Act, Confirm, or Stop? Smarter behavior for AI assistants, wearables & robots — Amit Desai, Roku

An assistant can make mistakes less costly without improving its recognition accuracy. This talk's worked example assigns different costs to playing the wrong song, asking for confirmation and declining an unclear request. Playing the wrong song costs more because the user must notice, stop it and ask again.

The proposed Outcome User Cost Heuristic chooses confidence thresholds according to those costs. Low confidence can trigger a request to repeat, intermediate confidence a confirmation, and high confidence immediate action. The thresholds depend on the interaction: showing choices on a television changes the burden of confirming compared with a voice-only exchange.