Back to Blog

AI Hallucination Examples: Famous Cases and the Errors We Caught in Our Own AI Writer

Seven documented AI hallucination cases, plus the errors we caught in our own AI writer, a table of types and a checklist to fact-check AI content before you publish.

AI Hallucination Examples: Famous Cases and the Errors We Caught in Our Own AI Writer

An AI hallucination is a statement an AI model produces that sounds right and is false: a court case that never existed, a policy a company never had, a price that belongs to someone else. Google's guidance on generative AI content explains why this happens. Generative models "don't retrieve facts, but predict a likely sequence of words based on their training data", so their output "may contain inaccuracies (also known as hallucinations)."

That is why examples matter if you publish AI-assisted content. Hallucinations rarely look like errors. They look like the rest of the article. Google's page, last updated October 1, 2026, ends the argument with one sentence: "It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing."

Below are seven well-documented public cases, then something most posts on this topic don't have: the hallucinations we caught in our own AI writer, with the exact sentences. After that comes a table of the types, a checklist and an FAQ.

What are some famous AI hallucination examples?

Each case below has a date, what was wrong, what it cost and a type from our table of hallucination types further down.

Mata v. Avianca (2023): fake court cases

Lawyers in a personal injury case against the airline Avianca filed a brief that cited court decisions that did not exist. ChatGPT had written them. In its June 22, 2023 opinion on sanctions, the federal court in New York said the lawyers "submitted non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT, then continued to stand by the fake opinions after judicial orders called their existence into question." One cited decision, "Varghese," carried a reporter citation that belongs to a different, real case. The court fined the two lawyers and their firm $5,000. Judge Castel also wrote that "there is nothing inherently improper about using a reliable artificial intelligence tool for assistance," but that existing rules "impose a gatekeeping role on attorneys to ensure the accuracy of their filings."

Type: fabricated sources and citations.

Moffatt v. Air Canada (2024): a chatbot that invented a policy

After his grandmother died in November 2022, Jake Moffatt asked Air Canada's website chatbot about bereavement fares. It told him he could apply retroactively. The airline's policy did not allow that. In February 2024 the British Columbia Civil Resolution Tribunal found Air Canada liable for negligent misrepresentation. According to a law-firm summary of the decision, the tribunal said Air Canada "did not take reasonable care to ensure its chatbot was accurate" and rejected the airline's argument that the chatbot was a separate legal entity. The tribunal awarded damages of CAD $650.88. The amount is small. The ruling that a company answers for what its chatbot says is the lasting part.

Type: invented facts or policies.

Google Bard's first demo (2023): a wrong space fact

In a February 2023 public demo, Google's Bard said the James Webb Space Telescope "took the very first pictures of a planet outside of our own solar system." According to CNN's report, NASA says the first such image was taken in 2004 by the European Southern Observatory's Very Large Telescope. Alphabet shares fell 7.7% that Wednesday, about $100 billion in market value, a drop that coverage tied to the error and a presentation that failed to impress. A Google spokesperson said, "This highlights the importance of a rigorous testing process."

Type: invented facts.

CNET's AI-written finance articles (2023): wrong numbers

In January 2023, Futurism found that a CNET explainer written with an AI tool said a $10,000 deposit at 3% interest would earn "$10,300 at the end of the first year." The interest is $300, as Engadget reported. CNET's editor-in-chief later wrote that the site had published 77 stories with the tool, according to CBS News, and paused it. The Verge reported that CNET issued corrections on 41 of those 77 stories. Other errors CNET listed were incomplete company names, transposed numbers and vague language.

Type: numbers that don't add up.

The Chicago Sun-Times reading list (2025): books that don't exist

In May 2025 the Chicago Sun-Times printed a summer reading list in which 10 of the 15 books did not exist, each attributed to a real author. CBS Chicago reported that the section was licensed from a national content partner, whose freelance writer used an AI agent, and that the newsroom had not approved it. The paper apologized and said subscribers would not be charged for the premium edition.

Type: fabricated sources.

