Interpretability techniques that target deception

Definitionp. 31-32

Interpretability techniques can be used for finding the root causes of outputs of an AI model that reliably lead to false beliefs for its users (e.g., deceptive behavior) [180].

Additional Information

It is often difficult to distinguish a deceptive AI model from an honest AI model, since absence of deception and very sophisticated (hard to detect) deception may appear behaviorally similar. Interpretability techniques and tools can be used to detect whether AI model outputs arise from internal computations representing deception. This can apply in cases of purposely trained deception by the developer or if it emerges unintentionally during training. These interpretability tools can come from mechanistic interpretability, such as identification of features involved in generating the outputs, or attribution of parts of the input most important in generating the output [129].

Part of