Implicit confirmation embeds the ack in the next move

What it is

Implicit confirmation does not open a separate “is that right” turn. It embeds what was heard inside the next sentence that is already doing the work. A campus-shuttle dispatcher says “okay, rerouting via the tunnel, five minutes to the west gate” while changing the route. If the rider hears that “tunnel” is wrong, they must cut in at once; if they do not, the system takes silence as consent. Confirmation has not vanished. It has become continue by default, stop me if I am wrong.

Why it happens

Implicit confirmation packs read-back and progress into one sentence and drops a floor exchange. The information structure is: already-filled slots appear as given information in a presupposition or adjunct; the new action is the focus. The listener has to run a verification scan while the dialogue advances — matching the embedded slot against what they just said. On a mismatch they still have to seize the floor before the system walks further.

So reliability does not hinge on whether the system spoke the word. It hinges on whether the user noticed and could interrupt in time. Embed too shallowly (“okay” then change the route) and there is no confirmation. Embed too densely (three slots in one fast synthetic sentence) and the scan cannot finish. Silence is rewritten: under explicit confirmation, silence is pending; under implicit confirmation, silence is a go. If people treat the next sentence as a progress broadcast and do not scan the slots, the error executes.

Studying it

Present the same misrecognised slot two ways: a standalone “is it X,” versus X nested in the next action sentence. Dependent measures: embedded-error detection (did the user interrupt and correct before execution), time to detection, and whether undetected errors actually executed. Independent variables: position of the nested slot (sentence-initial given, medial, final focus), speech rate, barge-in on or off.

Eye-tracking or a dual task will depress detection: while watching a road or a progress bar, an embedded proper name slides past more easily. Wizard-of-Oz can swap a neighbour value on the content turn (west gate / east gate) and count how many people stop the implicit sentence. Do not treat “no later complaint” as a detection success — many people notice only after the action has already run.

Where it stops holding

When the user cannot hear the synthesis, or when the next sentence is already doing something else (the map has started drawing, the number has started dialling), the action itself steals the verification scan and implicit confirmation decays into a notice. For hard-of-hearing listeners, non-native listeners, and noise, a proper name buried in an adjunct is the first thing to become unintelligible. If the implicit sentence executes irreversibly before it finishes, an interrupt that arrives has nothing left to confirm. Implicitly confirming every slot at once dumps a whole form’s verification onto one auditory glance.

Applying it

• Use implicit confirmation on low-risk moves that can still be corrected in the next breath: put one hearable slot in the next sentence, then continue. Park the slot in a prominent position, not a fast subordinate clause.
• The implicit sentence must be interruptible, and an interrupt must stop in a state that can still change — not “already executing, just listen.”
• A bare “okay” is a receipt, not a read-back. Do not count it as implicit confirmation.
• How to check: in test, nest a neighbour value in the next sentence and see whether the interrupt happens before execution. A correction that starts after execution has begun means the implicit confirm did not work.

Related

• Same group: M2.03.1 Explicit confirmation is reliable but slow · M2.03.3 Confirmation strength must match consequence
• Nearby: M2.08 Explicit and implicit confirmation · M3.04 Barge-in · M3.09 Barge-in and interjection
• Search terms: implicit confirmation · embedded acknowledgement · grounding in dialogue