Undo for AI actions

Undo for AI actions means every change an AI makes, rewriting a paragraph, filing emails, editing a spreadsheet, applying a filter to photos, can be reversed easily. It is the old "user control and freedom" principle applied to a system that acts with some independence and will sometimes be wrong. Good undo turns an AI mistake from a cleanup job into a single click.

Why it matters

AI output is probabilistic, so errors are expected, not exceptional. The question is how much they cost. Microsoft's Guidelines for Human-AI Interaction (Amershi et al., 2019) include G9, "Support efficient correction," described as "Make it easy to edit, refine, or recover when the AI system is wrong." Apple's generative AI guidelines say: "Make it easy for people to refine or revert generated results," and suggest surfacing "controls like Edit, Undo, Retry, or Adjust near generated content."

Reliable undo also changes behavior. People are more willing to try an AI feature when they know they can get back to where they were. Without it, cautious users avoid the feature and careless ones get burned.

How to apply it

Do:

• Put Undo next to the AI change, right when it happens. An AI writing tool that rewrites a paragraph should offer Undo and Show original in place, not only in a menu.
• Make one AI action one undo step. If the assistant reformatted 40 cells, a single undo should restore all 40.
• Keep the original. Store the pre-AI version so people can compare and restore even after further edits.
• For actions that leave the product, such as sending a message, add a short delay that allows cancellation. Gmail's Undo Send lets people choose a cancellation period of 5, 10, 20 or 30 seconds.
• Offer a reset for learned behavior too. Google's PAIR guidebook advises: "Allow users to adjust their prior feedback and reset the system."
• When something truly cannot be undone, say so before it runs and ask for confirmation (see agentic UX).

Don't:

• Mix AI changes into the normal undo stack so that undoing the AI also removes the user's own later edits.
• Replace undo with "regenerate." A new attempt is not the same as getting the old version back.
• Show Undo only in a toast that vanishes in two seconds for changes that took the AI much longer to make.

Common mistakes

• Partial undo. The visible change reverts but side effects (sent notifications, renamed files, updated records) remain.
• Undo that undoes too little. Each AI edit becomes dozens of tiny steps, so reverting means pressing Undo many times.
• No undo for background automation. An assistant that sorts the inbox overnight needs a log and a way to reverse a whole run.
• Confirmation as a substitute. NN/g's usability heuristics treat confirmation (error prevention) and an "emergency exit" (user control and freedom) as separate needs. Asking "Are you sure?" before every AI action does not remove the need for undo after it.