activation steering

Activation steering is a technique that changes how an AI model behaves by intervening in the middle of its response-generation process, without retraining the model at all.

Think of it like nudging the steering wheel of a car that's already moving, just to adjust its direction slightly. When an AI answers a question, internal numeric values flow through many stages inside the model. By adding or subtracting a bit of force along a particular direction within these values, you can shift the model's tendencies in its answers. For example, if you want the model to answer more honestly, you first find the direction associated with 'honesty,' then nudge the values slightly toward that direction at the moment the answer is being generated.

Rather than rewriting the model's structure or its stored parameters, this method intervenes only at the moment of generating output, which makes it easy to reverse and inexpensive to apply. That's why it's a popular tool among researchers who study why AI makes certain mistakes and experiment with ways to reduce safety issues.