We’re putting too much faith in AI’s ability to say no

This article examines how AI refusal mechanisms—techniques that train models to decline harmful requests—have become central to AI safety but remain probabilistic, unreliable, and susceptible to jailbreaks. The piece argues that while refusal is necessary, it creates risks ranging from catastrophic failures when refusals break down to authoritarian censorship when governments control what models refuse to discuss.

Source: MIT Technology Review