SearchResearch (10/28/26): How do you know when the AI is wrong? What then?

You can't fact-check everything... 

... there's just too much.  It's the proverbial drinking from a fire hose... even if you're just a dog on the internet.  (Slightly obscure reference.)  

The past few posts have been on a related problem--how to use AI to help find relevant information and check on the accuracy of things.  

But if everything is suspect and you're drowning in content, how do you know when to spend the time and effort on checking?  

You might remember this example from a while ago. Here's the photo I took: 

When I asked Google Lens what this tree was couple of months ago (July, 2026), here's the answer I got... 

I know for certain that this a California Buckeye (Aesculus californica)--you can get all of the details from the authoritative source, CalFlora.  

As I said back then, Google search-by-image (aka Google Lens) gets it wrong. I can't really blame it because it looks a lot like a California Sycamore (Plantanus racemosa) in its fall foliage.  But it's only early July, so that can't be right.  

Even stranger, when I ask Google Lens today (Oct 18, 2026), the answer is different yet again... now it says it's a 

"... a small deciduous tree, which appears to resemble a distressed or stressed Japanese maple (Acer palmatum) or a similar small ornamental tree with thinning, browning foliage."  

That's also very wrong.  

Regular Reader Ramón pointed out that "...you know a lot about trees, so you know that it's wrong... but I don't live in California--how would I know that the answer is wrong?"  

Great question. 

To generalize a bit: When should you trust ANY AI-generated result? 

The teacher in me wants to say "never--just check it all!"  But that's not realistic.  Like the dog in the top image, there's just too much stuff flowing across your screen to check it all.  

Here are four heuristics I've found to help you decide when to turn off the fire hose and pull out the magnifying glass:

1. The "Load-Bearing Fact" Test

Ask yourself: If this fact is wrong, does my entire argument or project collapse? If an AI tells you the capital of North Dakota is Bismarck to settle a trivial debate, a hallucination just costs you pride. But if you are using a specific fact or observation as the foundational premise of a research paper, a syllabus module, or a legal argument, it requires deep verification. Fact-check the structural beams, not every speck on the wallpaper.

2. Missing the Meta-Context 

As demonstrated by the Buckeye masquerading as a fall-foliage of the Sycamore in the middle of summer, AI often ignores the meta-context. 

If an AI generates an answer that conflicts with the physical laws of time, space, or seasonality, dig deeper. Large language models lack common sense about the physical world; when an answer feels disconnected from basic reality—like a pristine winter coat recommendation for a July hike in the Swiss Alps—trust your intuition and verify. (Note that your intuition is based on what YOU bring to the table. The more you know, the better your intuitions will be.)  

3. "Too Neat and Tidy" Warning

Real-world data is messy, contradictory, and full of caveats. AI systems, however, hate that.  They're statistically tuned to be highly agreeable and narrative-driven. If an AI provides an answer that perfectly resolves a notoriously complex debate, or gives a suspiciously clean historical timeline without any friction, it is time to stress-test the source. Reality rarely fits into a perfectly symmetrical box; if the answer reads like a Hollywood script resolving a plot hole, it’s highly suspect. Verify.  

4. The SIFT Trigger: High Emotions Cause Checking

Applying the principles of the SIFT verification method (hat tip to Mike Caulfield), pay attention to your own internal reaction. Does the AI's response immediately validate your deepest held beliefs, or make you unexpectedly outraged? 

Or... When you feel a strong emotional surge of "I knew it!" upon reading a result, that is precisely the moment you need to stop, step back, and triangulate the claim with an authoritative, independent source. When YOU feel completely validated is exactly the moment your guard needs to go up.  Verify that feeling.  

In both cases (very positive or very negative high emotions), you should breathe in and start a bit of checking.  

In short..  

You don't need to fact-check the weather forecast for your afternoon walk. But when the stakes are high, the context feels slightly off, the answer is too perfect, or your emotions spike, that’s your cue to stop reading and start researching.

And, as I've said before--the more you know, the better your intuition will be.  I have pretty good knowledge about plants and trees, so I can smell a suspicious result when it comes over the digital transom.  

Pro tip: To inoculate yourself against AI errors, spend the time to become fairly knowledgeable about the topic.  You need to know enough to know when things seem shady, weird, or just a bit too-perfect.  

