ProductsMenu
Resources
About

Updated

AI Copy Fact-Check Statistics: What the Research Shows

By Kaivalya, founder of Maitreya Labs. We make Copy Desk, an AI copywriting tool that checks every statement in a draft against the facts you give it. Published 20 November 2026. Updated 20 November 2026. Every figure was checked at its source on 20 November 2026.

AI makes factual errors often enough to check every draft before it goes live. In 2025, EBU and BBC found 45% of AI news answers had a significant issue. OpenAI's system card showed o3 at 33% on questions about people. Rates fall lowest when the model works from text you supply.

Key Statistics at a Glance

Eleven figures from named public studies. Each comes from a different task, so they measure different things and cannot be averaged into a single number.

FindingFigureSourceYear
News answers with a significant issue45%EBU and BBC2025
AI search: incorrect source citationsOver 60%Tow Center, CJR2025
o3 hallucination rate on PersonQA33%OpenAI system card2025
Hallucination range across 26 models22% to 94%Stanford HAI, AI Index2026
Summaries adding unsupported facts (lowest)1.8%Vectara leaderboard2026
Legal questions hallucinated58% to 88%Journal of Legal Analysis2024
Legal tool incorrect answersOver 17% to over 34%Stanford RegLab, HAI2024
AI citations made up (GPT-3.5)55%Scientific Reports2023
Court decisions on AI-invented content2,313 casesAI Hallucination Cases2026
AI finance articles needing corrections41 of 77CNET / The Verge2023
Fake references in a government reportOver a dozenThe Guardian, AP2025

AI Answers on News and Search

Two large 2025 studies put hard numbers on how often AI tools get news answers and source citations wrong.

EBU and BBC, "News Integrity in AI Assistants", published October 2025. Twenty-two public service media organisations in 18 countries and 14 languages took part. Journalists checked 2,709 answers about the news from ChatGPT, Copilot, Gemini and Perplexity. Forty-five percent of answers had at least one significant issue, and 81% had an issue of some form. Sourcing was the biggest cause: 31% had significant sourcing problems, with missing, misleading or wrong attribution. Twenty percent had significant accuracy problems, such as made-up details or wrong quotes. Every assistant fell between 18% and 22% on accuracy.

Tow Center for Digital Journalism, published in Columbia Journalism Review, March 2025 (Klaudia Jazwinska and Aisvarya Chandrasekar). Eight AI search tools were given excerpts from news articles and asked to name each source: headline, publisher, date and address. That produced 1,600 queries in total. Together, the tools gave incorrect answers to more than 60% of queries. Perplexity answered 37% incorrectly. Grok 3 answered 94% incorrectly.

Model Benchmarks and Why the Rates Differ

Hallucination rates in published benchmarks range from 1.8% to 94%. The task tested explains most of that spread.

OpenAI, "o3 and o4-mini System Card", April 2025. On PersonQA (questions about public facts about people), the hallucination rate was 16% for o1, 33% for o3 and 48% for o4-mini. On SimpleQA (4,000 short fact questions), it was 44% for o1, 51% for o3 and 79% for o4-mini. OpenAI noted that o3 makes more claims overall, so it gets more right and also more wrong. A newer model generation carries no guarantee of lower error rates.

Stanford HAI, AI Index Report 2026 (Responsible AI chapter). On a new accuracy benchmark, hallucination rates across 26 top models ranged from 22% to 94%.

Vectara hallucination leaderboard, updated 22 October 2026. Models summarise a given article from a set of over 7,700 articles. The leaderboard counts summaries that include claims absent from the source article. The lowest rate was 1.8%. GPT-4o was at 9.6%. Claude Haiku 4.5 was at 9.8%. Even with the source text in front of it, a model adds facts.

That range, 1.8% to 94%, follows the task. Summarising a text you supply gives the lowest rates. Answering from memory, naming a source or citation, and questions about people or court cases give the highest. Figures from different studies measure different tasks, so they stay separate on this page.

If you are picking a model for copy, ChatGPT vs Claude for copywriting compares the two for copywriting.

Made-Up Sources, Citations and Case References

Citations, sources and legal questions show some of the highest rates on this page.

Matthew Dahl, Varun Magesh, Mirac Suzgun and Daniel E. Ho, "Large Legal Fictions", Journal of Legal Analysis, 2024. The researchers asked models checkable questions about real US federal court cases. The models hallucinated between 58% (ChatGPT 4) and 88% (Llama 2) of the time.

Stanford RegLab and Stanford HAI, May 2024. Over 200 open-ended legal questions were put to legal research tools built for lawyers. Lexis+ AI and Ask Practical Law AI gave incorrect information more than 17% of the time. Westlaw AI-Assisted Research gave incorrect information more than 34% of the time.

William H. Walters and Esther Isabelle Wilder, Scientific Reports, 2023. ChatGPT wrote short papers on 42 topics. The 84 papers held 636 citations. Fifty-five percent of GPT-3.5 citations were made up. Eighteen percent of GPT-4 citations were made up. Of the real citations, 43% (GPT-3.5) and 24% (GPT-4) had substantive errors.

