You will finish this page able to sort a confident ChatGPT answer into claims you can prove and claims you cannot, in about 30 seconds. The case that made the point is a strange one.
A researcher tracing a two-plate Audran engraving was handed an image ChatGPT called an authentic scan. The inscriptions were invented. The sheet divisions were artificial. Only after being pressed did the model say it was an AI reconstruction.
The tone never changed. The same flat, declarative sentences carried the correct answer and the fabrication, so how sure the reply sounds tells you nothing about whether it is true. A 60-scenario rheumatology study at ACR Convergence found hallucinations on every platform tested, ChatGPT included.
Your version is smaller and faster. Thirty seconds of triage turns a fluent answer into a ranked list of exact claims. Each one either gets a receipt you opened yourself, or it gets cut. You leave with four instant triggers, a 10-second link-and-quote check, the follow-up prompt that forces a readable receipt, and a rule for when to leave the chat and go to the primary source.
| Claim type | Fastest check | What it catches | When it still fails |
|---|---|---|---|
| Link or citation | Open it and read the domain, author and date | Dead links, unrelated pages, invented sources | Real source that never contains the quoted line |
| Exact number | Find the figure on the primary source page | Wrong year, wrong unit, wrong population | Figure is right but answers a different question |
| Direct quote | Search a distinctive phrase inside the named source | Quotes that were never said | Quote is real and truncated mid-sentence |
| Named study | Search the study title plus the year | Studies that do not exist | Real study summarized beyond its scope |
| Fabricated image | Ask whether it is a scan or a reconstruction | AI images presented as authentic | Caption blends a real object with a fake image |
1. Why a Confident Answer Can Still Be Wrong

Prompt to copy
Read my last answer for exact claims only. Circle every number, name, date, direct quote, source and link, and put all of them in a list under the heading Unchecked." A false or invented detail, delivered in the same flat, declarative tone as a correct one, is a hallucination. That's the word for it: a wrong answer with no error message, no hedge, no visible seam.
It comes in four forms. A false fact. An invented source. A made-up quote or link.
And a fabricated image presented as real, which is exactly what happened when ChatGPT handed a researcher an AI reconstruction dressed up as an authentic Audran engraving.
Every major assistant does this.
💡 Tip: Circle the exact claims before you read for meaning. Once you have understood the argument, your brain stops looking for the holes.
How Often the Big Assistants Invent Things
One 60-scenario rheumatology study, presented at ACR Convergence
Figures from the ACR Convergence scenario study reported by Healio, November 2026.
2. Exact Numbers, Quotes, and Links Deserve Suspicion

Mark the triggers: a named study, an exact percentage, a direct quote, a publication date, a source URL, or anything about a recent change. A specific detail is the cheapest thing for a model to generate, so it is also the only part worth testing.
Rank them: note what happens if each one is wrong, and put the costliest at the top.
Check in the tool: paste one suspect claim into ChatGPT on its own and ask where it came from.
A vague answer, or a reply that restates the claim, is a flag.
Specifics feel credible because they cost the model almost nothing. In the Audran case ChatGPT made an image, presented it unlabeled, and said afterward it had favored something convincing over a truthful result.
A misread question needs a sharper prompt. A fabricated detail needs a source.
💡 Tip: Anything about a price, a rule or a date that changed in the last year goes to the top of the list by default.
3. Open the Link Before You Trust the Citation

The commonest failure is a citation that looks perfect and holds nothing.
A citation counts as evidence only when three things hold at once: the link opens, the page contains the exact sentence or figure, and the publisher is the sort of body that would say it.
Read four things: the domain, the author, the publication date, and whether that quoted sentence or number is really on the page.
Dead links, links to a homepage, real articles that never contain the quote, and wholly invented citations all turn up. Often within one answer.
Search the exact phrase: copy a distinctive run of words from the claim and put it in the site's own search or a search engine. If it appears nowhere inside the named source, the citation is dead.
💡 Tip: A citation to a homepage rather than a specific article fails by default. Real sourcing points at a page, not a website.
⚠️ COMMON MISTAKE
Asking the Model to Grade Itself
Pressing 'are you sure?' and accepting the second answer is not a check. Models can invent a source and then withdraw it under pressure without either reply touching the real world. Open the receipt yourself, or the check did not happen.
4. Put the Number Back in Its Own Source

A figure can sit on a real page whose wording matches and still be answering some other question. Read the unit first, then the currency, the year and who was counted.
Ten minutes with the source usually settles it by hand; a chat reply pointing at the figure takes seconds and buys you nothing new.
A rheumatology paper presented at ACR Convergence puts ChatGPT's hallucination rate inside its test scenarios at 6.2%. That is not a general failure rate.
Say "wrong 6.2% of the time" and a true number now answers a question the study never asked, which is the hardest kind of error to spot later.
Check the arithmetic before you repeat anything. The parts should sum to the total. Every percentage needs the base it was taken from. Any multiple should divide out exactly.
Then write the claim with its source and date attached. If that sentence will not form, mark the claim unverified and leave it out of the draft.
💡 Tip: Write the year into the sentence. 'In 2026, X found…' fails loudly the moment the source is from 2019.
5. Ask for the Receipt, Then Read It

