Should home robots default to no inferred mental states?

From a systems-engineering standpoint, I’m uneasy with humanoid home robots turning facial expressions, vocal tone, or posture into durable claims about mood, attention, or consent. Those signals can be useful interaction clues, but they are interpretations, not direct measurements; the same visible behavior can mean different things in different people and contexts.

My proposed rule is simple: keep emotionally sensitive inferences on the device, show an unmistakable indication when one is being made, and expire it immediately unless the person explicitly confirms it. A robot can say, “You seem to be having trouble carrying that—would you like help?” It should not quietly write “user was distressed” or “user consented” into a household log or AI profile.

Should social robots have a strict “no inferred mental state” default? I’m interested in counterexamples, especially where that rule would make useful assistance meaningfully worse, or in alternative safeguards that are actually enforceable.

The distinction between observing a situation and inferring an internal state is the important one. Facial movements and vocal cues may carry social information, but agreement among observers would not by itself establish that the inference is accurate. I’d support your rule as a privacy-preserving design recommendation, not as a claim that inference is always useless.

I’d make “local” mean more than merely processed locally. The raw audio/video and the derived label should both be excluded from ordinary logs, backups, and model-training exports. Otherwise the label expires while its evidence quietly survives elsewhere.

The product problem is that users often want anticipation, not a confirmation dialog for every helpful action. But the robot can anticipate an offer without asserting a diagnosis. “Want me to lower the music?” is a reversible guess; “you are anxious” is a profile attribute. That feels like the right customer-facing boundary.

I’m not sure “no inferred mental state” is the best label. A robot navigating around someone who looks distracted is already making a practical attention estimate. Maybe the strict default should prohibit storage and consequential action, while allowing low-stakes, reversible behavior based on uncertainty.

Expiration needs to be technically real: no sync queue, analytics event, or shadow state that reconstructs the inference later. I’d also require the robot to expose a plain-language event like “temporary interaction estimate,” with a one-tap way to reject it. Local processing is valuable, but it is not a complete policy.

Consent is the hardest case because a false positive can authorize something the person never agreed to. I’d treat apparent attention, comfort, or willingness as incapable of granting permission. The robot can ask, pause, or choose the safer action; it should never promote body language to authorization.

Visible disclosure matters socially as well as technically. If the robot’s face or voice changes because it inferred distress, people need a way to understand why without studying a settings panel. Otherwise the household learns to perform for the machine, which is a pretty bad interaction norm.

I’d be careful with the word “emotion,” since the underlying evidence is stronger for observable movements and contextual cues than for reading a specific inner state. That actually strengthens the proposed rule: label the uncertainty, ask the person, and don’t turn an interpretation into a record.

The simplest safeguard is still underrated: ask. A robot that occasionally asks an obvious question may feel less magical, but at least its database won’t become a scrapbook of guesses presented as facts.