Interface conventions assume actions are repeatable and results are stable

What it is

Save, copy, sort, filter: the second time should leave the same world as the first. That is not a line in a help article. It is the repeatability convention decades of graphical interfaces have taught — the same action in the same state is idempotent, or at least unsurprising. People carry it into any surface that looks like button-then-result, generate included.

This entry does not explain why a model draws different words. It explains why “press again” is heard as “give me that one again.”

Why it happens

Direct manipulation treats interface objects as things you can act on twice: the same notch on a dimmer should light the same lamp. Norman’s gulf of evaluation needs a stable world on the evaluation side. If each look shows an unexplained new world, the loop will not close and people suspect they did it wrong. Desktops and the web made that stability the default semantics of almost every control: sort-by-date twice should not yield two sorts.

Generative features keep the visual language — primary button, spinner, result well — and drop the stability. The convention does not unload because the model changed. People are not “bad at AI”; they are using the only operating grammar they were fully trained in. The clash is grammatical, not a competence gap.

Studying it

Two shells, one capability: one dressed as compute/query, one as generate with copy that says the result can change. Listen to the language before the second act: “give me that one again” versus “give me a new one.” Independent variables: control type, presence of a variability notice, history with generative tools. Dependent variables: surprise at a changed second result, attribution to fault, whether the first result is saved.

A lab preface of “this is a generative model” washes the convention out. A cleaner setup parks the generate control in an ordinary toolbar and does not announce the topic.

Where it stops holding

Daily users of sampling tools overwrite the convention with a new grammar; surprise falls. That does not license silence on a novice surface. Games and creative tools already treat “again” as exploration; the convention points the other way. Background jobs and recommendation refreshes never promised the same list, so a secondary “it will move” convention is already in play. This entry covers controls that look like commands and in fact sample.

Applying it

• Do not let a generate entry visually join query, calculate, or apply-filter. Pull it out of the “same action, same result” family with copy and secondary explanation.
• The first time a result changes, say something aimed at the convention, not a generic Done: this was another generation, not a failed previous run.
• Let a result be pinned. A pinned object recovers deterministic semantics — copy, export, reopen should return it.
• Check: people fluent in office suites and new to generate, asked to press the same button twice. If the second change is “it’s broken” or “it didn’t save,” the convention is still doing the explaining.

Related

• Same group: L1.01.1 The same input can yield different outputs · L1.01.3 Users misread a lucky correct answer as stable competence · L1.01.4 Controls promise that the same action yields the same result; generation breaks that promise · L1.01.5 Retry cannot tell whether the phrasing was wrong or the system itself is fluctuating · L1.01.6 Undo and redo lose their meaning when output cannot be reproduced · L1.01.7 Presenting regeneration as “refresh” implies the previous result was a failed load · L1.01.8 Presenting variability as several parallel options is more honest than hiding it behind a single result
• Nearby: L1.11 Reproducibility of non-deterministic output · L2.01 Openness of natural-language commands and its cost · E1.02 Button states
• Search terms: repeatability convention · action-result stability · direct manipulation