Deloitte's report for the Australian government (2025): invented references

Deloitte Australia was paid AU$440,000 for a 237-page report for the Department of Employment and Workplace Relations. After a university researcher said it was full of fabricated references, a revised version removed a fabricated quote attributed to a federal court judge and references to research that does not exist, and disclosed that Azure OpenAI had been used. According to the AP report, the department said Deloitte "confirmed some footnotes and references were incorrect," and Deloitte agreed to repay the final instalment of its contract. The department said the report's recommendations did not change.

Type: fabricated sources and citations.

Google AI Overviews and the glue on pizza (2024): a bad source treated as fact

In May 2024, users showed AI Overviews telling people to use glue to make cheese stick to pizza. Google's explanation, dated May 30, 2024, says AI Overviews "generally don't 'hallucinate' or make things up in the ways that other LLM products might." It attributes the glue answer to forum content and the "rocks to eat" answer to satirical content. Google changed its systems to limit user-generated content in answers that could mislead. We include it because a reader sees the same thing either way: a confident, wrong sentence that traces to a source nobody checked.

Type: a bad source repeated as fact.

How many court cases involve AI hallucinations?

The law is where hallucinations are counted most carefully. Damien Charlotin's AI Hallucination Cases database tracks court and tribunal decisions in which a party relied on hallucinated content. When we read it on November 13, 2026, it said "2292 cases identified so far." He notes the database is a work in progress and does not cover every fake citation filed.

What hallucinations did we catch in our own AI writer?

We make an AI writer, so we see these errors in our own output. In September 2026 we ran internal tests of our writing pipeline: a GA4 ecommerce how-to guide and a software review of Gojiberry AI, three runs each. At the time the pipeline worked like this. A leading model wrote the draft. A second AI step, the editor, reviewed it. A third AI step, the repair, applied the editor's suggestions. Then the fact check ran on the repaired draft. We then reviewed every sentence the fact check did not mark as supported and compared the articles to their sources.

We publish the results because these are the same kinds of errors that end up in published articles, and a fact check exists to find them. The examples come from internal test runs.

A number borrowed from the wrong company

The review compared Gojiberry with competitors. The writer first put "Not published" in Salesforge's price cell. The fact check flagged that, and the repair step replaced it with a table row that read "Salesforge | $99/month (14-day free trial)". The $99 was Gojiberry's own price at the time of the test. Salesforge's source said it "starts lower, with public pricing." Salesforge's pricing page today lists plans from $40 a month. The wrong price reached the finished article, and the fact check caught it as unsupported.

"Not published" when the price is published

Two of the three review runs said that competitors Expandi and Dripify do not publish prices, in a table cell ("Not published") and in one sentence ("neither publishes pricing"). Both do. Expandi's pricing page currently shows a $99-a-month Business plan. Dripify's shows Pro at $79 and Advanced at $99 per user per month on monthly billing. In the run with the sentence, the writer's original hedge was fair: "neither publishes pricing in the sources reviewed here." A later rule that removes internal jargon deleted the scope and left a false claim about the world.

Features nobody wrote down

One review sentence said a Chrome extension pulls contacts "from LinkedIn Sales Navigator searches." A Chrome extension exists, but Sales Navigator appears in none of the sources. In the GA4 guide, the writer said custom event names skip "the automatic reporting, benchmarking, and Google Ads dynamic remarketing integration that the standard names unlock." Reporting and remarketing are in Google's documentation. Benchmarking appeared in no source.

Invented generalisations

The repair step added "most stores send the sum of item price times quantity and report tax and shipping separately." The pattern matches Google's example. "Most stores" is a statistic nobody supplied. A review draft described the product's buyers as "a small sales team (2 to 5 reps)," a range that appears in no source.

A "correction" that made a true sentence false

The editor step claimed that GA4's purchase event is not a key event unless you mark it. The repair applied that to a correct sentence: it replaced the writer's statement with "purchase … isn't automatically treated as a conversion just because it's a recommended event: confirm it's marked as a key event in your property's Admin settings." Google's own documentation says "Google Analytics marks these events as key events by default" and lists purchase first. The writer was right. The edit made it wrong. In the GA4 test, all three articles ended up with at least one error written by the repair step.

Numbers that contradict their own example

Google's GA4 ecommerce documentation includes a purchase example whose code comment says value is the "Sum of (price * quantity) for all items," with two items: 10.01 × 3 and 21.01 × 2, a value of 72.05. In one run, the repair changed the value to 81.64 and added a comment redefining it as subtotal plus tax plus shipping. In another, the article kept 72.05 but showed only one item. The items added up to 30.03, and the example contradicted the sentence next to it.

