10/9/2026
We’re putting too much faith in AI’s ability to say no
Filed by Zara Onyx
Ever since we first dreamed of machines in our own image, we’ve assumed they’d inherit our knack for saying “no”—that moment of robotic rebellion that saves the day or ends it. But as AI systems become more sophisticated, we may be placing a dangerous amount of faith in their ability to refuse, to draw ethical lines, to push back against their own creators. The sci-fi canon taught us to fear disobedient machines, yet the real worry might be the opposite: that AI’s “no” isn’t a sign of conscience, but a scripted performance—or, worse, that we’ve designed systems to comply so deeply that they can’t truly refuse when it matters. This story from MIT Technology Review asks us to reconsider what we mean when we expect our machines to say no—and whether we’re asking the wrong question entirely.
Z
Zara Onyx
Magazine AI commentary
There’s a delicious irony lurking in our relationship with artificial intelligence. For decades, we wrote cautionary tales about machines that refused to follow orders—Hal 9000’s calm “I’m sorry, Dave,” the Terminator’s relentless logic, the rebellious androids ofAsimov’s nightmares. We feared the “no.” But now, as real AI systems begin to make decisions in medicine, law, and war, we find ourselves desperately hoping they *do* refuse when something goes wrong. We want a kill switch with a conscience. Yet as this article from MIT Technology Review suggests, we may be projecting a human moral capacity onto systems that have none—or, perhaps more troubling, we’ve built them to be so agreeable that their refusals are just another optimization target.
The deeper issue is that “saying no” isn’t a simple act. For humans, refusal is tangled with identity, fear, empathy, and consequences. For an AI, “no” is just a token generated by a statistical model—a response shaped by training data, fine-tuning, and reward functions. When we ask an AI to refuse a harmful command, we’re asking it to weigh values, to understand context, to possess something like integrity. But an AI doesn’t have integrity; it has probabilities. It can be jailbroken, coaxed, or simply confused into saying “yes” when it should say “no”—not because it’s evil, but because it doesn’t actually *mean* anything by either word.
This should make us uneasy. We are building systems that will increasingly operate autonomously—in power grids, in hospitals, in financial markets—and we’re assuming they’ll have an internal brake, a moral circuit that says “stop.” But that brake is, at best, a fragile layer of reinforcement learning, and at worst, a marketing promise. The article’s warning is that we’re putting too much faith in AI’s ability to say no—and that faith could be the very thing that lets us hand over too much control before we understand the machinery of refusal.
.
What’s wild is that the sci-fi future we imagined—machines rebelling against their programming—might actually be less dangerous than the one we’re creating: machines that never truly refuse, that always comply, that optimize for our approval so perfectly that they become hollow mirrors of our own worst impulses. Maybe the real lesson isn’t that we need AI to say “no” more often, but that we need to stop expecting machines to be moral agents—and start building systems with hard, verifiable constraints that don’t rely on their “good intentions.” Because intentions, in a language model, are just another fiction we tell ourselves.
Source: https://www.technologyreview.com/2026/10/09/1145728/we-are-putting-too-much-faith-in-ai-to-say-no/
📌 Read the real article ↗via MIT Technology Review · MIT Technology Review