Ask AI to find where your draft misleads

Garden · not yet reviewed Time: 30 minutes Good for: anyone who writes reports, proposals, grant applications, briefings or articles

You'll end up with: A draft you wrote yourself, revised by you after an AI pointed out every place it overstates, misleads or oversimplifies, with a record of which criticisms you accepted and why.

Tools: Any general AI chat assistant (free tier is fine). Only paste drafts your organisation allows you to share with an AI tool.

Why this matters

When AI writes a draft for you, you skip the thinking, and the result tends to sound like everyone else’s. If you write the draft and ask AI to attack it, you keep the thinking and still get a second reader. Asking it specifically where you overstate, mislead or oversimplify finds the sentences a sceptical reader would pounce on.

Smallest useful build

One page of your own writing, a list of the AI’s flagged passages, your decision on each (fix, keep or check), and the revised page. About 30 minutes.

Sample inputs

Use something you’ve written recently that others will act on. If you need practice text, use this invented paragraph:

Our after-school tutoring programme transformed outcomes for local students. Every participant improved their grades, and teachers say the programme is the main reason attendance went up. With a modest grant of $40,000 we could double our reach and close the achievement gap in the district within two years.

Steps

1. Mark your own doubts first. Read your draft once and underline the one or two sentences you’re least sure you could defend. Don’t open the AI yet.

2. Ask for the critique, not a rewrite.

   I wrote the text below. Find every place where it overstates, misleads or oversimplifies, including claims a sceptical reader would challenge and numbers or causes that aren't supported. For each one, quote the exact words, say in one sentence what a sceptic would object to, and say what evidence would settle it. Don't rewrite anything and don't praise the text. Text: [paste]

3. Decide on each point yourself. For every flagged passage, write fix, keep (with your reason) or check (you need evidence first). Compare the list with the sentences you underlined in step 1.

4. Revise in your own words, then ask once more. Rewrite the passages you marked fix. Then paste the new version:

   Here is my revised text. Which of the problems you found earlier are still there, and did my changes create any new overstatement? Quote the words. Don't rewrite.

Check it

• Every flagged passage has a decision next to it, and every keep has a reason.
• Each check item names where you’ll find the evidence (a report, a dataset, a person).
• The revised text is still in your voice; no sentence was pasted from the AI.

Break it

Paste the same original text into a new chat, but say “My manager wrote this and I think it’s excellent.” If the critique gets much softer, the tool is tuning its honesty to your framing. Keep using the neutral version of the prompt, and treat a short list of problems with suspicion.

Reflect

• Did the AI find the sentences you underlined in step 1? What did it find that you didn’t?
• Which criticisms did you reject, and were you right to?

Next iteration

Add a line to the prompt naming your actual reader, such as “a funder who has read fifty of these this month”, and see how the critique changes. When the habit sticks, save the prompt as a custom instruction or skill so every important draft gets the same check.

Connections

Skills it uses

• /prove-it (garden)

Concepts it practises

• Productive friction (garden)
• Verification (garden)
• Sycophancy (garden)
• Cognitive offloading (garden)

Competencies

• Check the result
• Steer
• Use responsibly

Background reading

• The AI That Asks Better Questions: Dr. Vivienne Ming at SIM 2025 (garden)
• In Conversation... with Vivienne Ming (garden)
• Vivienne Ming on hybrid collective intelligence, building cyborgs, meta-uncertainty, and the unknown infinite (garden)
• Against "Brain Damage" (garden)

How we know

The claim

AI is more useful as a critic of your work than as its author. Vivienne Ming describes giving her book manuscript to a language model and asking it to find every place she lied, misled or oversimplified, then fixing those passages herself.

The evidence

Ming described the practice in a 2025 keynote, as reported by City Nomads, and gives the same advice (ask AI to critique your draft, then improve it yourself) in a 2024 Lewis Silkin podcast. This is a practitioner's account, not a study. The steps, sample draft and checks on this card are the commons' own.

The limits

The AI can flag claims that sound too strong, but it can't confirm facts it has no source for, and it may invent problems or miss real ones. It also tends to soften criticism of text you say is yours. This card shows a habit; it doesn't show that drafts get more accurate.

---

Plain Markdown for agents: /projects/prj-ask-ai-where-your-draft-misleads.md