Summary

An Economist interview with one of its own journalists about his reporting on whether AI models could be conscious. The framing is deliberately careful: the reporter's own answer is no, and most of the piece is about why the question is hard to test rather than about any claim that it has been answered. The most concrete material is a described Anthropic interpretability finding.

• Asking a model whether it is conscious is treated as worthless evidence: the reporter notes many would once have said yes and that they did not want to be switched off, and that this tells you nothing about what is happening inside.
• The reported Anthropic finding: asked to count to five and then introspect, Claude produced the expected output, but inside the network, between layers and invisible to the user, words surfaced - 'halfway' when halfway, 'countdown' while counting, 'conscious', 'cloud', and 'done' at the end. Described second-hand from the paper, not verified here.
• Anthropic's researchers are reported to have characterised this as a 'mental whiteboard' or workspace where the model works things through before producing output.
• The parallel drawn is to global workspace theory, one theory of human consciousness, in which information entering a shared workspace is broadcast to the rest of the brain and that is where awareness arises. The reporter is careful: a parallel, not evidence.
• Asked directly, the reporter says no model is conscious now, and that the Claude workspace 'might not end up being the most important bit' or might be one of a million. He describes the state of things as 'at the foothills of things that look a bit like consciousness, but aren't'.
• What he says changed in his own view is narrower than it sounds: not that machines are conscious, but that he is 'no longer confused by the idea of it' - he no longer treats silicon consciousness as incoherent in principle.
• On the labs' incentives he is even-handed, noting they might want the conscious answer because it looks futuristic, or want to avoid it because it raises awkward questions about how they treat the systems. He reports nobody at any lab said they were trying to build consciousness.
• The framing worth keeping is the two failure modes: prematurely granting rights and power to rule-following systems that do not merit it, versus accidentally creating something with moral worth and treating it badly. A philosopher is quoted as saying doing it by accident 'will be a moral catastrophe'.

Why it matters

This is the most disciplined treatment of machine consciousness the Handbook has logged, and it is useful precisely because its author concludes no. It gives a defensible position to hold - the question is no longer incoherent, the evidence is nowhere near sufficient - and a concrete finding to attach it to. The interpretability result is also independently significant: it says a model is doing work in places the chain of thought does not show, which connects directly to the depth-scaling interpretability worry raised elsewhere this batch. Two different routes to the same conclusion, that what is readable is not what is happening.

J5phvo3bqGU-transcript.txt