AI Hallucination Cases database, maintained by researcher Damien Charlotin, checked 20 November 2026. The database holds 2,313 court and tribunal decisions dealing with AI-invented content, including fake case citations.

When AI Errors Reached Print

Publishing AI-written content without a fact-check has carried a public cost. Here are two documented cases.

CNET, January 2023. CNET published 77 finance explainers written with an AI tool. After its own review, it issued corrections on 41 of the 77 (count reported by The Verge, 25 January 2023). Some needed substantial correction; others had wrong numbers and vague wording. CNET paused the tool.

Deloitte Australia, October 2025 (reported by The Guardian and the Associated Press). A report worth AU$440,000 for Australia's Department of Employment and Workplace Relations held references to research papers that do not exist and a made-up quote from a court judgment. A university researcher spotted them. The department re-uploaded the report with over a dozen false references removed. Deloitte agreed to repay the final instalment of its fee. The revised report notes that a GPT-4o based tool was used.

What These Numbers Mean for Your Ad, Email and Landing Page Copy

Your copy is full of the exact slots these studies test: a price, a date, a result, a customer count, a quote. Here is what the research means for a draft going live.

  1. Copy carries the same kind of slots the studies test: prices, dates, results and counts. When your brief leaves one empty, a general AI tool fills it with something that fits the surrounding copy.
  2. The lowest error rates come when the model works only from text you supply. A draft built from your own facts, prices and proof statements starts safer.
  3. A made-up figure in an ad reaches customers before anyone checks it. The CNET and Deloitte Australia cases show the cost lands on the publisher.
  4. Named sources, quotes and citations show high error rates in the studies above: 55% of GPT-3.5 citations were made up (Scientific Reports, 2023) and more than 60% were wrong on naming news sources (Tow Center, 2025). Any study, review, case or quote your copy cites needs a manual check against the original.

For more on the mechanism behind this, see why AI copy invents numbers.

A Fact-Check Routine for AI Copy

This routine works with any tool and takes a few minutes per piece. Run it before any draft goes live.

  1. Mark every number, price, date, name, quote and claimed result in the draft.
  2. Find each one in your own source: price list, offer sheet, website or brief.
  3. Cut or source anything you cannot find in your own material.
  4. Check the currency sign and the unit on every figure.
  5. Open the original for any study, quote or citation and check the year and the exact figure.
  6. Ask the AI tool to list every claim it sourced from your brief and every claim it could not match.
  7. Read the final copy once more after the last edit, because edits add new claims.

To write ad copy from your own facts, put your brief together before you open any tool.

How Copy Desk Handles Facts

Copy Desk limits every statement in a draft to the facts you give it. After drafting, a check step confirms each statement against your brief. Any claim the tool cannot match to your brief is left out and listed in the hand-over as a [NEED: ...] line, so you can add the missing fact.

Three pieces are free to start. You add payment details later if you choose to continue. After that, $5 covers 30 days, one payment; it stops at the end of that period and renews only if you pay again. It runs on the web and inside Claude, ChatGPT, Gemini, Claude Code, Codex and Cursor through its MCP server.

How We Collected These Numbers

Eleven figures from named public studies, reports and published news reports, each checked at its source on 20 November 2026. Sources are named as plain text throughout the page. Figures from different studies measure different tasks and are kept separate: they cannot be compared or averaged.

We publish no Copy Desk accuracy rate on this page. There are too few real customer pieces yet to publish one honestly.

Questions About AI Fact-Check Statistics

How often does AI make up facts?

It depends on the task. EBU and BBC found 45% of AI news answers had a significant issue in 2025. OpenAI's system card showed o3 at 33% on questions about people in the same year. Rates on this page range from 1.8% when summarising a supplied article (Vectara, 2026) to 94% on naming news sources (Grok 3, Tow Center, 2025).

Which AI model hallucinates least?

The lowest rates come on summarising a text you supply. The Vectara hallucination leaderboard (2026) recorded the best-performing model at 1.8% on that task. Results vary by benchmark, and the leader differs from task to task.

Can I quote these statistics in my own work?

Yes. Name the source and year for each figure. Check the original study before you publish a number, since studies update and figures can be revised.

Do newer AI models make fewer mistakes?

The OpenAI "o3 and o4-mini System Card" (2025) shows o3 at 33% on PersonQA against 16% for the older o1. OpenAI explains that o3 makes more claims overall, so it also makes more wrong ones. A newer release carries no automatic accuracy improvement.

How do I fact-check AI copy fast?

Mark every number, name, date and quote in the draft, then find each one in your own source material. The full 7-step routine is in the section above.

Ready to Write Copy From Your Own Facts

Explore the free copywriting tools on Maitreya Labs. You will find free AI caption and script generators for YouTube, TikTok, LinkedIn and more, plus an ad copy length checker.

Start with 3 free pieces