Because... the more you know, the more you know.  

Keep searching. 

SearchResearch (10/12/26): How to Verify Assertions in an Age of Infinite Noise

We live in a golden age of information...

Diogènes, the seeker-of-truth. Painting by Jean-Léon Gérôme (1860). P/C Wikimedia.

... and an equally catastrophic era of unverified claims.

Every day, we are bombarded with sleek infographics, AI videos, confident tweets, corporate press releases, and persuasive AI-generated assertions.

They sound authoritative. They look polished. But how often do we stop and ask: Is this actually true, correct, and valid?

In the past couple of weeks, two Regular SearchResearchers (thanks, Ramón and jeffp!) have asked versions of "but how do I know what's true and verifiable?"

It's a great question. So let me try another version of an answer. (I will note that many others have written about this "how to determine.." question as well. See citations below for more perspectives on this.)

In our rush for answers, we often confuse fluency with factuality. A well-written sentence feels true because our brains love cognitive ease.

But the reality is that every day I read assertions that sound authoritative: a statistic in a news story, a quote on social media, a confident paragraph from an AI chatbot. Some are true. Some are almost true. Some are true but don’t support the conclusion they’re attached to. The trouble is that all three look identical on the screen.

So here’s this week’s Challenge:

When someone hands you a claim, how do you decide whether it’s true, correct, and valid?

I’ll give you my answer below, but try think about your own verification routine first. I suspect most people discover they don’t have one. Do you? 

Notice: True, Correct, and Valid are three different questions

We tend to lump them together as “is it right?”, but each one fails in its own way, and each needs a different check.

True, correct, and valid are three different questions

We tend to lump them together as “is it right?”, but each one fails in its own way, and each needs a different check.

	What you’re really asking	How it typically fails
True	Does the claim match the world? Did this thing actually happen, does this person exist, was this said?	Fabricated quotes, misattributed photos, events that never occurred
Correct	Are the details accurate and complete? Right number, right date, right units, right context?	A real statistic from the wrong year; a real quote with the crucial second half cut off; percent vs. percentage points
Valid	Does the conclusion actually follow from the evidence offered?	Correlation sold as causation; one study generalized to everyone; a true fact used to support an unrelated claim

A claim can pass one test and fail the others. “The popularity of the first name Killian is highly correlated with Automotive recalls on air bags” is true and correct, but “kids named Killian causes air bag recalls” is not a valid conclusion. A misremembered or misunderstood statistic might support a perfectly valid argument, but the argument is still built on a wrong number.

A great example of a "spurious correlation."
This website by Tyler Vigen has wonderful (and crazy) examples.

The practical upshot: before you verify anything, decide which of the three you’re checking. Most verification failures I see come from checking the easy one (does a source say this?) and assuming the other two come along for free.

A six-step verification process for claims and assertions

Here’s the routine I actually use. It’s not fast for every claim, but it scales in a reasonable way: a trivial claim takes 30 seconds, a contested one takes an afternoon.  Here’s a good process to follow: 

1. Pin down the claim. Try rewriting it as one precise, checkable sentence. “Coffee is bad for you” isn’t checkable without a major research effort. “Drinking more than 4 cups a day raises heart-attack risk in adults” is. Vague claims can’t be verified, only argued about. Also note who is asserting it and what they want you to do with it. Also notice if the claim is somehow connected to the person making the claim. Do they have an interest in having you believe it?
   
   

2. Stop and notice your reaction. If the claim makes you feel vindicated or outraged, slow down. Strong emotion is the single best signal that you’re about to skip a step or jump to an unwarranted assessment of accuracy.
   
   

3. Trace the evidence upstream. Follow the claim back to its origin: the original study, dataset, transcript, filing, or photo. Every hop between you and the source is a chance for a claim to drift or context to vanish. If you can’t find an origin, that’s a finding in itself.
   
   

4. Read laterally. Most people practice vertical reading—staying on a single page, scrolling up and down, reading an article from top to bottom to judge its credibility. Don’t judge a source just by staring at it in isolation. Open new tabs and see what other people say about the source and the claim. Who runs this site? Who funds this group? What do people with no stake in the outcome say about the claim?
   
   

5. Triangulate. Look for independent confirmation. In particular (as I’ve said many times) find sources that don’t just copy each other. Ten articles citing the same press release are one source, not ten. And look actively for disconfirming evidence.  This is the step everyone skips.
   
