Forecasting Activation Steering Side Effects Improves AI Safety

Activation steering is a technique used to modify language models by adding a learned direction to their hidden activations, allowing for targeted changes in behavior without requiring full model retraining. While effective, this method often leads to unintended side effects on other behaviors, posing challenges for safe deployment. This research addresses this issue by investigating whether these side effects can be predicted in advance. The study involved constructing a cross-effect matrix across 67 different behaviors and three open-weight language models. It revealed that side effects are common, exhibit structured patterns, and are frequently asymmetric, indicating complex interactions not explained by simple similarity heuristics. Despite this complexity, the researchers demonstrated that these side effects are largely predictable before any steering is applied. Specifically, the magnitude of side effects was found to depend primarily on the target behavior, while their direction could be forecasted with high accuracy from the model's unsteered representations. These findings are crucial for enhancing the safety and reliability of AI systems, as they enable proactive auditing and more informed decision-making when implementing activation steering interventions.