Invent less, hedge more
The latest models are noticeably better at saying "I am not sure" and at declining to produce a citation they do not have. They still do it.
Does AI make things up?
Does AI make things up? Yes. It is called hallucination, and it looks exactly like knowledge.
Every large language model invents things: citations to papers that do not exist, statistics with no source, features a product never had, court cases nobody filed. It happens because the model produces what fits the pattern, and a plausible-sounding specific fits the pattern whether or not it is real. This page explains what hallucination is, where it happens most, why newer models reduce it without ending it, and how to work with AI so the invented parts never reach your work.
Free to start. Perplexity Sonar Pro and Auto Verification on paid plans are built for this.
What it is
Pattern completion without a truth check
A language model is trained to continue text plausibly. It has read a vast amount and compressed it into patterns; it has not stored a database of facts it can look up. When you ask for a citation, it produces something with the shape of a citation. When the real one is in its patterns, you get the real one. When it is not, you get a convincing fake, delivered in the same tone. The model has no internal signal that distinguishes the two cases, which is why it does not warn you.
"Hallucination" is the accepted term; "confabulation" is more accurate, because the model is not seeing things, it is filling gaps. Either way, the practical fact is the same: any specific detail a model produces from memory might be invented, and the fluency of the sentence tells you nothing.
Where it happens most
Six hot spots
| Hot spot | Typical invention | Defence |
|---|---|---|
| Citations and references | A real-sounding author, year and title that do not exist | Open every citation. Use a model that retrieves and links. |
| Statistics | "73% of small businesses…" with no origin | Ask for the source and its date; find the original study. |
| Niche technical details | A library method, a config flag, an API parameter that was never there | Check the docs; run the code. |
| Recent events | Confident description of something after its training cut-off | Use a search-native model; check the date. |
| People and companies | Biographies with invented roles, products, controversies | Never repeat without a source; this is where reputational damage happens. |
| Legal and medical specifics | Case names, dosages, statute numbers | Professional verification, always. |
Why it is not fixed
Newer models reduce it; nothing eliminates it
Search before answering
Models that retrieve pages and cite them, such as Perplexity, ground answers in something you can open. They can still misread a source; you check.
A second, independent check
Auto Verification in MultipleChat re-checks factual claims independently on paid plans. It catches what the first answer got wrong; it is not a guarantee.
The last line
Opening the source is the only step no model can do for you. Everything else makes it faster.
The fix
The workflow that keeps invented facts out of your work
The same method works for facts, code, advice and writing. It costs one or two messages in MultipleChat because every model reads the same conversation, files and Project.
Ask a second model the same question
Switch the model from the selector and ask again; the history carries over. Pick a model from a different family: ChatGPT then Claude, or Claude then Gemini. Same-family models share blind spots.
Ask it to attack, not to agree
"Argue against this answer as strongly as you can" produces more than "do you agree?". Models are agreeable by default; you have to invite disagreement.
Read the disagreement as a map
Where the two agree, confidence is warranted. Where they differ, that is the claim, the assumption or the line of code to check. Usually it is the one that mattered.
Get a source or a test for anything you will act on
Facts get a source you open (Perplexity cites; Auto Verification on paid plans runs an independent pass). Code gets a test you run. Advice gets a third model or a person.
Keep it in a Project
The question, both answers and the sources stay together, so the next time the topic comes up the checking is already done.
Free is the start. The latest models are the difference.
Checking answers is where the latest models earn their price.
The free plan gives you the major model families and enough messages to make second opinions a habit, which is the whole point of this page. The paid plans add the latest model in each family, which disagree in more useful ways and invent less; Perplexity Sonar Pro for sourced facts; and Auto Verification, an independent verification pass that runs when factual confidence matters, without you orchestrating it.
Newer models hold your whole Project in view instead of the last few messages, reason through trade-offs instead of picking the first plausible answer, and are wrong less often and with less confidence. For the questions on this page, that is the difference between advice that sounds right and advice you can act on.
Pro from $20 a month, cancel any time. The free plan stays free.
- Your first Project and file uploads
- The major AI families to try the workflow
- A small daily message allowance
- The latest ChatGPT, Claude, Gemini, Grok and Perplexity models, all on the same Project
- Far more messages a day, so a working session does not stop halfway
- Modes where several models draft, challenge and verify each other's answers
- Auto Verification: an independent check on factual claims
- Perplexity Sonar Pro for sourced, current facts
- Collaboration modes where models draft, challenge and verify each other
Start here
Messages that expose invention
Use on any answer with specifics in it.
SourcesGive a source I can open for each specific claim. If you do not have one, say "no source" for that claim.
ExistenceDoes [paper / feature / case] actually exist? Find the original and quote it.
CertaintyMark each specific in your answer as: verified, likely, or generated to fit.
RetrieveSearch for this and cite what you find. Do not answer from memory.
Cross-checkHere is another model's answer with specifics. Which of them can you confirm with a source?
StripRewrite this answer removing every claim you cannot source.
Honest answers
What people ask
Does ChatGPT make up sources?
Yes, and so do all the others. Fake citations are the most common form of hallucination. Open every reference before you use it.
Is hallucination getting better?
Yes, with each generation, and retrieval-based models reduce it further. It has not gone away and will not fully.
Which AI hallucinates least?
It changes with each release. Retrieval models with citations are structurally better for facts. The safe answer is to check any specific regardless of model.
Can AI detect its own hallucinations?
Somewhat, if asked directly for certainty marks, and better if a different model checks it. It cannot do it reliably unprompted.
How does MultipleChat help?
Perplexity retrieves and cites, other models cross-check on the same context, and Auto Verification on paid plans runs an independent pass on factual claims. The invented parts get caught before they reach your document.
What is free?
The major model families including Perplexity Sonar, and a daily message allowance. Sonar Pro, Auto Verification and the latest models are on paid plans.
Assume every specific might be invented until you have opened its source.
Retrieval, cross-checking and verification make that fast. Free to start.
Also: how to fact-check ChatGPT · can AI lie to me? · research with AI