AI for Everyday Work: Writing, Research, Meetings & Data · Research and verification · lesson 13 of 17 · 11 min

Verification and fact-checking in five minutes

Links aren't proof

A citation shows where the AI says information came from. It doesn't guarantee the source says exactly that, that it's current, or that it's trustworthy. AI can also occasionally produce references that don't exist, especially when asked for academic papers without real search. Verification is the step that turns AI research into something you can safely use.

The 5-minute verification routine

For the facts that matter most (usually 3 to 5 per piece of work):

Open the source. Find the sentence or number. Does it really say that? (Use your browser's find function.)
Check the date. When was it published or updated? Is it still current?
Check who's behind it. Official body, respected publication, or someone selling something?
Cross-check. Can you find the same fact in a second, independent reliable source?
Record it. Note the source link and date next to the fact in your document.

Verification prompts

Ask the AI to help you check its own work:

For each factual claim in your answer, list: the claim, the source link, and the exact sentence from the source that supports it. Mark any claim you can't directly support.

What in your answer might be out of date or uncertain? What should I double-check?

These prompts don't replace opening sources, but they point you to the weak spots quickly.

Spotting common red flags

Very precise numbers with no source ("73.4% of shoppers...")
Sources that don't mention the claim when you open them
Old dates for fast-changing topics (prices, rules, technology)
Only one type of source, such as only company blogs
Academic references you can't find in a library or scholarly search
Quotes attributed to famous people that you can't find in a reliable source

Worked example

Leila, a university student in Manchester, asks an AI for five sources about social media and mental health for an essay. Following the routine, she searches each in her library: three exist and are relevant, one exists but was published in a different year, and one can't be found at all. She uses the three verified sources, corrects the year on the fourth after reading it, and drops the fifth. Her tutor's feedback praises her referencing. Her university's AI policy allowed AI for finding sources, but not for writing her essay, so she wrote it herself.

Sample verification note

Keep notes short and consistent, for example: "Claim: new rule takes effect in April. Source: official government page, updated [date]. Exact wording found. Verified." or "Claim: 40% of customers prefer X. Source: vendor blog, no methodology. Not verified; removed." Notes like these make your work easy to defend if someone asks.

Scale checking to the stakes

Low stakes (internal brainstorm): a quick sense check.
Medium stakes (client email, blog post): verify key facts and names.
High stakes (anything legal, financial, medical, published widely, or about real people): verify every fact with primary sources and get expert review where needed.

Hands-on: the claim-check table

Paste this after any AI research answer, then do the opening yourself.

Make a claim-check table for your answer:
Claim | Source link | Exact supporting sentence from the source | Date of source | Confidence (high/medium/low)
Only include claims that matter for my decision. Mark any claim with no direct supporting sentence as UNSUPPORTED.

Then, for each high-stakes row, open the link and use your browser's find function (Ctrl+F or Cmd+F) to search for a distinctive word from the "exact sentence".

Before: a polished paragraph saying a regulation "requires all businesses to register by June" with one link.

After: a table showing the regulation applies only above a revenue threshold, the linked page dated two years ago, and a newer official page (found on cross-check) with an updated date. You caught a costly mistake in four minutes.

Checking AI-generated images and media

Realistic AI images, audio and video are easy to make. Before sharing surprising media:

Look for the original source (reverse image search, the publisher's own channels).
Check for provenance labels or content credentials where platforms show them.
Be cautious with emotionally charged content that's spreading fast.

Keep a verification note

Next to each key fact in your document, record the source and the date you checked it. It takes seconds and makes your work defensible when someone asks, "where did this come from?"

Pitfalls

Verifying only the facts that "feel" wrong; confident errors feel right.
Citing the AI tool as your source rather than the original.
Skipping verification under time pressure; that's when mistakes slip through.

Video lecture: Verification and fact-checking in five minutes

Lecture coming soon · 10 chapters · about 8 minutes. Read the full transcript below.

Verification in five minutes
Why links aren't proof
The five-minute routine
The claim-check table
Example 1: the '73.4 percent' statistic
Example 2: the registration deadline
Watch me do it
Red flags
Scale it to the stakes
Recap and try this now

Lecture transcript

Verification in five minutes

Here's a story that happens more often than you'd think. Someone asks an AI for statistics for a presentation. The answer is beautifully written, with a neat little citation after each number. The deck goes to a client. And the client's analyst clicks one of those links and finds that the page says something completely different. Nothing was invented on purpose. But nobody checked. In this lecture you'll learn a five minute verification routine that turns AI research into something you can safely use. You'll see a simple check and a costly near miss, and watch me verify a real answer, claim by claim.

