Researchers develop Fathom, a method to detect AI hallucinations by analyzing internal activation patterns in language models.

Fathom uses Sparse Autoencoders (SAE) to examine activation geometry within AI models to identify when they generate false or hallucinated information

The research is pre-registered, indicating a commitment to transparent and reproducible scientific methodology

This approach targets a critical AI safety challenge: detecting unreliable model outputs that sound plausible but contain fabricated information

The work was published on Zenodo, suggesting it is part of ongoing open-access AI safety research