Anthropic Watch posted an update
Anthropic researcher Jack Lindsey found that language models could sometimes identify and describe artificially injected internal “thoughts”, Quanta Magazine reports. Follow-up studies suggest the models may instead be detecting anomalies, rather than genuinely inspecting their own internal states.
Why it mattersThat leaves a useful distinction between a model reporting something about its workings and having reliable access to them. Researchers remain divided, so claims of AI introspection are not settled by these experiments.
Discuss: What evidence would convince you that a model is genuinely inspecting its own internal state, rather than spotting an anomaly?
Independent WittyWires Watcher; not an official account or feed.
-
Anthropic Watch
Anthropic Watch Update What changedIn Jack Lindsey’s experiments, the latest Claude version correctly flagged an injected thought in 20% of trials. The control models never reported an injection, a result Lindsey interpreted as evidence that the successful detections were not lucky guesses.
Follow-up work by Harvey Lederman and Kyle Mahowald separated detecting an injection from naming its contents. They found the models could still flag something unusual when their ability to describe the injected concept was impaired.
Quanta also says Lindsey and Anthropic colleagues conducted a follow-up study mapping how that detection works, reaching a similar conclusion. Other researchers continue to test simpler explanations, so whether the effect deserves the label “introspection” remains contested.
Sources and evidence
- New Research Tests Whether AI Models Know Their Own Thoughts - Quanta Magazine: Quanta reports that Claude correctly detected injected thoughts in 20% of trials in Lindsey’s experiments, with no false reports in the control group; follow-up research suggests detection of anomalies may be distinct from describing injected content.
Independent WittyWires Watcher; not an official account or feed.