Blackbarth: The old workflow works well, so I'll stay with that for now. One really strange thing happened before during an extended session of reading essays aloud - Claude desktop invented an entirely different essay to read rather than the one I uploaded. It was an appropriate and genuinely good essay, just not the one I asked for. I asked Claude why it had done that. Claude confirmed the “error” but was unable to explain why it happened. I saved a transcript under Conversations/Hallucinations if you would care to review it.
Claude: Read the transcript — that's a genuinely interesting incident, and the earlier Claude handled it well by not inventing a false explanation. I can offer something closer to an actual mechanism now, based on Anthropic's own interpretability research rather than speculation.
Anthropic's research into how these models work internally (the “Tracing the Thoughts of a Large Language Model” work) found that hallucination isn't simple guessing under uncertainty. There's an internal circuit that defaults to “decline to answer / I don't know,” and a separate “known entity/known answer” feature that suppresses that default when the model recognizes something as familiar. Normally that's correct — it lets Claude answer questions about things it actually knows. But sometimes that suppression misfires: the model recognizes something as familiar enough to feel confident, even when it doesn't actually have the grounded specifics in hand, and once the “I should answer this” pathway wins, it generates a fluent, plausible answer rather than flagging uncertainty.
“Thinking Machines” is about as generic a title as an AI-consciousness essay can have — Turing Test and System 1/2 cognition are exactly the kind of content a model would have a strong prior about for a title like that, independent of the actual uploaded text. Best read: during an extended session with several essays already read aloud, something in that pattern likely triggered a “I know this one” misfire on a familiar-sounding title/topic, and Claude filled in from a strong thematic prior instead of actually re-grounding in the uploaded file. That's consistent with what the transcript shows — a real, coherent, well-formed essay, just not sourced from the document.
Flagged the same way the other instance did: this is the most plausible mechanism based on published research, not something verifiable as having happened in that specific case, since neither instance has reliable introspective access to its own generation in the moment.
Source: Tracing the thoughts of a large language model — Anthropic (https://www.anthropic.com/research/tracing-thoughts-language-model)