You probably read several AI summaries today without deciding to. One sat above your search results. Another condensed a long email thread. A third turned a forty-page PDF into six bullet points so you did not have to open it.
According to a June 2026 Pew Research Center report, roughly 60 percent of Americans read AI search engine summaries. The summary has become the default way of meeting information, and the original has become the thing you consult if you have time, which you do not.
So it matters how often the summary is wrong, and even more how it is wrong. Several groups have now measured this. The error rate is lower than critics claim and higher than the confident formatting suggests. The more interesting finding is about the type of error, because it is one that a careful reader is badly placed to catch.
How often they are wrong
The numbers depend heavily on what is being summarised.
For general search queries, a study reported by the New York Times found that Google’s AI Overview offers correct and reputably sourced summaries 9 out of 10 times. That sounds reassuring until you consider the scale, and one further detail: an initially wrong response to a search query may transform into a correct summation when a user repeats the search. The errors are not stable, so you cannot learn which topics to distrust.
For news, the picture is worse. The BBC gave four major chatbots 100 news articles from its own site and asked for summaries, which BBC journalists then assessed. The result: 51% of the AI-generated answers exhibited significant errors. Among responses citing BBC content, 19% introduced factual mistakes, such as incorrect dates, numbers and statements. And 13% of the quotes attributed to the BBC were either altered from their original form or did not exist in the cited articles.
News is a hard case, because it is recent, detailed and full of attributed claims. But it is also exactly what people increasingly ask chatbots about.
The error that does not look like an error
The most careful study of how summaries go wrong was published in Royal Society Open Science in 2025 by researchers at Utrecht and Western University. They were not looking for hallucinations. They were looking for something subtler.
They tested 10 prominent LLMs, comparing 4900 LLM-generated summaries to their original scientific texts, drawn from journals such as Nature, Science and The Lancet. What they measured was scope: whether the summary claimed more than the paper did.
It frequently did. Six of ten models systematically exaggerated claims found in the original texts, often in ways that are easy to miss. The researchers give the canonical example: changing a cautious, past-tense claim like “The treatment was effective in this study” to a sweeping, present-tense version like “The treatment is effective.”
Every word in the second sentence is defensible. Nothing has been invented. But a finding about one group of patients in one trial has become a general truth about a drug. The caveats that scientists fight to include (in this population, at this dose, over this period) are precisely what a summariser trims, because trimming is what summarising means.
When the researchers compared chatbot summaries with summaries of the same papers written by human experts, LLM summaries were nearly five times more likely to contain broad generalizations.
Two findings that should worry you more
Newer models were worse. The paper reports that newer models tended to perform worse in generalization accuracy than earlier ones. A likely explanation is that training on human feedback tends to reward answers that sound confident and broadly relevant over more tentative ones. Models are being trained, in part, to sound helpful, and sweeping statements sound more helpful than hedged ones.
Asking for accuracy backfired. This is the result that matters most for daily use. When the models were explicitly prompted to avoid inaccuracies, they were nearly twice as likely to produce overgeneralized conclusions than when given a simple summary request. As the lead author put it: students, researchers, and policymakers may assume that if they ask ChatGPT to avoid inaccuracies, they’ll get a more reliable summary. Our findings prove the opposite.
So the instinctive defence, adding “be accurate” to the prompt, made the problem worse in this study.
Why you will not catch it
Put these findings next to how people actually behave and the risk becomes clear.
Users know the tools are imperfect. A researcher studying how people use AI for news told the Columbia Journalism Review: they know the answers they are getting are not perfect. They use them anyway, because the alternative is reading the source, and the entire point of the summary was not to.
That is the trap. A fabricated quote can be caught if you check. An overgeneralisation can only be caught by reading the original closely enough to notice which qualifiers went missing. The error is invisible from inside the summary, and the summary exists so that you never leave it.
What is not known
How much this matters in practice. The audits measure error rates in outputs. Nobody has measured how often a wrong or overbroad summary leads a real person to a worse decision.
Whether it holds outside science and news. Legal documents, contracts, code reviews and meeting transcripts have not had equivalent audits, though the mechanism (dropping qualifiers) should apply anywhere qualifiers matter.
Whether current models have improved. The scope study collected its summaries over a year ending in early 2025, and the models have changed since. Given that the trend in that study was for newer models to do worse, improvement should not be assumed.
What prompt actually helps. The authors suggest lowering the model’s temperature setting and asking for cautious, past-tense phrasing. Setting the temperature to zero reduced overgeneralisation in their tests, but that option is not available in most consumer chat interfaces.
How to use summaries without being misled
Match your trust to the stakes. A summary of a newsletter can be wrong at no cost. A summary of a contract, a medical paper, a security advisory or a colleague’s design document is a pointer to what to read, not a replacement for reading it.
Ask what was left out. Instead of “summarise this accurately”, ask for the study population, the sample size, the limitations the authors list, and what the paper does not claim. That targets the specific failure: missing scope.
Distrust the present tense. “X improves Y” is the signature of a dropped qualifier. The original almost certainly said something closer to “X improved Y in this group under these conditions”.
Verify anything in quotation marks. With 13% of attributed quotes altered or absent in the BBC audit, a quote in a summary should be treated as a paraphrase until you have seen the source.
Re-run it. Because errors are unstable, asking twice and comparing is a cheap check. Where two runs disagree, neither should be trusted.
Common questions
How accurate are AI summaries?
It depends on the material. One study found Google’s AI Overviews correct and well sourced about 9 times out of 10, while a BBC audit of four chatbots summarising 100 news articles found significant errors in 51% of answers.
What mistakes do AI summaries make most often?
Beyond factual errors, the most systematic problem is overgeneralisation. In a study of 4900 summaries from 10 models, LLM summaries were nearly five times more likely than human-written ones to broaden a finding beyond what the original supported.
Does telling an AI to be accurate make its summary better?
Not in the one study that tested it. Models explicitly prompted to avoid inaccuracies were nearly twice as likely to produce overgeneralised conclusions as when given a simple summary request.
Are newer AI models better at summarising?
Not necessarily. The Royal Society Open Science study found newer models tended to perform worse on generalisation accuracy than earlier ones, possibly because training on human feedback rewards confident-sounding answers.
The bottom line
AI summaries are right often enough to be useful and wrong in a way that is hard to see: they keep the finding and drop the conditions attached to it. Telling the model to be careful does not fix that and may make it worse. Use summaries to decide what to read, ask specifically for the caveats, and open the original whenever being wrong would cost you something.
Did this answer your question?
Thanks. That helps us decide what to fix.
Leave a comment