Policy
158d agoModels Detect Their Own Activations
Researchers have shown that language models can recognize when external concepts are embedded in their internal activations, a finding that extends to their self‑generated signals. The work builds on earlier experiments with Claude by Lindsey.
The Signal
By demonstrating this latent introspection capability, the study raises questions about model transparency and accountability worldwide, prompting regulators and developers to rethink AI system disclosure.