People act polite to AI without thinking it through — but does that prove they're not really thinking at all?

This explores whether people's consistent social behavior toward computers and AI, such as politeness, reciprocity or treating a device like a person, proves they're responding on autopilot. The corpus answers fairly directly: no. Observed behavior alone can't prove mindlessness, and treating it as proof is one of the field's known weak points.

The sharpest version of this argument is Lee's critique of the CASA tradition Can mindlessness in CASA effects be measured independently?. The usual reasoning goes like this: people behave socially toward a machine, they 'know' it isn't a person, so they must be doing it mindlessly. The trouble is that the social behavior serves as both the evidence for mindlessness and the thing mindlessness is supposed to explain. Without a separate measure of whether people are actually on autopilot, taken apart from the behavior itself, the claim can't be proven wrong. Any social response at all counts as confirmation. Behavior that changes in systematic ways shows that *something* consistent is driving it. It doesn't tell you whether that something is automatic or deliberate.

When researchers actually ask people, the picture gets more complicated. In interviews with owners of Nomi, a small dashboard robot for cars, most described engaging consciously and reflectively Do people interact mindlessly or mindfully with in-car AI?. They named it, dressed it up and gave it social agency, which are exactly the behaviors CASA would read as mindless, yet they did these things on purpose. The researchers suggest that as voices and faces get more convincing, the old split between mindful and mindless responses may stop being useful. People can choose to treat a machine socially while fully aware of what it is, much as people knowingly get absorbed in a novel.

There's a lateral thread worth pulling. Work on reading users' mental state from small behaviors like gaze, hesitation, typing rhythm and interaction speed Can AI systems read cognitive state from interaction patterns alone? points to the kind of independent measure Lee says is missing. If automatic processing really drives a social response, it should leave a separate trace, for example quicker and less hesitant reactions, that can be checked apart from the social behavior itself. The same caution applies to popular 'System 1' framings of human–AI interaction Why do people trust AI outputs they shouldn't?. They're useful for spotting traps like overtrust, but they too can slip into reading 'automatic thinking' backward from outcomes.

The corpus here is thin: two notes speak to the question directly and the rest is adjacent. Still, the takeaway is clear. 'People act socially toward AI' and 'people act mindlessly toward AI' are separate claims, and much of the research has treated them as one.