Human edits should be marked, or afterwards no one can tell which stretch came from whom

What it is

In the handed-over draft, the numbers were changed by a person, the argument written by a model. Three months later an audit asks “who is responsible for this sentence”; the file has no colour, comment, or author track. Marking human edits requires that a range a person touched can still be read afterwards. Without it, responsibility, training data, and merge rules lose their object.

Whether regenerate will cover edits needs to know where the edits are. Marks are the perceptual layer of that rule.

Why it happens

Artefact and patch share one text stream; the seam is invisible on the page. Human memory covers only a short editing session; after leaving the window or handing to someone else, “which stretch did I change” is gone. If the system does not record it either, authorship collapses at character grain into an anonymous mix.

Responsibility, copyright notices, and “human preference” used to update models all depend on that seam. With no seam, an error can only be assigned to the whole system or the whole person; training then treats human patches as model output, or model output as human demonstration.

Studying it

Have people edit several places; after a delay or a change of person, ask them to mark “which are human edits.” Compare no marks, inline highlight, a sidebar author track. Dependent variables: identification accuracy, responsibility judgments, whether regenerate can avoid human-edited regions. Independent variables: amount of editing, delay, whether the person changed.

The change-of-person condition is the hard one. Original authors can guess from memory; that is not after-the-fact distinguishability.

Where it stops holding

A one-shot private draft that is never sent and never trained has a weak external need for marks, though they still help regenerate avoid those spans. Tiny edits (one typo) can live on a character-level track without a loud highlight that interrupts reading. When co-editing already has an author track, generation should write in as one “author” on the same track, not a parallel system. This entry does not treat a threshold that keeps people from editing at all.

Applying it

• Record author at character or sentence grain: model / current user / someone else. Visible to editors by default; on send-out, optional to keep.
• Regenerate, export, and accountability views should all read the same marks, not each keep their own.
• Do not let “one-click polish” strip marks from human-edited ranges without a prompt.
• Check: a day later, have someone else open the file and ask “which sentences did a person change.” If they cannot point, the marks have no after-the-fact use.

Related

• Same group: L3.12.1 Whether a generated artefact is framed as a draft or a finished piece decides whether the UI should lean editor or viewer · L3.12.2 Regenerating after the user has taken over will overwrite human edits unless a merge rule is explicit · L3.12.4 The higher the takeover threshold, the more people settle for a result they dislike · L3.12.5 Editing a generated artefact can take more skill than writing from scratch, because the existing text must be understood first
• Nearby: L3.13 User Feedback Loops on Generation Quality · L3.04 Content Labeling
• Search terms: human edit marks · mixed authorship · who wrote this span