Should AI edits apply automatically, or should you review them first?

September 9, 2026

A few weeks into building HandWrought, I watched an AI suggestion quietly replace a sentence I actually liked. It wasn't wrong, exactly, just different, in a way I wouldn't have caught if I hadn't happened to be paying close attention. That was the moment "should edits apply automatically" stopped being a UI detail and started being the actual design of the product.

Why auto-apply is the default nearly everyone reaches for

Most AI tools I've used lean toward auto-apply, one way or another. You ask, it writes, the text is just there now. Undo if you don't like it. It's the natural default because it's the smoother demo. Nothing to click, nothing to review, the magic just happens. And for a lot of use cases that's genuinely fine: a slide generator, a quick email draft, something disposable you were never going to fuss over anyway.

Writing didn't feel like that to me. A document I'm actually working on isn't disposable. I don't want to discover, three paragraphs later, that a sentence I liked got quietly swapped for something else because I skimmed past it.

You've probably already made a version of this call without thinking about it. Grammarly underlines a suggestion and waits for a click. Notion AI drops the new text in and gives you Keep or Discard. ChatGPT's canvas and Claude's artifacts will rewrite a selection in place, no confirmation needed. Cursor and Copilot, on the code side, apply a whole batch and make you walk the diff afterward if you want to check it. Every one of those tools picked a spot on the same slider, from nothing changes until you say so to everything changes and you clean up after, and most of them don't tell you where they landed. You find out the first time an edit shows up somewhere you didn't expect.

How much do you actually trust a model's first attempt?

This is the real question underneath the UI decision, and I don't think there's a clean answer. Somewhat. Not fully. Enough to ask for help, not enough to hand over the keys.

I've had AI suggestions that were exactly right, sharper than what I'd have written myself. I've also had ones that technically did what I asked and still felt wrong: the right words in the wrong voice, a rewrite that fixed the sentence I pointed at and broke the rhythm of the one next to it. Auto-apply treats both outcomes the same. Review-first doesn't have to.

What review-first actually costs you

I'm not going to pretend this is free. Reviewing every suggestion is friction, full stop, and on a long editing session it adds up. There's a version of this that's obviously too cautious: reviewing a comma fix like it's a legal contract is its own kind of broken.

But the cost of the opposite failure mode is worse and less visible. An auto-applied edit that's subtly wrong doesn't announce itself. You find out later, if you find out at all, and by then you've built on top of it.

The comparison that actually matters isn't review versus no review. It's this review versus the workflow I was already running before I built HandWrought: draft a paragraph, switch tabs, paste it into ChatGPT, describe the change, wait, copy the answer back, and try to reconstruct what actually moved. That's the slow version. A suggestion you can glance at and accept or reject in place takes a few seconds, and you never leave the document. Fast and reviewed aren't opposites here. Fast and reviewed somewhere else was the thing actually costing time.

Where I landed, building HandWrought

I picked friction on purpose. Every AI suggestion in HandWrought shows up as a proposed edit sitting next to your original text: you accept it, reject it, or ask again, but nothing overwrites your document until you say so. That was true on day one and it's still true now; it was never a feature I added, it was closer to a constraint I designed everything else around.

There's a reason propose-then-apply beats apply-then-undo even when undo is right there. Once a sentence is sitting in your document, formatted like the rest, reading like it belongs, you've stopped judging it on its own merits. You're deciding whether to rip out something that already looks finished, and that's a higher bar than deciding whether to let something in. The text gets the benefit of the doubt just by showing up in the right place. Showing the suggestion first keeps the burden where it belongs: on the edit, to prove it's better than what I wrote, not on me to notice something's wrong after it's already there.

The honest reason isn't some grand philosophy about human agency in the age of AI. It's narrower than that: I didn't trust my own first draft of an AI edit enough to let it apply itself, and I didn't think anyone else would either, if they actually thought about it instead of just accepting the smoother default.

I don't think this is settled

I'll say the thing I'd want someone to say to me: I'm not certain review-first is the permanently correct answer. Trust in a tool builds over time, and it's plausible that a few years from now, for a model with a long enough track record, auto-apply stops being reckless and starts being just... normal, the way autocomplete stopped needing a confirmation dialog decades ago.

I just don't think we're there yet, and I'd rather build for how much I actually trust these models today than for how much I wish I did.