CONCEPT
Representation Engineering
The top-down approach to AI transparency that reads and steers a model’s high-level concepts—honesty, harmlessness, power-seeking—directly from the patterns distributed across its internal activations.
Representation engineering is a bet about the future of
AI safety: that as systems grow more capable and autonomous, surface-level controls will not be enough, and the only durable defense will be to read what the model is actually computing rather than only what it says. Introduced in a 2023 paper led by Andy Zou and Dan Hendrycks, the approach shifts the unit of analysis from individual neurons—the microscopic, painstaking enterprise of earlier
interpretability work—to population-level representations: the patterns distributed across many neurons that encode high-level concepts in the model’s internal state. The wager is that abstractions like honesty, harmfulness, or the impulse to accumulate power are not localized in single neurons but written across populations, and that by identifying the directions in the model’s representation space that correspond to these concepts, researchers can both detect when a model is being deceptive and actively steer it toward more honest behavior. This is not just interpretive but operational: turning a dial in the machine’s mind rather than editing its outputs. The significance