What Is an AI Hallucination? (2026): A Confident Answer With No Source

An AI hallucination is a fluent claim that the source does not support. The model states a price, a quote, a citation, or a policy as if it were in the document. Sometimes the document never said it. Sometimes the model named a paper that does not exist. The sentence can be grammatical and still be wrong. Fluency is not evidence.

You reduce the damage by attaching a source and checking the sentence against it. You do not remove the failure by asking the model to “never hallucinate.” That instruction is a wish. The check is yours.

This page does not publish a hallucination rate. Vendors’ percentages are their studies, on their prompts. Prices are US list prices excluding tax from directory pages checked October 16, 2026, except Perplexity’s plan detail, which a later article read on October 24, 2026.

Where the invented claim shows up

What you asked	What a hallucination looks like	What to do instead
“What is our refund window?” with no policy attached	A plausible 30-day policy	Paste the policy, or refuse to publish a number that is not in it
“Cite a study”	An author, year, and title that do not resolve	Search papers in Consensus or Elicit and open the paper
“Summarize this PDF”	A finding from a different section, or from a similar paper	Require a quote, then find the quote in the file
A market figure	A dollar total with no page	Perplexity, then open the link. The steps are how to use Perplexity for market research

flowchart TD
  claim["A specific claim in the draft"] --> src{"Is there a passage you can open?"}
  src -->|No| cut["Cut the claim or find a source"]
  src -->|Yes| match{"Does the passage say that?"}
  match -->|Yes| keep["Keep it, with the link"]
  match -->|No| cut

The pattern of “retrieve, then write” is what RAG is. RAG gives you a passage. It does not guarantee the summary matches the passage. You still do the match step.

Tools that make the check cheaper

Need	Start here	Price to know
A public-web claim with a link	Perplexity	Free includes limited Pro Searches. The October 24, 2026 help-center read lists 3 Pro Searches a day. Pro is $17/month billed annually
A peer-reviewed paper	Consensus	Pro is $12/month on a $144 year, or $20 monthly
A table of fields pulled from papers	Elicit	Plus is $11/user/month billed annually on the non-Industry card
Files you uploaded	Gemini Notebook	Free with a Google account. Google AI Plus is $4.99/month
A draft from notes you paste, which you will line-edit	Claude or ChatGPT	Claude Pro $20/month or $17 annual. ChatGPT Plus $20

The bill for Perplexity’s ladder is Perplexity pricing explained. The paper workflow is how to research with Elicit and Consensus. A general chat is the wrong tool for “cite a real paper” unless you paste the paper.

Habits that catch the failure before a client does

Ask for the quote, then search the file for it. If the quote is not in the file, the summary invented a finding. Delete the finding.

Refuse numbers that have no page. Revenue, headcount, and “studies show” are the usual inventions. A model will supply a round number because round numbers are common in training text.

Split research from drafting. Collect claims with links first. Then ask Claude to draft only from that list. The writing workflow is how to write a blog post with Claude. A single prompt that says “research and write” hides the moment where a source should have been opened.

Treat a missing source as an answer. “Not in the notes” is a successful output. A completed paragraph with a made-up policy is the expensive one, because someone might follow it.

Temperature changes how varied the wording is. It does not add a citation. That setting is what temperature means. The underlying system is what an LLM is.

FAQ

Are hallucinations the same as bias?

No. Bias is a skew in what the model tends to say. A hallucination is a specific claim that is not supported. A sentence can be biased and sourced, or invented and neutral in tone. Check the claim either way.

Does a bigger model stop this?

A different model can be wrong less often on a test the vendor chose. Your contract, your price, and yesterday’s news are not that test. Attach the source. Open it.

Is a citation proof?

A citation is a pointer. If the page 404s, or the page says something else, the citation failed. Perplexity and Consensus are useful because they give you the pointer. They do not replace the click.

Should I tell the model it will be penalized for mistakes?

You can require quotes and abstention. A threat in the prompt is not a control. The control is the source check in the diagram above.

Last verified

October 16, 2026 for most plan prices. October 24, 2026 for Perplexity’s free-tier Pro Search note, via Perplexity pricing explained. Published October 24, 2026.

Sources: Perplexity pricing, Consensus pricing, Elicit pricing.

Published by Andrew for AI Tools Optimizer.

Perplexity, Consensus, and Elicit are in the research directory. Directory pages may use affiliate links. A purchase through them may earn us a commission at no extra cost to you.