If people say 'thank you' to a computer on autopilot, how would you ever prove the autopilot part, not just the thank-you?

This explores the claim behind 'Computers Are Social Actors' research: that people are polite to computers because they are acting mindlessly, on autopilot. The question is what evidence could show that mindless state directly, as something separate from the politeness itself. The corpus has a sharp answer to the first half. Most of this research works backward. It sees people thanking a computer, concludes they must have been mindless, and then uses mindlessness to explain the thanking. Can mindlessness in CASA effects be measured independently? calls this circular, and therefore impossible to disprove. For mindlessness to explain anything, it has to be measured as a separate step in the chain, apart from the outcome it predicts. If the only sign of mindlessness is the social response, the theory can never be wrong.

So what could that separate measure look like? The corpus doesn't offer a ready-made mindlessness test, but it points to two kinds of signal. The first is how people behave while they interact, as opposed to what they end up doing. Can AI systems read cognitive state from interaction patterns alone? describes systems that track where people look, how long they hesitate, and how fast they move as ongoing signals of their mental state, without interrupting them to ask. Applied here, a person on autopilot should look different moment to moment: quicker responses, less hesitation, less scanning, compared with someone thinking carefully. If those signs predicted who acted socially toward the computer, mindlessness would finally be measured on its own terms. The second is a manipulation study. Can readers tell truth from fabrication without evidence signals? found that readers given no cues about a text's source accepted fluent fabrications as readily as true claims, then regained their judgment when shown evidence cues. Hand people something that should break the autopilot, and see whether the social response goes away. That is the experimental design a mindlessness account would need.

The idea becomes more useful once you notice that AI research keeps hitting the same trap from the other side. When a model passes a test of reasoning about other people's minds, did it actually reason, or did it match patterns? Can language models solve ToM benchmarks without real reasoning? shows that surface pattern-matching can pass these benchmarks, so a correct answer can't prove the mental process behind it. That is the CASA problem in reverse: you can't read a hidden state off the output it is supposed to explain. Can language models genuinely monitor their own thinking? reaches a similar point about whether models can monitor their own thinking. The test it recommends is whether a model's self-explanations hold up when you ask what it would have done differently, which goes beyond whether they sound plausible.

The most transferable standard comes from work on model introspection. Can language models actually introspect about their own states? argues that a self-report counts as real introspection only when you can trace a causal chain from the internal state to the report. For human mindlessness, that suggests changing the person's state first, for example with time pressure or distraction, then measuring that state independently, and only then checking whether social behavior follows. Asking people afterward whether they were on autopilot wouldn't settle it, because their answers have the same circularity problem. Can any falsifiable theory of consciousness apply to LLMs? adds a warning that applies here too: a theory of an inner state that only looks at outputs turns out to be trivial. Lee's critique makes the same point about mindlessness.

The broader takeaway: whether the subject is a person thanking a computer or a model passing a mind-reading test, claims about hidden mental states become testable only when a separate measure or a deliberate manipulation sits between the state and the behavior. Lee shows that most CASA research skips that step. The corpus names it as a gap, but it doesn't show anyone filling it yet.