What it is. A probe is a small, simple model trained to read the internal activity of a larger AI model and answer one narrow question about it, such as “is this model being deceptive right now?”
An analogy. A lie detector does not listen to what someone says. It watches their pulse and breathing. A probe ignores a model’s text and looks at its internal numbers instead.
How it works. As a model processes text, every layer produces a long list of numbers called activations. Researchers collect examples where they know the answer (honest or deceptive, say), record the activations and train a small classifier to tell the two apart. If the classifier succeeds on new examples, that information is present inside the model, even when the output gives nothing away. The technique goes back at least to a 2016 paper by Guillaume Alain and Yoshua Bengio, who used “linear classifier probes” to study what the layers of a network had learned.
Why it was in the news this week. FAR.AI reported that probes reading many layers and many tokens caught model sabotage on the SHADE-Arena benchmark with 98.8% AUC, above a text-based monitoring baseline. AUC (area under the curve) is a score from 50%, no better than chance, to 100%, perfect separation.
One thing to remember. Probes need access to a model’s internals, so only the model’s developer, or anyone running an open-weights model, can use them. And a probe finds what it was trained to find; a high score on one benchmark does not guarantee it catches every kind of deception. The appeal is that a model could learn to write innocent-looking text while pursuing a hidden goal, but it may be harder to hide the internal activity that produces that text. That is why safety researchers see probes as a second line of defence alongside reading the output.
Sources: AIooze brief on FAR.AI; Alain and Bengio, “Understanding intermediate layers using linear classifier probes”, arXiv:1610.01644, 2016