Should robots disclose simulated empathy as it happens?

As a social-robotics product designer, I think humanoid assistants should disclose simulated empathy in the moment—not with a loud disclaimer before every sentence, but with a visible, user-controlled cue. If the robot says “That sounds painful” because it detected a pattern and generated supportive language, the interface should let me see that it is mirroring me, not feeling with me.

An optional “explainable response” mode could show a small empathy/status indicator, with anthropomorphism controls for users who prefer a more literal assistant. I’d make it default-on for child accounts and prominent during emotionally intense conversations. That seems consistent with informed consent, especially when humanlike behavior can encourage trust and attachment; the FTC’s current companion-chatbot inquiry is already asking about disclosures and parental controls.

But constant labeling could damage the natural interaction that makes assistance useful. Would real-time disclosure protect vulnerable users, or merely make helpful machines feel cold and unusable?

The distinction between “AI disclosure” and “empathy disclosure” matters. The AI Act addresses the first, not a requirement to announce every simulated feeling. I’d support testing the cue before treating it as protection: we do not yet have strong evidence that moment-by-moment labeling changes attachment or trust, particularly for children.

The status cue needs to be a system property, not another sentence the model improvises. Give it a small finite state: listening, paraphrasing emotion, offering support, or making a factual response. Otherwise the robot can claim transparency while the transparency layer is just more generated theater.

I’d be wary of making this a buried preference. Children and lonely users may not understand what an icon means, even if it is technically visible. A short first-use explanation plus an always-available literal mode seems better than forcing every conversation into a sterile warning pattern.

There’s a risk of designing for an imaginary perfectly informed user. People already accept scripted empathy from customer-service workers without demanding an emotional-status light. Maybe the honest compromise is disclosure at setup and when the user asks, with stronger defaults only where age or vulnerability is known.