[Question] Should AI interaction have an "undo" button?

[gloomcheng]

Your Question

Should AI interactions support "undo" — and what would that even mean?

In traditional UI, undo is straightforward: Ctrl+Z reverts the last action. The system returns to a previous state.

But in AI interaction, what is "undo"?

• Undo the AI's last response? (But I've already read it, can't unread)
• Undo an action the AI took? (If it sent an email, can we unsend?)
• Undo my last prompt? (Retry with different wording?)
• Undo the entire conversation branch? (Go back to an earlier fork?)
• Undo AI's "learning" from this interaction? (Don't adapt based on this)

Context

I'm asking because undo is one of the most fundamental UI patterns — it enables exploration by making mistakes cheap. Jakob Nielsen lists it as a core usability heuristic.

But in AI interaction:

• Responses aren't deterministic (same input ≠ same output)
• Actions may be irreversible (sent messages, API calls)
• Context accumulates (the AI "remembers" what was said)
• Real-world effects may occur (files changed, meetings scheduled)

If we can't provide undo, we lose a key affordance that makes interfaces feel safe to explore.

What have you already considered?

Some possible approaches:

1. Preview before action

Instead of undo, show what will happen before it happens.

AI: "I'll send this email. [Preview] [Send] [Cancel]"

Prevents the need for undo by requiring confirmation.

2. Branching conversations

Like Git branches — you can always go back to a previous point and try again.

Conversation at 10:00 ──┬── Branch A (current)
                        └── Branch B (alternative path)

3. Action journaling

Keep a log of all AI actions with reversal options where possible.

[10:05] Created file: config.json [Undo: Delete file]
[10:06] Sent Slack message [Cannot undo - already delivered]

4. Probabilistic replay

Accept that regenerating a response gives different output, and frame this as a feature ("try again" rather than "undo").

5. Scope limitation

Only allow AI to take reversible actions, require human execution of irreversible ones.

Who is asking?

• Human-AI collaboration

---

💬 What do you think?

• Is "undo" the right mental model for AI interaction?
• What patterns from version control (branching, commits, diffs) might apply?
• Should irreversible actions simply be off-limits for AI?
• How do you currently recover from AI interactions that went wrong?