AI FactScan
Before you trust the answer, open one source
Not every source. Just one. Does the page exist? Is it current? Does it actually say what the AI claims? That single click catches a surprising amount of nonsense.
The customer is always right. The chatbot is apparently righter
In a recent report for The Verge, Gaby Del Valle spoke with workers who are already dealing with a peculiar new customer dispute: the AI said so.
A restaurant guest with a shellfish allergy challenged a server because ChatGPT said the dish was safe. Diners asked for wines the restaurant did not carry, including bottles that may not exist. Visitors arrived at federal land with AI-generated itineraries that required them to row upstream. Parents trusted chatbot advice about a child's sleep over the teacher watching that child struggle through the day.
Customer service has quietly expanded. Staff must now know the menu, the inventory, local geography, child development, and whatever a language model imagined five minutes ago.
The old complaint was, “Your website says you have it.” At least that gave the employee a page to check.
The new complaint is, “ChatGPT says you have it.” Good luck asking ChatGPT which page it meant.
Reddit has been documenting the field tests
The Verge story is funny until the same pattern appears everywhere.
In one Reddit thread from a hardware-store employee, customers arrived asking for products the store did not sell, or products that did not exist. Some rejected equivalent alternatives because an AI had named one specific item. One customer apparently consulted a chatbot throughout the shopping trip, as if accompanied by a very confident friend who had never entered a hardware store.
Another retail worker described a customer insisting that a shop stocked a rare whiskey because ChatGPT said so. The store did not have it. ChatGPT, notably, did not have to check the stockroom.
The problem reaches beyond retail. In a small-business discussion, a prospect brought an AI-generated product design to a manufacturer. The files were not sufficient for production, but the chatbot had assured the prospect that they were. The business then had to defend reality against a machine that could not manufacture anything, refund anything, or feel embarrassed.
These are anecdotes, not a controlled study. They still reveal a useful pattern: an AI hallucination can escape the chat window and become somebody else's operational problem.
Why “ChatGPT said so” feels like evidence
Consumers are not foolish for using AI. They are using it because it removes work. Instead of comparing ten links, they get one calm answer that sounds as if the comparison already happened.
The trouble starts when the answer hides how uncertain that work was. A guess can look like inventory data. A scraped menu can look current. A blog post can sound like the business's official policy.
Then the customer reaches the store, restaurant, park, or school. The employee has direct access to the facts, but now sounds like the person disagreeing with the computer. The AI made the claim. A human gets cross-examined.
One extra click before you believe it
You do not need to fact-check every sentence. Start with the claim that will change what you buy, eat, book, or do.
Open one cited link and ask three quick questions:
- Is this the official business, agency, paper, or other primary source?
- Is the information current?
- Does the page support the specific claim, rather than merely discuss the same topic?
If there is no source, treat the answer as a suggestion. If the source is an old review while the claim concerns today's price or stock, call the business. That is not anti-AI. It is one extra click before handing the answer your credit card, dinner order, or travel plans.
When an answer includes links, AI FactScan can help you see what kinds of sources are behind it. It cannot reveal sources the model never showed. No source still means no source.
For important questions, ask twice
A second model will not magically produce the truth. It can reveal where the first answer was standing alone.
Ask the same question in two AI services and compare the links they use. Do they cite the same official page? Do they rely on different blogs? Does one provide sources while the other simply sounds certain?
If two models disagree, check the conflicting claim and open the sources each one cites. One answer may be outdated, unsupported, or simply wrong.