Editable AI output

Problem Context

• An AI system has produced text, table content, code, message copy, policy analysis, customer response, document summary, or another draft artifact.
• Users can improve the generated output directly instead of only changing the original prompt.
• Some generated claims are grounded in citations, retrieved sources, tool outputs, or conversation context that may no longer support the edited text.
• The final output may be copied, saved, applied to a record, sent to another person, or used in a higher-risk workflow.
• Reviewers may need to distinguish model-generated content from human-edited content after the edit is complete.

Selection Rules

• Choose editable AI output when the primary object is generated output after creation and users need to revise, review, save, copy, or apply that output.
• Use prompt box when users are still composing the request before generation rather than editing the generated result.
• Use inline edit when the object is an ordinary record field or table value with no generated provenance, source mapping, citation preservation, or regeneration boundary.
• Use streaming response when users mainly need to monitor partial generated output as it arrives, before the answer has settled into an editable draft.
• Use citation display when users need claim-level source previews; editable AI output should link edited claims to citation display when claims are source-backed.
• Use source grounding display when the task is whole-answer evidence coverage rather than editing one draft.
• Show generated draft, user edited, source mapping, citation preservation, unsupported edit, tracked change, stale source, unsafe edit, review required, and conflict edit states before apply.
• Require deliberate accept edit, reject edit, save draft, copy edited, undo to generated, regenerate from edit, and apply output actions with clear consequences.
• Do not make the generated answer look final while silently accepting edits, dropping citations, rewriting the prompt, or applying output without review.
• Do not let regeneration overwrite manual edits unless the user can compare, keep, discard, or recover the edited version.

Required States

• Generated draft state with original generated content, creation time, model or run reference, and source coverage visible.
• User-edited state that marks changed text separately from original generated text.
• Source mapping state that shows which edited claims still map to citations, retrieved sources, or tool outputs.
• Citation preservation state for unchanged or still-supported claims.
• Unsupported edit, unsafe edit, stale source, and source drift warning states.
• Tracked change state with accept edit and reject edit controls.
• Review required state that blocks apply output until unsupported, unsafe, or stale content is resolved.
• Partial edit and selected-range edit states for revising only part of a generated answer.
• Conflict edit state when regeneration, collaboration, policy update, or source refresh conflicts with manual edits.
• Save draft, copy edited, undo to generated, regenerate from edit, and apply output states.
• Mobile editor state that preserves source access, review status, edit actions, and final apply controls in constrained space.

Interaction Contract

• The generated draft, user edits, tracked changes, source mappings, citations, review status, and final output state are modeled as distinct data rather than one mutable text blob.
• Editing text does not silently update the original prompt, erase the generated baseline, or imply that citations still support changed claims.
• Accept edit commits a tracked user change into the draft while retaining its author and evidence state.
• Reject edit restores the affected range to the generated draft or previous accepted text.
• Undo to generated restores the original generated version or a selected range without deleting saved versions.
• Regenerate from edit uses the current accepted draft as context and warns when regeneration may conflict with manual edits.
• Apply output is unavailable or blocked when unresolved unsupported edits, unsafe edits, stale sources, or conflicts remain.
• Copy edited and save draft preserve visible review warnings or metadata when the destination requires provenance.
• Source mapping opens the relevant citation, source grounding, tool output, or evidence panel for the edited claim.

Implementation Checklist

• Store generated baseline text, edited draft text, selected ranges, tracked changes, author, timestamp, source mappings, citation IDs, review status, and apply state separately.
• Represent unsupported, unsafe, stale, conflicted, accepted, rejected, copied, saved, regenerated, and applied states as explicit values.
• Keep citation preservation logic claim-aware so unchanged claims keep citations and changed claims require revalidation.
• Connect edited claims to citation display, source grounding display, tool-use visibility, and audit records where those are the actual evidence surfaces.
• Require review before apply when an edit changes factual claims, removes qualifying language, changes numerical values, or alters policy-sensitive text.
• Add compare controls for generated draft versus edited draft and for edited draft versus regenerated output.
• Protect manual edits from accidental overwrite by regeneration, source refresh, collaboration updates, browser back, or mobile navigation.
• Test long generated text, selected-range edits, pasted content, citation deletion, unsupported claims, stale source warnings, review gating, keyboard editing, screen reader status, copy behavior, and mobile wrapping.

Common Generated-UI Mistakes

• Letting users edit a final-looking AI answer without generated-versus-user-edited status.
• Dropping all citations when one paragraph is edited.
• Preserving citation markers on claims that no longer match the cited source.
• Treating answer editing as prompt editing and hiding the generated baseline.
• Allowing apply output before unsupported edits, unsafe changes, stale sources, or conflicts are reviewed.
• Regenerating over manual edits without compare, warning, version recovery, or undo.
• Using ordinary inline edit for generated content that needs provenance and source review.

Critique Questions

• Can users tell which text is generated draft, user edited, accepted, rejected, unsupported, unsafe, stale, or ready to apply?
• What happens to citations and source mappings when users change a factual claim?
• Can users undo to generated, compare tracked changes, save a draft, copy edited text, and regenerate from an edit without losing work?
• Which edits block apply output and who can clear review required states?
• Does the surface explain when a source is stale, permission-limited, or no longer supports an edited claim?
• Would prompt box, inline edit, streaming response, citation display, or source grounding display better match the task?