Privileged Self-Access Matters for Introspection in AI
This addresses the practical question of AI introspection for researchers and developers, but is incremental as it builds on existing definitions.
The paper tackles the problem of defining introspection in AI by proposing a 'thicker' definition requiring privileged self-access, and shows through experiments with LLMs reasoning about temperature parameters that they can appear to have lightweight introspection but fail under this stricter definition.
Whether AI models can introspect is an increasingly important practical question. But there is no consensus on how introspection is to be defined. Beginning from a recently proposed ''lightweight'' definition, we argue instead for a thicker one. According to our proposal, introspection in AI is any process which yields information about internal states through a process more reliable than one with equal or lower computational cost available to a third party. Using experiments where LLMs reason about their internal temperature parameters, we show they can appear to have lightweight introspection while failing to meaningfully introspect per our proposed definition.