Latent Fact-Checking: Detecting Misinformation through Activation Engineering
The episode explores 'Latent Fact-Checking,' a paper detailing how to detect misinformation within AI models. Researchers define veracity as a geometric property in the model's representation space. They identify a specific 'falsehood direction' using Contrastive Activation Addition, pinpointing the optimal layer to outperform traditional prompting methods for fact-checking.