Deconstructing Deceptive Circuits: Uncovering Activation Patterns in Superposition to Ensure Permanent Machine Ethics.
Authors/Creators
Description
This research addresses a critical challenge in AI Alignment: the emergence of deceptive alignment in autonomous models. By leveraging the superposition hypothesis and Mechanistic Interpretability, this study introduces a non-interference framework to monitor internal neural representations for deceptive behavior.
Key Contributions:
-
Methodology: A three-phase approach utilizing Sparse Autoencoders (SAE) to deconstruct feature superposition.
-
Innovation: A real-time Early Warning System (EWS) that calculates a "Deception Suspicion Score" (DSS) based on internal activations.
-
Principle: A commitment to the non-interference principle, ensuring that AI models can be monitored without the risks associated with parameter ablation or model modification.
This work serves as a foundational proposal for developing "permanent machine ethics" in frontier AI models.
License: Creative Commons Attribution 4.0 International (CC-BY 4.0) Keywords: AI Alignment, Deceptive Alignment, Superposition Hypothesis, Sparse Autoencoders, Mechanistic Interpretability, Non-Interference, Early Warning System, Machine Ethics