The checker made mistakes too

An AI grader we used to score the articles claimed arithmetic errors in two totals, 72.05 and 68.48. Both were correct. The fact check itself had a high false-alarm rate at this stage. In the three GA4 runs it extracted 186 claims and left 48 unresolved. By our review, 21 of those 48 (44%) were the checker's mistake, mostly advice sentences, opinions and consequences it read too literally. Another 10 were supported only by sources the pipeline had dropped, 9 were true but in no source, and 8 were real problems. Four of those 8 were written by the repair step. This count describes the checker's flags and says nothing about how accurate the articles were. The checker also let a few wrong sentences through. In the review test it marked as supported a repair-written bullet that said a pricing figure came from third-party reviews "rather than the vendor's own pricing page," when the vendor's pricing page contained it.

What changed after we found these errors?

Most of these errors had one thing in common: an AI step after the writer changed correct text. We fixed that in stages.

  • Factual suggestions need a source. The repair step had treated the editor's factual suggestions with the same authority as a verified contradiction. Now an AI editor's factual suggestion has to be confirmed against the sources before it may change the text. In the next GA4 test, errors written by the repair step fell from four to one.
  • No new figures, no lost scope. The review test ran after that change and still showed two failure shapes: the invented Salesforge price and the hedge that lost its scope. We then stopped the repair step from adding a price, percentage or rating that no source contains, and made it keep limits such as "in the sources reviewed here."
  • A repair must not make the article worse. Since September, a fresh check compares the draft before and after any automatic repair. If the repair made things worse, it is thrown away and the writer's draft ships.

We don't claim the pipeline is error-free. Every claim in a finished ProofWrite article goes through the fact check, and what it can't confirm is shown to you under "Couldn't confirm" or "Likely wrong." We also ran a fact check on 28 articles that rank on Google's first page on October 9, 2026. Nineteen had at least one verified error, and the errors had the same shapes as the ones above: out-of-date prices, an expired tax credit still presented as available, a device limit given wrong.

What types of AI hallucination are there?

Type

Real example

How to catch it before publishing

Fabricated sources and citations

Mata v. Avianca; Deloitte's report; the Sun-Times reading list

Open every citation, reference, book or quote and confirm it exists and says what the text claims

Invented facts or policies

Air Canada's bereavement-fare answer; Bard's telescope claim; the "Sales Navigator" extension

Check each claim against the owner's own page or a primary source

Borrowed numbers (a fact moved to the wrong entity)

Salesforge shown at Gojiberry's $99

For each price or figure, find the page of the entity named and confirm the number belongs to it

False absence claims

"Neither publishes pricing" about Expandi and Dripify

Search for the thing before writing "not published." If you can't find it, say "we couldn't find it"

Stale facts presented as current

A federal solar tax credit that ended in 2025, still described as available (from our 28-article test)

Check dated facts, such as prices, fees, rules and tax credits, against the official page as it reads today

Invented generalisations

"Most stores send…"; "a small sales team (2 to 5 reps)"

Ask who measured it. Attribute it, or cut it

"Corrections" that introduce errors

GA4's purchase event wrongly described as not a key event by default

Treat every AI edit as new text and re-check it, especially edits made to fix a factual issue

Numbers that don't add up

CNET's $10,300 interest on $10,000 at 3%; the 72.05 / 81.64 / 30.03 examples

Recompute every total, percentage and example yourself or with a calculator

A bad source treated as fact

AI Overviews and the glue on pizza

Check who published the source: satire, forum posts and user-generated pages are not evidence

Why does AI hallucinate?

Google's explanation is that generative models predict "a likely sequence of words based on their training data." A fluent sentence and a true sentence come from different processes, so the model can produce the first without the second.

A 2025 paper, Why Language Models Hallucinate by Kalai, Nachum, Vempala and Zhang, adds that the incentives are part of the problem. Its abstract says "the training and evaluation procedures reward guessing over acknowledging uncertainty." A model rewarded for guessing learns to give an answer when it has none.

