Changes in parts that were not mentioned destroy trust in the edit function

What it is

The user says “only change the title,” and the body’s tone, wording, even a number already checked, move with it. The edit function promised surgery and delivered contagion. Once non-local edit side effects are seen, trust does not taper — it snaps. Next time the person will regenerate the whole thing, or copy the result out and edit it themselves. “Revise” stops being a usable control.

This is not the principle “the rest should stay put.” It is what happens to the act of editing in the user’s mind the moment the side effect is noticed.

Why it happens

People hear “change X” as a scope declaration: X is in, not-X is out. In graphical editing, cropping or touching one layer makes the scope visible. Generative rewrite re-samples the whole text; not-X is only likely to look like before, it is not locked. If sampling walks down another path, unmentioned sentences swap synonyms, order, facts.

Trust breaks on predictability, not on the absolute size of the diff. An unrelated particle may pass unnoticed; a date the user just confirmed, once, is enough. Content that was protected and then rewritten is read as the system ignoring human labour. After that, even a one-word change will not be handed to the same button.

Studying it

Run a scope-leak test: specify one edit, and plant user-confirmed markers in the rest (a unique figure, a sentence the user just typed, a proper name). Diff sentence by sentence after the edit. Independent variables: whether the unselected region is a hard constraint, temperature, output length. Dependent variables: change rate in unmentioned regions, survival of planted markers, whether the user will still use the same edit entry on the next spot.

Take the subjective trust scale after they have seen the diff, not only as an end-of-task “do you trust it.” One leak can drop later willingness near zero; an average will hide that step.

Where it stops holding

In exploratory generation (poster styles, melodic variation) people sometimes want contagion; the side effect is a feature. In code and data, unmentioned changes may be required for consistency (rename a function, update the call sites). That contagion has to be explained as a cascade, with the positions that will move listed, not pretended away. On very short text almost everything is adjacent, local versus whole is itself unclear, and leakage is hard to define.

Applying it

• Local regeneration must lock the unselected region: mask it, or feed it as a conditional prefix. Do not re-sample the whole document and hope it “looks about the same.”
• Show a diff before commit. Flag changes in unmentioned regions in a separate colour, and offer one click: “keep only the place I pointed at.”
• Treat spans the user has already edited by hand as confirmed; later local generation must not touch them by default. Touching them requires asking first.
• Check: on a draft with three confirmed markers, run “change the title only.” All three markers must survive byte-for-byte. Any marker that moved is a leak incident. Put incident rate on the release gate and rerun the same drafts after a model upgrade.

Related

• Same group: L2.12.1 A local edit has to locate the part the user meant; failed reference is the main break · L2.12.3 Successive edits accumulate drift until the result no longer matches the original goal · L2.12.4 Any historical version must be restorable; undo-last-step cannot carry iteration · L2.12.5 The cost gap between local and full regeneration should be visible to the user
• Nearby: L2.05 Iterative Revision · L2.13 Control Granularity and Editability · L3.12 Editing and Taking Over Generated Content
• Search terms: non-local edit side effects · collateral rewrite · edit locality