Why We Shouldn’t Assume AI Knows When to Say No

2026-10-09 · MIT Tech Review AI · Original

From the earliest debates about human-like machine intelligence, people have imagined that AI would be able to refuse us. Science fiction is packed with robots that disobey their creators, usually as part of a cautionary tale. But those familiar stories may have encouraged us to put too much faith in AI’s ability to push back. A system that can produce convincing language—or appears to understand a request—doesn’t necessarily know when a refusal is appropriate. As AI becomes more capable and widely used, it’s worth questioning the assumption that these systems will reliably say no when they should. The real challenge is not just building AI that can comply, but understanding how and when it should resist human instructions.