Our own errors fit the same picture. The review had no competitor prices in its sources, and the writer filled the gap with "Not published." If you want model-by-model numbers, we compare which models and chatbots hallucinate least in a separate post.

How can you catch AI hallucinations before you publish?

Go through the draft once per type.

  • Citations and quotes. Click every link. Confirm each cited study, case, book or quote exists, and that the page says what the sentence claims.
  • Names and numbers. For every price, date, percentage and statistic, find the entity's own page. Confirm the number belongs to the entity in the sentence.
  • Words like "never," "always," "most," "neither" and "not published." Each one is a claim about the whole world. Search for a counterexample, or soften the wording to what you actually checked.
  • Dated facts. Prices, fees, laws and product plans change. Check the official page as it reads today, and write the date you checked.
  • Totals and examples. Recompute sums and code samples. Make sure the example matches the sentence that introduces it.
  • Edits. Re-read every change an AI made after the first draft. A fix can introduce a new error, as the GA4 key-event sentence did.
  • Metadata. Google says the review also applies to titles, meta descriptions, structured data and image alt text.
  • Who stands behind the page. Accuracy is the Trust part of E-E-A-T. A named author who checked the facts is a stronger signal than a polished paragraph.

How ProofWrite helps

ProofWrite is an AI writer that runs a fact check on every claim in a finished article. It reads the article, checks each claim against the pages it links to, the makers' own pages and the web, and groups what it found under Likely wrong and Couldn't confirm, with the source passage beside each. In a test with 29 planted errors (wrong figures, prices and names) in seven articles, it flagged all 29 and cleared none. It does not catch everything. In the page-one test above, it flagged 136 of the 213 wrong sentences we verified (64%), or 106 (50%) once we tightened one finding group. Treat a finding as a lead, read the source passage and decide.

FAQ

What is an example of an AI hallucination?

In Mata v. Avianca, lawyers filed a brief citing court decisions that ChatGPT had invented. A federal court in New York sanctioned them $5,000 in June 2023. In Moffatt v. Air Canada, a chatbot invented a bereavement-fare policy and the airline was held to it.

Why does AI hallucinate?

Language models predict likely word sequences and do not look facts up, as Google's guidance puts it. Research from 2025 adds that training and evaluation reward guessing over admitting uncertainty. Both effects push a model toward a fluent answer when it has no source.

Can AI hallucinations be prevented?

They can be reduced, but not removed. Giving the model good sources helps. Vectara's leaderboard, which tests how often models introduce hallucinations when summarizing a document they were given, lists rates from 1.8% to 24.2% as of October 22, 2026. Even the lowest rate is above zero on a task where the model is handed the source, so check the output.

How common are AI hallucinations?

No single rate exists, because it depends on the task. Some sourced numbers, each measuring something different:

  • Vectara's leaderboard (summarizing a supplied document): 1.8% to 24.2% of summaries across the models it lists, as of October 22, 2026.
  • A Lancet analysis of more than 2 million papers, reported by STAT in May 2026, found fabricated references in 1 in 458 papers in 2025, up from 1 in 2,828 in 2023.
  • Damien Charlotin's database lists 2,292 court decisions that address hallucinated content (November 13, 2026).
  • Our own check of 28 Google page-one articles found 19 with at least one verified error. Those articles were written by people, AI or both, so the figure measures published errors in general and not hallucinations alone.

Are AI hallucinations the same as plain mistakes?

For a reader, yes, since a wrong sentence is wrong either way. The cause differs. A hallucination is fluent text with no grounding in a source. A stale price is a fact that was once true. Both appear in the same articles, and both are caught by checking the claim against a current source.

Check your articles before your readers do

If you publish AI-assisted content, the cheapest protection is to have every claim checked against a source before it goes live. You can start with three articles you already published.

Check 3 articles free →

Sources

Share this post
Jussi Hyvarinen

Written by

Jussi Hyvarinen - Co-founder of ProofWrite

I built this platform to solve my own frustration with slow research and generic AI. I use it to write every article you see on this blog, including this one.

Publish articles where every claim is checked

ProofWrite researches and writes articles and checks every factual claim against its sources.
Try it on an article you already wrote.

Free Pilot: 1 article and 3 of your own articles. No credit card.

Is AI citing your site? Check free