Frequent low-risk confirms train a blind yes

What it is

If the same “shall I continue?” lands on cheap actions dozens of times a day, the answer becomes a ritual instead of a check. People train an unchecked “yeah,” and that “yeah” later attaches to a rare expensive action — “call your ex?” eats the same token. The wording of the confirm is not the bug; frequency trains confirmation into a reflex. Extra asks on low-risk actions look safer and actually take the gate off high-risk ones.

Why it happens

When a confirm almost never marks real risk, the turn is filed as a skippable formality. Attention leaves the echoed content; a polarity token is dropped at the end. Latency shrinks below lexical access, which is evidence the check has stopped. Speech has no visual “uncheck this box” brake, so habituation runs further. If the same yes/no channel serves both skip-track and place-call, the reflex transfers: the channel is shared, the risk is not.

Matching consequence by single-action labels, ignoring how often that label appears, fills the low-cost cell with “just in case” confirms. Each sentence is reasonable alone; together they train “yes” as the default key. The one time a check is needed, checking is no longer what the person is doing.

Studying it

Habituate people on high-frequency cheap confirms, then insert one action with different consequence, and count how many emit “yes” without checking content. Dependents: miss rate on the insert, latency of “yes” on the confirm turn (shorter than normal lexical access looks like reflex), and barge-in rate on confirms.

On the log side, plot confirm-turn latency per skill. A peak glued to the endpoint-detector floor means checking has likely stopped. Co-occurrence of “confirms per day” with “said yes and later found the error” is more informative than confirm coverage as a safety metric.

Where it stops holding

In banking or medicine every confirm is institutionally real; habituation still happens, but it cannot be fixed by deleting confirms — change the form (repeat the amount, not yes/no). Rarely used skills never form the reflex. Click-through agreements are the visual version of the same mechanism: the channel is not the root, repeated low-information confirms are. New users still check the first few times; the habituation curve has a start, and labs that only measure trial one will understate the field.

Applying it

• Take yes/no off high-frequency cheap paths: skip track, volume, lights. Do the act, speak the result, undo if needed.
• Reserve spoken yes/no for truly rare or truly expensive acts, so the question is surprising again. If a confirm must stay, change its form: repeat a number or a name; do not reuse the skip-track “yeah.”
• Do not patch a cheap skill's safety by adding a confirm; that patch's frequency will contaminate the expensive skill's channel.
• How to check: median latency of “yes” on confirm turns. If it is already shorter than recognizing and understanding one syllable, checking has stopped. Insert a confirm whose content was quietly changed; misses mean the ritual is in place.

Related

• Same group: M2.09.1 Consequence is reversibility and cost together, not one severity knob · M2.09.2 Reversible actions can replace a pre-confirm with undo · M2.09.4 Actions that affect other people escalate on their own
• Nearby: M2.03 Confirmation strategies · M2.08 Explicit and implicit confirmation · M2.01 Information density of prompts
• Search terms: habituation · ritual yes · confirmation fatigue