Does Perplexity hallucinate? What the data shows
Last updated: July 26, 2026
Yes. Perplexity hallucinates in a distinctive way: it cites real, working sources but can attach a claim the source never actually made. In the March 2025 Tow Center study of eight AI search engines, Perplexity was the most accurate yet still gave incorrect answers on 37% of tested queries.
Does Perplexity hallucinate, and how often?
Yes, Perplexity hallucinates — and because it is built around citations, its errors are especially deceptive: they arrive wrapped in a real footnote.
In the Tow Center / Columbia Journalism Review study (March 2025, 1,600 queries across eight AI search engines), the engines collectively answered incorrectly on more than 60% of queries. Perplexity was the best performer, yet still wrong on about 37% of tested queries (CJR, 2025). "Best of the group" here still means more than one in three answers had a citation problem.
Perplexity's underlying language model matters too — it runs on models like GPT and Claude that carry their own hallucination tendencies. But its signature failure is not in the base model; it is in how it stitches sources together.
Best-in-class is not accurate
Perplexity leading a benchmark is a real credit — its retrieval and citation UX are genuinely strong. But leading a group where every engine failed most queries is a low bar. A 37% error rate is not a rate you can build a legal brief or a diagnosis on.
Perplexity's signature failure: the synthesis hallucination
Perplexity's defining failure is unique among these models: it cites a real source, and the claim it attaches to that source is wrong. The URL exists, the page loads, the author is credible — but the sentence Perplexity wrote was never in the page. The citation is real; the support is not.
This is why we say a real citation is not a supported claim. Independent reviews, including GPTZero's June 2024 investigation into what it called "second-hand hallucination," describe both this synthesis error and a related one where Perplexity retrieves a page that itself contains AI-generated errors, then restates them as cited fact (GPTZero, 2024).
Worked example. Ask Perplexity for the year-over-year revenue growth of a public company, and it may answer "revenue grew 18%," footnoted to the company's actual 10-K on the SEC site. You open the 10-K — it is the right filing — but the real figure is 11%, and "18%" appears nowhere in it. The footnote passes a glance-check; the number is invented. Because the citation looks impeccable, most readers never open it.
How to fix Perplexity hallucinations: manual and automated
The fix for Perplexity is counterintuitive: the citation being real is exactly why you must open it. A working link is not evidence the claim is in the page.
Manual checks:
- Open every footnote and locate the specific sentence or number in the source. If you cannot find it, the claim is a synthesis hallucination.
- Distrust precise figures most — exact percentages, dates, and dollar amounts are where synthesis errors concentrate.
- Check whether the cited page is a primary source or itself AI-generated content, to avoid second-hand hallucination.
Automated with Mephistopheles: paste any Perplexity answer into /chat. Mephistopheles does exactly what the synthesis failure requires — it checks whether the claim is actually supported by the source, independently, rather than trusting that a citation exists. It returns a per-claim verdict (supported / contradicted / disputed / unverifiable) and a hallucination-risk score. On our ~290-claim benchmark it catches ~88% of factual errors while passing ~98% of true statements clean (~2% false-alarm rate).
Frequently asked questions
Perplexity shows sources — how can it still hallucinate?
Because citing a source and being supported by that source are different things. Perplexity can attach a real, working URL to a claim the page never actually makes — a synthesis hallucination. The footnote passes a glance-check while the claim is invented. Always open the source and find the exact sentence.
Is Perplexity more accurate than ChatGPT search?
In the March 2025 Tow Center study, yes — Perplexity was the most accurate of eight engines, wrong on about 37% of queries versus 67% for ChatGPT Search. But every engine failed most queries, so "most accurate" still leaves a serious error rate. Verify high-stakes answers.
What is a second-hand hallucination?
It is when Perplexity retrieves a web page that already contains AI-generated or incorrect content, then restates that error as a cited fact. GPTZero documented this pattern in June 2024. The citation is real, but it points to a source that was itself wrong. Prefer primary sources when verifying.
How do I verify a Perplexity answer?
Paste it into Mephistopheles. It independently checks whether each claim is actually supported by its cited source, catching synthesis hallucinations that a working link disguises, and returns a per-claim verdict with a risk score.
Verify what your AI just told you.
Paste any AI answer and Mephistopheles checks each claim against independent sources — no sign-up to try.