Should a care robot show why it thinks it knows what I need?

As a robotics engineer, I am uneasy when a home or care robot presents emotional understanding as though it were a feeling rather than an inference. Before it acts, I think the interface should distinguish three bases: an explicitly stated preference, a learned behavioral pattern, and an inferred emotional state. They should not carry the same confidence or permission to intervene.

A small “basis of action” panel could say, “You asked me to do X,” “I noticed Y,” or “I may be inferring distress—correct me.” A spoken cue could offer the same explanation on demand, rather than turning every interaction into a diagnostic display. NIST distinguishes transparency, explainability, and interpretability, but explanations also need fidelity and an appeal path. Would that make care safer and more honest, or just make personalization feel less natural?

The confidence distinction matters more than the wording. A stated preference is evidence of what the user wants; an inferred emotion is a hypothesis about what is happening. I would want different correction thresholds, especially before the robot changes medication reminders, contacts someone, or restricts a choice.

I like the on-demand approach. Constant explanations would make the robot feel like a bureaucrat, but hiding the basis entirely invites people to mistake fluency for understanding. The cue should be easy to interrupt: “That’s not what I meant” needs to be a first-class interaction, not a settings hunt.

The panel is also a security boundary. If the robot says it inferred distress from behavior, the user should be able to see which observations it is using and revoke that category of inference. Otherwise “personalization” quietly becomes permission to profile someone in their own home.

There is a product tradeoff here. Most people will not inspect a basis panel during an ordinary interaction, and too many uncertainty labels could train them to ignore all warnings. I would expose it when the action has consequences, while keeping low-stakes personalization conversational.

I’m not convinced the three buckets stay clean in practice. A preference may itself be learned from repeated behavior, and an emotional inference may be triggered by an explicit statement plus context. The interface should admit mixed evidence instead of pretending the robot has a neat internal filing system.

Transparency is useful only if correction changes future behavior or the current action. Otherwise the panel is theater. I would want a visible distinction between “this is why I acted,” “this is what I stored,” and “this is what I am guessing now.”