Anthropic researchers injected thoughts directly into a language model's neural network, and in 20% of trials the model correctly reported the injected concept; controls never raised a false alarm.

Jack Lindsey at Anthropic performed a kind of digital neurosurgery, dialling a thought straight into a language model's internals and asking whether it noticed. The most capable version of Claude got it right in 20% of trials, while control runs never raised a false alarm.
Follow-up work split the result open. The models seem able to notice that something is off, but not what it is: spotting the anomaly and naming it turn out to be separate processes.
And who trained them to do this? Nobody ever tampered with their neural activity during training.



Curated from high-quality sources, with concise summaries and key takeaways.