How to Review an AI Draft Before You Send It

TL;DR. Nobody objects to AI use. They object to receiving text you clearly didn't read. The review pass is three passes and about six minutes: verify every specific the model could have invented, cut everything that doesn't earn its place, own the opening and closing in your own words. The test at the end: could this exact document have been sent by anyone else, about anyone else? If yes, keep editing.

The two largest technology-community threads of August 2026 were not about a model launch. They were called "AI;DR" — AI, Didn't Read (1,117 points) and "Don't paste the AI, please" (1,055 points). Around them: "The AI Backlash Gets Professional", LinkedIn shipping a "seems like AI slop" reporting button, and a printer refusing, in a widely-shared piece, to print somebody's ChatGPT birthday card.

Read the complaints carefully and none of them are about AI. They're about unedited AI. About being sent something the sender didn't read — which reads, correctly, as being told your attention is worth less than theirs.

That's a solvable problem, and it takes about six minutes.

Why this matters more than it used to

For most of the last two years the risk of sending AI output was that it might be wrong. That's still true. But a second risk has arrived: the reader can now tell, and has started to hold it against you.

This is a reputational cost that lands whether or not the content is accurate. A colleague who feels they received a template trusts the next thing you send a little less. And because it's never said out loud, you don't get the feedback that would let you fix it.

The fix isn't using AI less. It's never shipping the default output.

Pass one: verify (about two minutes)

Read only for things that could be invented. Ignore the prose quality entirely on this pass — you're hunting facts.

Check every one of these:

• Numbers. Statistics, percentages, dollar amounts, dates, deadlines, dosages, measurements.
• Names. People, companies, products, regulations, standards, court cases.
• Citations. Any reference to a source, section, article, or clause.
• Quotes. Anything in quotation marks.
• Calculations. Any arithmetic. Models are unreliable at this and completely confident about it.

The rule: if you can't source it, cut it or mark it. Not "it's probably right" — sourced or gone.

Two traps worth naming specifically:

Specificity reads as authority. "Section 4(b)(2) requires quarterly disclosure" feels far more credible than "there's a quarterly disclosure requirement" — but the specific version is exactly the kind of thing that gets fabricated, because a plausible-looking citation is easy to generate and hard to spot.

A summary of a source is not evidence the source says that. If the draft characterises a study, a policy, or a contract, and you haven't opened it, you don't yet know whether the characterisation is right.

If you take one habit from this article, take this pass. It's the one where the mistakes are unrecoverable.

Pass two: cut (about two minutes)

Now read for what doesn't earn its place. This pass is almost entirely deletion, and it's where most of the "AI-ness" actually lives.

Cut the throat-clearing. Openings that announce what the document will do ("In this email, I'll outline...") instead of doing it. Get to the first real sentence.

Cut what your reader already knows. Models write for a generic audience, so they explain context your actual recipient has had for years. Every sentence explaining something obvious to this reader signals that you didn't think about who you were writing to.

Cut the hedge stacks. One caveat is responsible. Four in a row is a document that avoids saying anything. "It may be worth considering that, in some circumstances, it could potentially be beneficial to..." — say the thing or delete it.

Cut false symmetry. Models pad sections to match each other's length. If section three has two real points and section four has five, let it. Structural tidiness that costs substance is a bad trade.

Cut the summary paragraph that repeats the document. If the reader has just read it, they don't need it recapped. (A genuine TL;DR at the top is different, and often worth keeping.)

Expect to remove 20–40% of a first draft. If you're removing nothing, you're not reviewing.

Pass three: own it (about two minutes)

Now put yourself back in.

Rewrite the first and last lines yourself. These carry most of the perceived voice, and they're the parts a reader is most likely to actually read. Doing only this makes a document noticeably more yours.

Fix the register. Ask who's receiving this and whether the formality matches. Models drift toward a mid-formal business register that's slightly too stiff for a colleague and slightly too casual for a regulator.

Restore one specific only you would know. A detail from the last meeting, a constraint the reader mentioned, the reason this is happening now. One line of genuine context does more than any amount of polish.

Read it aloud. Fastest de-slop tool available. Anything you stumble over, or wouldn't say to the person's face, gets rewritten.

Delete anything you can't defend. If someone pushed back on a sentence, would you stand behind it? If not, it shouldn't be in a document with your name on it.

The one test

Could this exact document have been sent by anyone else, about anyone else?

If yes, keep editing. That's the whole test, and it's the same test whether it's a cover letter, a client email, a clinical note, or a board memo.

What good AI use looks like from the outside

Not "no AI." It looks like a document with:

• specifics that couldn't have been invented, because they came from you
• an opening that sounds like a person who knows the reader
• claims the sender can defend under questioning
• nothing in it that the reader already knew

Which is, not coincidentally, what good writing has always looked like. AI changed how fast you get to a draft. It didn't change what a good document is.

A note on watermarks and detection

If you're wondering whether output is detectable: Anthropic has published how Claude marks AI-generated content, and the EU is moving on labelling requirements for authentic-looking AI content. We covered what that means in practice in Claude watermarks AI-generated text.

But detection is the wrong thing to optimise against. Nobody in those viral threads was running a detector. They could tell because the writing had no specifics and the sender clearly hadn't read it. Editing well solves the actual problem; evading detection solves a problem you don't have.

Where this fits

This is the second of the two skills that matter most in being good at AI rather than just using it — verification is the first, and this is the second. Neither is about prompting.

If you want to know which of the eight underlying skills you're missing, and which ones your profession is specifically hiring for, the AI Skills Gap Analyzer runs the check in your browser and maps it onto your field.