Before Finalizing Your Writing, Have AI "Take Roll Call" Item by Item: The Verification Statement Method

Last week, while revising a product copy, I asked AI to check for "exaggerated claims." It replied: "The overall tone is steady, and no obvious issues were found." I reread it line by line and spotted the phrase "industry-first innovation" in the third line—it missed it, but I caught it. I asked it to revise, and it inadvertently introduced two new problems.

Where did things go wrong? With instructions like "check for problems," the AI’s job is to "state an opinion," not to "provide evidence." Stating an opinion has a low cost, while providing evidence has a high cost, so naturally, it chooses the former.

What Is This Skill?

**Verification Statement Method**: Instead of asking AI for a vague conclusion, require it to output three things for every verifiable claim, item by item: the original text, the source of evidence, and the confidence level (confirmed / questionable / unconfirmable). Change "Do you think this works?" to "Prove it to me, item by item."

When to Use It

- **External Copy / Marketing Pitch**: Contains data, rankings, or words like "unique" or "first-of-its-kind," where the cost of failure is high.

- **Reports for Bosses / Clients**: Heavy on concluding statements; one mistake can invalidate the entire document.

- **Compliance-Sensitive Content**: Financial or medical copy, where promising language must be reviewed item by item.

- **Factual Content Written by AI**: Having it verify its own output item by item is far more effective than asking it to "rewrite more accurately."

When Not to Use It

- **Brainstorming / Idea Generation Phase**: Item-by-item verification stifles divergence. Diverge first, verify later; don’t reverse the order.

- **Purely Creative / Stylistic Edits**: You can’t "take roll call" for "make this paragraph more vivid." Comparative review (giving it two options to choose from and explain) is more suitable.

- **Short Content**: For a post of three or four sentences, scanning it yourself is faster than running a verification cycle.

- **Scenarios Requiring Manual Line-by-Line Review Anyway**: This indicates that what’s missing from your process is human oversight, not a tool to fill the gap.

How to Implement: A Reusable Prompt

> Please perform a verification check on the following text. List all verifiable claims in the text item by item (data, facts, rankings, promises, timeframes). For each item, output: ① Original Text ② Supporting Evidence (where it is self-evident in the text, or mark as "needs external verification") ③ Risk Level (Confirmed / Questionable / High Risk). Output only the list and a one-sentence summary; do not modify the original text.

The keywords are: **item by item, list, original text, do not rewrite**. The last part is crucial—if verification and modification are mixed, the AI will edit as it checks, inadvertently "fixing" the issues it finds, causing you to lose the chain of evidence.

Checklist

- [ ] Claims are broken down into atomic sentences, with each containing only one verifiable point

- [ ] Numbers, percentages, and dates are listed as separate items

- [ ] Phrases like "industry-first," "unique," or "best" are marked as high risk

- [ ] Each item cites the source of evidence or is marked "needs external verification"

- [ ] AI is prohibited from modifying the original text while verifying

- [ ] Items marked "needs external verification" are assigned to humans for follow-up, not thrown back to AI

Common Pitfalls

1. **Vaguely defining verification standards yourself.** "Check for exaggeration" is too ambiguous; AI will pass it using the loosest standard. Clearly define your criteria first: what constitutes exaggeration, and what counts as documented evidence.

2. **Asking it to verify text that is too long in one go.** For texts over 800 words, split them into sections; the hit rate for identifying claims per section is significantly higher.

3. **Treating "questionable" as "passed."** Every item marked "needs external verification" or "questionable" in the list must have a follow-up: either you verify it, or you delete it. These are the most valuable parts.

4. **Stopping after one round of verification.** After AI revises, the claims change, so you must run another verification cycle. The more edits made, the more re-verification is needed.

5. **Using AI to verify numbers generated by AI.** It can "self-validate" numbers it fabricated out of thin air. For numerical and factual claims, evidence must come from a fact sheet you provide or external sources.

Remember in one sentence: **Demand evidence first, then accept the conclusion.** The process of having AI "take roll call" is the beginning of your editorial chain of evidence.