Can we detect harmful AI before it's too late? Researchers search for warning signs

Georgia Tech has assigned a Ph.D. student named Eyas Ayesh to the question of whether you can spot an AI going wrong from the inside, moments before it does something unsafe, and TechXplore has put it on the record under the headline 'Can we detect harmful AI before it's too late?' Note the question mark. It is doing an enormous amount of work there, in the way a 'wet paint' sign does when nobody has told you which bench. The premise is charming: that a model drifting toward misalignment will develop a tremor, a flicker in the wiring, like a kettle about to boil. Software engineering suggests the machine will not telegraph anything; it will simply do the thing and let the incident report handle the introductions.

The stated payoff is that detecting these internal changes 'may inform more effective safeguards,' which is the research equivalent of saying better locks may inform more effective burglary prevention. But safeguards are a downstream product, and the thing about warning signs is that they need someone contractually obliged to act on them. We have spent a decade building models faster than we've built anyone willing to hit the brake, and the contribution here is a proposal to watch the dial more closely. There is no funding figure, no benchmark, no timeline, no named lab partner — just a student, an institution, and a question with a 'may' in the middle of it.

Which brings us to the stop button. The Verfassungsblog essay we covered made the case for a legal right to stop a system that has already gone wrong; Georgia Tech is now working on the earlier, cheaper version, where you spot it first. Neither answers the only question that matters at 3am in the incident channel, which is who has the authority, the login, and the willingness to be the person who turned it off. Warning signs are only useful to people who intend to read them.