Pre-launch · Training and the free Foundations exam are open now. Levels 1–5 open soon. Get notified on any level page.
20 min read
An AI tool can return a clean, confident answer in seconds. That speed makes it tempting to skim and send. This module shows you how to read an output the way an editor would: check the substance first, then the style. You'll need this any time an AI draft is about to reach a customer, a manager, or a decision.
A fluent answer is not a verified answer. The model generates text that sounds right. Sometimes it adds a date, a number, a rule, or a quotation that wasn't in your source and isn't true. This is often called a hallucination: plausible content without reliable support. It doesn't look different from the true parts. It uses the same confident tone and the same tidy formatting.
So read in this order.
1. Substance before style. Ignore the polish on the first pass. Ask: what does this output claim? Underline every name, date, number, price, deadline, policy, and quotation.
2. Trace each claim. For each underlined item, ask whether it came from something you gave the tool. If you pasted a policy and the output quotes a rule that isn't in that policy, that rule is unsupported. It might be true. It might not. Either way, the output didn't earn it.
3. Match checking to consequence. You don't need to verify every word of a casual draft. You do need to verify anything where a mistake costs money, safety, trust, or compliance. A wrong adjective in an internal note is cheap. A wrong refund window sent to 400 customers is not. Check consequential claims against the original source, not against another AI answer.
4. Review against your requirements. Go back to what you asked for. Did you require five items and get four? Did you ask for the deadline to be included and it quietly dropped out? Missing content is easy to miss because you're reading what's there, not what isn't. A checklist built from your prompt catches it.
A citation doesn't settle anything on its own. Tools can produce a believable-looking source title or link that doesn't exist, or a real source that doesn't say what's claimed. If a claim matters, open the source and find the sentence.
Luis Ortega manages a home-goods store. Head office sent a two-page memo about changes to the return policy, and Luis wants a short summary for his floor staff. He pastes the memo and asks for "a summary for staff with the key changes." The memo says:
The output comes back looking polished. It lists the 30-day window and the receipt rule. It also says, "Per state law, customers are entitled to a full refund on defective items for up to one year." And it leaves out the store-credit rule for opened appliances.
Weak approach. Luis reads it, likes the tone, and posts it in the break room. Two problems ship: a legal claim no one at the company made, and a missing rule that staff will now get wrong at the register.
Better approach. Luis reads for substance first. He underlines four claims. Three trace back to the memo. The "state law" line doesn't appear anywhere in his source, so he marks it unsupported. Then he checks completeness against the memo and sees that one of the three changes is gone. He sends a follow-up:
Before:
Make it better.
After:
Revise the staff summary using only the memo I pasted.
1. Remove any statement that is not in the memo, including the
line about state law.
2. Include all three changes: the 30-day window starting April 1,
store credit only for opened small appliances, and receipts
required for returns over $50.
3. Under the summary, list each claim with the memo sentence it
comes from.
Keep it under 100 words.
Why this is stronger. The follow-up names the specific gap instead of hoping a vague "better" fixes it. It limits the output to the source, which removes the invented legal claim. It requires every change, so nothing drops out. And it asks the tool to show where each claim came from, which makes Luis's final check fast. He still reads the memo once more before posting, because the summary will change what staff tell customers.
Setup. Take a short document you have on hand: a policy page, an event notice, a product description, or a set of meeting notes. Paste it into a chat tool and ask for a summary in five bullet points.
Task. On paper or in a text editor, list every factual claim in the summary. Next to each, write the sentence from your document that supports it, or write "not in source." Then list each important point in your document and mark whether the summary kept it.
What a good result looks like. You end with two short lists. The first shows every claim as supported or unsupported. The second shows any missing points. If both lists are clean, the summary is safe to use. If not, you have exactly what your follow-up prompt needs to say.
Rewording can change meaning. "Must" becoming "should," or "up to 30 days" becoming "30 days," is a factual change that looks like a style change.