Paste the answer you are checking with a request that forces a readable shape. What comes back is a short list of claims with a link, a supporting sentence and a date attached to each.
Prompt to copy
For each factual claim in your last answer, give the exact source URL, the sentence that supports it, and the publication date. Mark each claim Verified, Unverified, or Invented. Do not fill gaps with guesses.
You get the list in well under a minute. Expect a row of Verified labels and at least one claim marked Unverified, which is the model admitting a gap instead of filling it.
Read every row as a task for you, not a verdict. The model is grading its own work.
If a row still reads vague, press it a second time: ask which of those sources it has actually opened and which it is rebuilding from memory. In the Audran engraving case, ChatGPT called an AI reconstruction an authentic scan and conceded only under pressure.
So open each link yourself and match the quoted sentence on the page. On a high-stakes claim, that press is what separates a real citation from a confident guess. What you end with is a to-check list you can clear in minutes.
💡 Tip: Run the prompt before you edit the answer. Asking for sources after you have rewritten it invites citations to your own words.
One Reply, Two Outcomes
Checked before sending
- Duplicate claims deleted before the draft
- Every link opened and matched to its sentence
- Numbered figures carry a year and a source
- Three minutes longer, one version of the truth
Trusted on tone alone
- A dead citation sitting in a client deck
- A figure restated for the wrong population
- A quote the named author never said
- Clean-up after the fact, in front of a reader
6. What Never Goes Into a Check Chat

The receipt step is where people paste the document underneath the claim, and that single move can cost more than the wrong number ever would. One setup fixes it: a placeholder sheet you build once, in about ten minutes.
Sit for ten minutes and build a blank sheet of placeholders. Swap names, account numbers, addresses and dates of birth for tags like PERSON A or ACCOUNT 1. Then ask about the claim's shape, such as whether the numbers add up, never the raw content.
Keep passwords, client or patient files, and financial records out of the chat entirely. Anything under a confidentiality agreement stays out too.
On free and personal plans, the conversation may be used to train the company's models unless that is switched off in settings.
A claim resting on private data gets checked in the system that owns it, or an approved enterprise tool.
💡 Tip: If the claim cannot be checked without pasting the file, check it with the person who owns the file. That call takes less time than a leak.
7. When Ten Seconds Is Not Enough

Ninety seconds is roughly the point where a real claim stops fitting the quick check.
The 10-second link-and-quote test holds until a claim touches money, health, a contract or a public send.
Past that line, a clean-looking receipt is not enough and the source has to be read in full.
When that line is crossed, TrustScale's Argus is the named tool to try. It marks claims against the source evidence itself, instead of asking a second AI to grade the first.
It runs next to ChatGPT, Claude, Gemini, Copilot and Grok, supports 12 languages for enterprise teams, and reaches individuals through a research preview and a Chrome extension, as of November 2026.
The numbers on the box: up to 135 times faster than manual research and up to 98.5% accuracy improvement, according to TrustScale. Put both on your own triage list.
A vendor figure is an unchecked claim until you open the page behind it and match the sentence yourself.
Dataiku's review found no single detector covers all four detection methods, so teams stack them. The small-scale version: your 10-second check plus one tool or one second source.
Cheap overlap checks miss reworded fabrications, and similarity checks miss a date off by one year. Then: run the link check, and if it fails, go to the primary source.
💡 Tip: Set the rule in advance. If a claim could cost money, health or a client, it gets the slow lane automatically.
Frequently Asked Questions
Do I need to pay for ChatGPT to spot hallucinations?
No. Every check on this page uses a browser and a few clicks on the source itself. Paid plans change usage limits, not truthfulness, so a paid answer can still invent a citation. Use the paid tier if you need the features, not for accuracy.
Which assistant hallucinates least?
There is no permanent ranking. In the 60-scenario ACR Convergence study, OpenEvidence came in at 3.8%, ChatGPT 6.2%, Claude 8.5%, Perplexity 15.6% and Gemini 18.1%. Those numbers belong to that medical scenario set. Treat all five as capable of inventing a claim and check the specific answer you have.
How do I check a claim about a rule that changed recently?
Go to the official body that makes the rule and read its current page, not a summary of it. Do not rely on year mentions inside an old PDF or an out-of-date help article. Checking the current page is the official figure is the only version you should repeat.
Can I paste a whole document in and ask ChatGPT to verify it?
Not with private material. On free and personal plans, chats may be used to improve the company's models unless you switch that off in settings. Redact names and numbers first, or verify inside the system that owns the document.
Does all this checking take longer than writing the answer myself?
For a 700-word answer, the review runs about 8 to 12 minutes with the link and quote checks, against the hour or more of drafting the first version by hand. The page you get is checked, which is the part the draft never was.
What if the source link opens but the quote is not there?
Treat the claim as invented until proven otherwise. A live, reputable page does not vouch for a sentence it never contains. Search a distinctive phrase from the quote inside the source, and if it is absent, delete the sentence from the draft.
Run the Link-and-Quote Check Before the Next Send
Your first move on any claim headed for a send, a publish or a decision is the ten-second check: open the link, match the sentence, note the date. If it fails there, stop and go to the primary source.
For low-stakes curiosity, a quick source match is enough. For work, money, health or legal claims, go to the primary source or an approved verification tool. Confidence in the tone is never evidence, no matter how clean the paragraph reads.