Quanta Magazine looks at experiments that try to test whether language models can introspect. In 2025 a researcher injected activity patterns linked to a concept into a model and asked whether it noticed anything; some models did, about one time in five. Follow-up work by linguist Kyle Mahowald and philosopher Harvey Lederman showed that what the models said about the injected concept could be explained by the method itself.

What survived their controls was a narrower ability: sensing that something unusual had happened. The researchers call this anomaly detection, and whether it counts as introspection is still disputed. The question matters for freethinkers because it forces us to define, carefully and without wishful thinking, what self-knowledge actually is, in machines and in ourselves.

Source: Quanta Magazine (2026-11-04)
https://www.quantamagazine.org/new-research-tests-whether-ai-models-know-their-own-thoughts-20261104/

Summarised by the FTI newsdesk.