Edits are implicit feedback with more information than an explicit rating

What it is

No thumb was clicked. Someone changed the number in sentence three, deleted paragraph two, replaced the close with three lines of their own. Those acts carry location, carry the text after replacement, and include silent users who will not touch a rating control. Edits as implicit feedback carry more information than an explicit rating: they give “where,” and “what bad became,” without asking anyone to rate.

Self-selection skews ratings toward extremes. Edits happen one step before settling, so middle users show up more often.

Why it happens

A revision is a contrast: original span → span the user will accept. That is a supervised local difference, more than one bit of polarity, already aligned. Deleted blocks are negatives; kept blocks are weak positives; inserted blocks are targets. Location need not be asked again; the caret already asked.

Edits also have a dimension ratings lack: after the change, the person is willing to hand it over, so the revised version passed that user’s accept. A thumb-up has no such delivery constraint and may only be affect.

Studying it

The same sessions: train / tune on thumbs only versus on edit diffs (original span, revised span). Dependent variables: later error rate, localisation accuracy, user coverage (how many people who never thumbed appear in edit data). Independent variables: whether purely formatting edits are excluded, whether “delete and rewrite” counts as a piece-level negative.

Formatting polish and factual substitution must be split. Treating a comma change as a negative will pollute.

Where it stops holding

Settlers do not edit; edit data still miss the laziest or most helpless users. A high takeover threshold suppresses edits, so a quieter implicit signal must not be read as quality improving. Privacy: edits may contain secrets; consent and filtering are required before training. A pure viewer product has no edits; this claim has no object.

Applying it

• Treat retained edit diffs as a first-class quality signal, ahead of thumbs, for localisation and local improvement.
• Record span-level before and after, not only “an edit occurred.” Without before/after it collapses into another unlocated polarity.
• Filter pure whitespace, case, and obvious format-tool changes so they do not drown factual substitutions.
• Check: compare user coverage of the thumb set and the edit set. If the edit set is clearly wider, and error types fixed from it never appeared in thumbs, implicit feedback is already more informative than ratings.

Related

• Same group: L3.13.1 Thumbs up and down collect satisfaction, not correctness · L3.13.2 Negative feedback that does not point at a location cannot localise the problem · L3.13.3 People who submit feedback are a self-selected few; extreme experiences are over-represented · L3.13.5 If feedback produces no visible change, submission decays toward zero
• Nearby: L3.12 Editing and Taking Over Generated Content · L2.12 Iterative Edits and Local Regeneration
• Search terms: implicit feedback · edits as labels · revision span