   

6. Check the reasoning, then calibrate. Now test validity: does the evidence support this conclusion, or something narrower? Then decide how confident you are, and say so in words: confirmed, likely, unsupported, false. “I couldn’t verify this” is an honest and useful answer.

Steps 1 through 5 mostly test truth and correctness.

Step 6 is where validity lives, and it’s the one that requires thinking rather than searching.

The special case: verifying what an AI tells you

Asking for AI answers break the usual cues. They’re fluent, well-formatted, and uniformly confident, so the usual signals we use to judge a web page tell you nothing.

The routine still works, with a few adjustments:

• Treat every citation as a claim, not as evidence. Open the link. Check that the page exists, says what the AI says it says, and is about the same thing. Mismatched or invented citations are still common.
  
  

• Watch for the plausible detail. AI errors cluster in specifics: dates, names, page numbers, exact figures, quotations. The gist is often right while the details are wrong, which is precisely a correctness failure.
  
  

• Don’t ask the AI to grade itself. “Are you sure?” mostly produces either a cheerful confirmation or a nervous reversal. Neither is verification. Check against an independent source instead.
  
  

• Ask for the upstream source. “Where does this number come from originally?” turns an AI answer into a starting point for step 3 rather than an endpoint.

Used this way, AI is a terrific lead generator and a poor final authority. That’s not a criticism; it’s just recognizing what your tools can do for you. 

 

Teaching this so it sticks

The research here is fairly clear: checklists don’t transfer to students, practiced moves do. Sam Wineburg and Sarah McGrew’s studies at Stanford found that professional fact-checkers outperformed both historians and Stanford undergraduates at judging websites, largely because they left the page quickly and read laterally instead of scrutinizing it. Mike Caulfield’s SIFT method (Stop, Investigate the source, Find better coverage, Trace claims to the original) packages those moves into something teachable.

Here’s what I’ve found works when I teach these skills in a classroom:

• Teach moves, not checklists. Long evaluation checklists (check the URL, the design, the About page...) take ages and reward the wrong things. A handful of quick moves, practiced until they’re habits, beat a 20-item rubric.

• Use real, messy claims. When you teach this skillset, use examples from this week’s feeds, including ones that turn out to be true. If every exercise is a hoax, students learn cynicism, not judgment.

• Separate the three questions explicitly. Give students a claim and have them label each part: what would make this true, correct, valid? Then have them find a claim that passes one test and fails another.

• Think aloud, then hand it over. Model a verification live, narrating your doubts and dead ends. Then let students do one while narrating theirs. The dead ends are the lesson.

• Time-box it. Give a 3-minute limit for a quick check. It forces prioritization and shows that most claims can be triaged fast. It also teaches you that not everything can be done rapidly.

• Grade the reasoning, not the verdict. A student who says “likely, but I couldn’t find the original” with a clear trail has done better work than one who guessed right.

• Close the loop with AI. Have students verify a chatbot’s answer, including its citations. It’s the most memorable lesson in calibrated trust I know. I've used this is my graduate AI classes... and have had students be surprised by the results.

Above all, make verification feel normal and quick, not like a special forensic event. The goal is a reflex: before I share or rely on this, let me take a quick look.

SearchResearch lessons

1. Decide which question you’re answering. True, correct, and valid fail in different ways. Name the one you’re checking.

2. Go upstream. The original source beats any number of summaries of it. If you can’t find the origin, say so.

3. Leave the page. Read laterally to learn what others say about a source before trusting what it says about itself.

4. Count independent sources, not total sources. Ten copies of one press release are one source.

5. Say how sure you are rather than just showing the answer. “Confirmed,” “likely,” and “couldn’t verify” are all legitimate, useful answers, and calibrate the degree of certainty.

Now it’s your turn. What’s your verification routine, and what’s the best lesson you’ve used to teach it? Leave it in the comments. I’ll collect the best ones for a follow-up post.

Keep searching!

--

Citations:

Wineburg, S., & McGrew, S. (2019). Lateral reading and the nature of expertise: Reading less and learning more when evaluating digital information. Teachers College Record, 121(11), 1–40.

Caulfield, M. (2019). Web Literacy for Student Fact-Checkers. Pressbooks.