Why links aren't proof

So why isn't a link proof? Four reasons. First, the source may not actually say what the AI claims. It might say something similar, or the opposite, or talk about a different country. Second, it may be out of date. A rule from three years ago might have changed. Third, it may not be trustworthy, like a company blog selling the very thing it's praising. And fourth, occasionally, especially when you ask for academic references without real search, the reference might not exist at all. A citation tells you where the AI says it got the information. Verification is how you find out if that's true.

The five-minute routine

Here's the routine. Use it for the facts that matter most, usually three to five per piece of work. Step one, open the source and find the sentence or number. Use your browser's find function. Does it really say that? Step two, check the date. When was it published or updated, and is it still current? Step three, check who's behind it. An official body, a respected publication, or someone selling something? Step four, cross check. Can you find the same fact in a second, independent, reliable source? And step five, record it. Put the link and the date next to the fact in your document. About a minute per fact. Five minutes for the facts that could hurt you.

The claim-check table

You can get the AI to help you check its own work. After a research answer, paste this: make a claim check table. Claim, source link, the exact supporting sentence from the source, the date of the source, and a confidence rating. Only include claims that matter for my decision, and mark any claim with no direct supporting sentence as unsupported. This doesn't replace opening the sources. But it does two useful things. It points you straight to the weak spots, and it gives you a distinctive phrase to search for when you open each page. If the exact sentence isn't on the page, you've found a problem in seconds.

Example 1: the '73.4 percent' statistic

A simple example. You're drafting a newsletter and the AI writes: seventy three point four percent of shoppers prefer brands that use sustainable packaging. It's specific, and specific numbers feel trustworthy. That's exactly the trap. There's no link. So you ask: which source says this? The answer is vague. It mentions a range of surveys. That's your red flag. A precise number with no clear source is treated as unconfirmed. You either find a real survey from a reputable organization and quote its actual figure, with the date and who ran it, or you rewrite the sentence without a number. Many shoppers say sustainability matters to them is less exciting, and much safer.

Example 2: the registration deadline

Now a costlier near miss, with illustrative details. An operations manager at a small import business in Dubai asks an assistant about a new registration requirement. The answer says all businesses must register by June and links to a government page. She runs the claim check table. The exact supporting sentence doesn't match: the page says the rule applies above a revenue threshold. And the page is dated two years ago. On cross check, she finds a newer official page with an updated threshold and a different date. Her business falls below it. Without the check, she'd have paid for a registration process she didn't need, or worse, told her board the wrong deadline. Four minutes of verification. Real money saved.

Watch me do it

Let me verify a real answer. I asked about the number of public holidays in a country for a planning document. I paste the claim check prompt. Three claims come back with links and exact sentences. Claim one: I open the link, which is an official government page, press control F, type a distinctive phrase from the sentence, and there it is, word for word, dated this year. Green. Claim two: the link goes to a news site. The sentence is there, but the article is from last year, and holidays can change. So I cross check with the official page. It matches. Green. Claim three: the phrase isn't on the page at all. That's unsupported. I remove it from my document. Then I paste each source and date next to the facts I kept. Done in about four minutes.

Red flags

Learn to spot the red flags. Very precise numbers with no source. Sources that don't mention the claim when you open them. Old dates for fast changing topics like prices, rules and technology. Only one type of source, such as only company blogs. Academic references you can't find in a library or scholarly search. And quotes attributed to famous people with no original source. Images and video deserve their own care, because realistic AI media is easy to make. Before sharing surprising images or clips, look for the original source, check for provenance labels or content credentials where platforms show them, and be extra cautious with emotional content that's spreading fast.

Scale it to the stakes

You don't need to verify every sentence of every answer. Scale it to the stakes. For an internal brainstorm, a light check is fine. For anything going to a client, published publicly, or used for a decision, run the full five minute routine on the key facts. And for legal, health, tax or financial matters, use AI research to prepare your questions, verify against official sources, and then get a qualified professional to confirm. The goal isn't to distrust everything. It's to put your attention exactly where a mistake would hurt.

Recap and try this now

Let's recap. A citation is where the AI says a fact came from. It's not proof. For the facts that matter, run the five minute routine: open the source, check the date, check who's behind it, cross check with an independent source, and record the link and date next to the fact. Use the claim check table to point you at weak spots, and treat the red flags seriously. Here's your try this now. Take the last piece of AI assisted research you used, even an old one. Pick three key facts and run the routine. If all three hold up, great. If not, you've just learned exactly why this habit matters.

Key takeaways

A citation shows where the AI says a fact came from; open it and find the exact sentence.
Use the 5-minute routine for the facts that matter: open, date, who's behind it, cross-check, record.
Ask for a claim-check table with exact supporting sentences and mark unsupported claims.
Treat precise numbers without sources, old dates and unfindable references as red flags.

Try it

Take an AI-assisted piece of work and run the 5-minute routine on its three most important facts. Record the results as verification notes.