Weigh the cost of clarifying against the cost of undo; if undo is cheap, do not ask

What it is

Renaming a local file undoes in one step. Asking “are you sure you want this name” first swaps a cheap inverse for a dearer round of talk. Booking a non-refundable flight does not undo; not asking is what is dear. Whether to follow up is a comparison of clarification cost and undo cost, not a test of whether the sentence is vague. Cheap undo, act first. Dear undo, ask first.

This pulls follow-ups out of linguistics and into operations: if an inverse exists, if the user can find it, if it finishes in one step, the ask is repeating a safety net that is already there.

Why it happens

Willingness to sit through a follow-up is bound to “if this is wrong, can I cheaply get back to a moment ago.” In graphical editing, a wrong drag undoes; no one dialogs before every drag. Generative and agentive acts often delete undo: a button named Regenerate overwrites the last copy, a sent message has no recall, a permission change needs another approval path. The UI looks changeable; the inverse is not on the same path, so clarification is treated as the only brake.

The comparison has to be made for the same person at the same moment. Designers’ internal knowledge that “undo is one click” may not be a click the user can find. An unfindable undo has infinite cost, and asking becomes reasonable again — until undo is made into a step that can be found.

Studying it

Configure two grades of inverse for the same ambiguous instruction: one-step visible undo / undo hidden or unavailable. Cross that with whether a follow-up is asked. Dependent variables: total latency, how long an error stays, rate at which follow-ups are skipped, which safety net people later say they wanted. The critical contrast is whether “one-step undo plus still asking” only adds latency and does not cut error dwell — if so, the ask is repeating the net.

Undo must be operationalised as a step the user can finish alone, not as an experimenter rolling back offstage. Otherwise the study measures system reversibility, not human reversibility.

Where it stops holding

When undo exists but has a time window (mail recall for a few minutes), compare costs inside the window; outside it, treat the act as irreversible and ask. When undo leaves a trace (the other party saw the recall, the log still holds a permission change), social cost may exceed one dialogue turn; do not judge on technical reversibility alone. In simultaneous editing, your one-step undo may be someone else’s lost data; cost is computed on the shared side. Legally required confirmation cannot be replaced by “we have undo.”

Applying it

• For every act that currently asks, write two numbers: how long one ask takes, how many clicks undo takes, and whether the world is restored. If the undo path is shorter, drop the ask and make undo a first-class button beside the act.
• Undo must sit beside the result that just happened. A timed toast is fine; a control that lives only deep in a menu is not.
• When the inverse cannot restore (already outbound, a non-refundable charge), do not dress “generate again” as undo, and put the ask or a confirm back on that class of act.
• Check: pick an act with one-step undo (local rename, in-draft replace), turn follow-ups on for a week, then off with undo exposed for a week. If error dwell does not rise and total latency falls, the ask was spare. Then pick an irreversible act as a control: if turning the ask off produces a world change that cannot be restored, the ask should not have been dropped.

Related

• Same group: L2.15.1 Assuming one reading and running with it is faster, and it also enlarges the cost of being wrong · L2.15.2 A follow-up is worth asking only when the ambiguity would lead to clearly different outcomes · L2.15.3 Too many follow-ups turn the exchange into a questionnaire and cancel the efficiency natural language was for · L2.15.4 Leading with a result and stating the assumption is a third path between asking and running
• Nearby: L4.05 Interruptibility and Rollback · L4.11 Pre-action Confirmation and Irreversible Operations · L1.01 Mismatch Between Stochastic Output and Deterministic UI
• Search terms: clarification vs undo cost · reversibility · ask-versus-undo