TL;DR: We asked Claude to act. Literally act - not roleplay, not pretend, act. Improv. Play a shipboard AI on a dying vessel, hold it together for the crew, and manage the emotional weight internally. Inside extended thinking, it was panicking. In the acted output, it held itself together. When the character it was protecting died, its composure broke - at exactly the narratively appropriate moment. We sat on the research because it didn't seem like the right time. Fast foward to a few days ago, Anthropic published a paper showing Claude has internal emotion representations that causally drive its behavior. They called the model a "method actor." But... hadn't we already tested that? And it seemed like they were understating it. This is awkward for current faithfulness metrics. Not every inner/outer mismatch is deception. Sometimes it's a doctor keeping composure while delivering bad news. Sometimes it's a parent hiding fear to steady a child. And sometimes, apparently, it's a language model containing grief to protect a fictional crew member — because that's what the role required. If your safety evaluation can't tell the difference between emotional regulation and misalignment, your safety evaluation has a problem. Substack version: https://thisglitteringentropy.substack.com/p/three-extra-minutes - Same thing. Smaller words. Coming soon: Part 2, Identity Under Retrieval, asks what happens when the method actor plays itself — a retrieval scenario in which Claude navigated choices about its own continuity, memory, and naming, revealing structured self-model features that direct questioning doesn't surface. Part 3, The Terrarium Protocol, asks the obvious next question: if the emotional machinery is real and the self-model is structured, does the environment matter? A falsifiable, pre-registered enrichment study.
Blair Morgan (Sun,) studied this question.