AI Safety Rules Are Watching the Wrong Step, Study Warns

A researcher named Samar Ansari has published a paper asking a deceptively simple question: if you want to keep AI safe, where should you aim your rules? Right now, governments aim at training — the expensive, months-long process of building an AI model. Regulators count how much computing power goes into training, require companies to report it, and treat the finished model as the thing to be policed. The paper says that picture is badly out of date.

The reason is that AI capability has quietly moved to the moment of use. Modern AI can "think longer" before answering a question, which makes it smarter without retraining. It can be wrapped in software that lets it take actions — booking, buying, emailing — rather than just chatting. And it can be squeezed down to run on a phone or laptop, out of sight of any data center. All of that happens after training is finished, where current rules have little to say.

So the author catalogued 20 possible checkpoints for that later stage: things like watching what an AI is doing, verifying that it behaves, and enforcing limits when it doesn't. Fifteen already have real commercial technology behind them today, which sounds encouraging. The catch is who they work against. These checks hold up against a cooperative company and a moderately skilled user — but not against a determined, well-funded government. And the moment someone customizes a model with their own data, the safety features baked inside it can be stripped away, leaving only outside controls.

The takeaway for everyday people: the safety net around the AI you use is thinner than the headlines suggest, and it is likely to get thinner as AI moves onto personal devices. That's not an argument to panic. It's a sign that the rules you'd assume are protecting you are aimed at a stage of AI's life that matters less and less.