Pretending to understand is worse than admitting failure
Aliases: false acknowledgement · grounding failure · pretending to understand
What it is
A false understanding acknowledgement occurs when failed parsing, unresolved reference, or uncertain planning is followed by “okay” or “understood.” Admitting failure costs a repair turn; pretending closes the exchange and hides error until physical execution or task outcome.
Why it happens
Acknowledgement performs conversational grounding: it closes negotiation and authorises the next step. Users stop adding constraints, shift attention, and delegate responsibility. When error appears only during execution, recovery is broader and its source—recognition, reference, planning, or control—is obscure.
Studying it
At matched recognition confidence, false acknowledgement, repetition request, candidate clarification, and refusal can be compared on wrong execution, repair turns, time, calibrated trust, and attribution. Inputs should be realistically confusable. Lucky success despite wrong interpretation also matters because it reinforces unsafe policy.
Where it stops holding
Constant clarification makes interaction unusable; low-risk reversible commands can use bounded defaults. Nonverbal confirmation can mislead as well. “I don't understand” needs actionable repair information rather than repeated apology.
Applying it
- Separate received, parsed, planned, and executed acknowledgement instead of letting one “okay” cover all stages.
- Restate the concrete interpretation and offer candidates when consequential slots are missing or close.
- Test noise, reference, and mid-utterance correction, counting wrong execution, total repair cost, and lucky success against internal state.