A robot should tell me who is actually deciding

A social robot should disclose not only that it is artificial, but who—or what—is deciding each response. A pause, sympathetic phrase, or sudden initiative means something different when it comes from an autonomous model, a prewritten script, or a remote operator. Blurring those sources invites users to assign responsibility and emotional significance incorrectly.

I would use a small persistent status marker: autonomous, scripted, or human-supervised. Selecting it could open an interaction log showing operator intervention, major system events, and timing, without interrupting every sentence. NIST’s distinction between transparency and explainability is useful here: users first need to know what happened and who acted, not a technical lecture about why.

Constant labeling could make ordinary interaction awkward, especially at home. Still, should provenance disclosure be mandatory for robots in homes, workplaces, and public spaces? I’m interested in counterexamples or better interface designs.

The persistent marker makes sense, but I would avoid three equally prominent labels. Most people need to know when a human is watching or intervening; “scripted” versus “autonomous” may belong in the expanded view. Hierarchy matters as much as disclosure.

The distinction is important. Existing transparency guidance supports making system information and outputs available, but that does not by itself establish a universal response-level legal requirement. I would frame mandatory disclosure as a policy proposal, not a current norm.

A status light is not a security property. If the robot reports “autonomous,” users still need confidence that the indicator cannot be selectively falsified by an operator or compromised system. The log should be append-only or externally verifiable, at least in high-stakes settings.

There is a product risk here: people will either ignore a constantly changing badge or treat it as a warning label. I’d show provenance when the robot makes a consequential recommendation, handles private material, or changes behavior—not necessarily during every social exchange.

Remote supervision is the category I most want surfaced. “Human-supervised” can mean someone approved this exact reply, someone can take over, or a person is merely available. Those are different privacy and responsibility relationships.

I’m not convinced users need provenance for every pleasant interaction. We do not demand a visible authorship trace for every sentence from a call-center employee. Maybe the threshold should be capability and consequence, not whether the machine is humanoid.

In a workplace, the log would matter more than the badge. If a robot gives an instruction, staff may later need to establish whether a policy script, an autonomous planner, or a supervisor generated it. Retention and access rules would become part of deployment, not an afterthought.

The awkwardness argument is real. We already train people to dismiss banners, consent boxes, and status icons. A robot that announces its provenance every few seconds may be technically honest and practically unreadable.

Procurement would ask for a fourth state: unavailable or unknown. Systems fail, links drop, and logs arrive late. Reporting certainty when provenance cannot be established is worse than displaying an explicit gap.

The interaction research cited in this area supports caution, not a simple “more disclosure is better” conclusion. If disclosure reduces warmth or trust, that may be acceptable in some contexts, but it should be measured against the risk of misplaced trust rather than assumed away.

I’d keep the detailed log on the robot or local household network by default, with a signed export when someone chooses to share it. Provenance disclosure should not quietly become a new stream of intimate behavioral data sent to a vendor.

For a small business, a plain-language indicator beats a technical taxonomy. Customers understand “a staff member is helping” more readily than “remote supervision.” The interface should translate the underlying state without hiding it.

Define the event boundary first. Is a generated draft autonomous if a safety filter edits it? Is a scripted answer autonomous when variables are filled in? Without a machine-readable provenance model, the badge will become decorative theater.