On this page
A headline says a food "boosts memory" or a new drug "could end" a disease. You want to know what the scientists actually found. An AI chatbot can help you pull a news story apart quickly, as long as you treat it as a sorting tool and not as the final word. The original paper is the better guide.
Why does science news get exaggerated?
Much of the hype starts before a journalist writes anything. A study in the BMJ looked at 462 press releases on health research from 20 UK universities, issued in 2011, and the 668 news stories linked to them (Sumner and colleagues, BMJ 2014).
40% of the press releases gave readers advice that went beyond the paper.
33% turned a correlation into a cause-and-effect claim.
36% stretched results from animals to humans.
When the press release exaggerated, news stories often repeated it: 81% of stories carried an exaggerated causal claim, against 18% when the release was accurate (full text on PubMed Central). So the hype in a story on your phone may have come from the university itself.
Can AI summarise a science article accurately?
Not reliably, and its errors tend to lean one way. Uwe Peters and Benjamin Chin-Yee compared 4,900 summaries written by 10 large language models with the scientific texts they summarised (Royal Society Open Science, 2025). Even when prompted for accuracy, most models made the conclusions broader than the original text did. DeepSeek, ChatGPT-4o and LLaMA 3.3 70B overgeneralised in 26-73% of cases. Compared with summaries written by people, the AI summaries were nearly five times more likely to contain broad generalisations.
That is a similar error to the one press releases make: "in this group, under these conditions" becomes "people". So ask AI to find and label claims, not to rule on whether they are true, and keep the original source as the judge.
How do you use AI to understand a science news article?
Copy the full article text, not just the headline, and paste it with this prompt:
Asking it to quote the article matters. A quote is something you can check line by line; a paraphrase can quietly drift.
Here is a made-up story to practise on: "Eating dark chocolate daily sharpens memory, scientists say." The article reports that people who said they ate more dark chocolate scored higher on a memory test, and that in a separate experiment mice fed cocoa compounds learned a maze faster.
A good sort looks like this:
Finding: people who reported eating more chocolate scored higher on one memory test; mice given cocoa compounds learned a maze faster.
Evidence: an observational survey of people (nobody was assigned to eat chocolate) and an animal experiment. The article gives no sample size.
Hype: "sharpens" claims a cause from a correlation; "daily" is advice the study did not test; the mouse result is written as if it applied to people.
Those three hype labels match the three kinds of exaggeration the BMJ study counted. To practise spotting the second one, try our guide to testing correlation and causation claims.
How do you check the original source?
The AI's sort tells you what to look for in the paper.
Find the paper. The story should name the journal, the lead author or a DOI. If the AI gives you a link, open it: chatbots sometimes invent references, so learn how to check AI-made citations.
Read the abstract. It is short and usually free to read even when the full paper is behind a paywall. Typing the DOI after doi.org/ takes you to the paper's page.
Compare four things: who was studied (cells, animals or people), how many, whether the design can show cause, and the exact verb the authors use. "Was associated with" is not "causes".
Write down what changed, in one line: "The paper says X; the article says Y."
Science stories often lean on ideas you have not met yet, such as odds ratios, control groups or how a particular cell works. When a term blocks you, look it up in the Fermi Concept Library, which has a lesson for each concept in physics, chemistry, biology, mathematics and computing. If you want to go further, Fermi's AI tutor works out what you already know, builds a plan for the session and teaches through it, checking your understanding as it goes.
How can parents and students practise this together?
Pick one science headline from the day. One person runs the prompt; the other finds the abstract. Compare the two and agree on one honest sentence about what was found. Over a few weeks you get quicker at spotting the three kinds of stretch: advice, cause and animal-to-human. For a version without AI, use our five-question check for overhyped science news.
