligence alone. Treating refusal as an automatic sign of judgment could encourage society to place more trust in AI than its safeguards can justify.