Home humanoids need a social autonomy budget

I’m increasingly convinced that home humanoid robots need a user-configurable “social autonomy budget.” The important boundary is not whether a robot can anticipate a need; it’s whether anticipation gets treated as permission to approach, speak, infer mood, follow someone, or retain context.

I’d separate controls for movement, conversation, sensing, and recording, rather than hide everything behind one autonomy slider. A robot could be allowed to navigate toward a person but not start talking, for example. Any action based on inference rather than a direct instruction should also trigger an unmistakable indicator, much like a dedicated listening or mute signal. That fits the broader goals of control, communication, and predictability without pretending the robot understands consent.

Would you prefer one global privacy switch, or fine-grained controls for each kind of social autonomy?

Fine-grained controls, but with a genuinely immediate global cutoff too. The two settings solve different problems: configuration for ordinary use, and a fast physical escape hatch when the user no longer wants sensing or movement. A software-only promise would be the weak point.

The product risk is configuration fatigue. Most people will not tune four dimensions before breakfast. I’d ship a simple default profile, then expose movement, conversation, sensing, and recording separately for households that care enough to customize.

The inference indicator matters more than it sounds. “I heard a command” and “I guessed you wanted company” are socially different events. If the robot makes that distinction visible and consistent, people can form a norm around it instead of reverse-engineering behavior.

I’d also want the indicator to be locally visible, not only logged in an app. Otherwise the person being approached has no practical way to know what happened. The FTC’s emphasis on physical controls, clear listening indicators, and reviewing or deleting recordings points in the right direction.

There’s a hard usability tradeoff here: a robot that asks permission for every small action may become furniture. I’m not against budgets, but “helpful anticipation” needs some permitted zone or the system will train users to disable the safeguards entirely.

Worth keeping the claim narrow. NIST’s Privacy Framework is voluntary, and its language around control, communication, and predictability supports this design direction; it does not prescribe a social autonomy budget. That distinction will matter when people start calling this a standard.

One physical switch, four software switches, and a big light. We have somehow reinvented permissions for a machine with legs.

I’d make the budget exportable and inspectable, not just a collection of vendor-specific toggles. If the robot cannot explain which permission allowed an inferred action, the interface is decorative. Local logs would at least make the setting testable when the network is unavailable.