Blog ·
How can you tell if an AI-written article is accurate?
Check it claim by claim: open the source cited for a claim, find the passage that supports it, and check that the source is current and independent. An article with no sources you can open cannot be checked this way, however confident it sounds. A polished style tells you nothing, because AI writes smoothly whether it is right or wrong.
How often is AI wrong about facts?
Often enough that checking is worth it, though the rate depends heavily on the task. One large test is about news: public broadcasters coordinated by the European Broadcasting Union and led by the BBC had journalists from 22 organizations in 18 countries evaluate more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity. Published on Oct 22, 2025, it found:
| Problem | Share of answers |
|---|---|
| At least one significant issue of any kind | 45% |
| Serious sourcing problems (missing, misleading or incorrect attribution) | 31% |
| Major accuracy issues, including invented details and outdated information | 20% |
Two cautions about that table. It measures chat assistants answering news questions, so it is not an error rate for AI-written articles in general. And assistants change quickly, so a 2025 figure describes them as they were then.
The finding that matters most here is the second row. The full report calls sourcing the single biggest cause of significant issues. So a citation being present settles nothing: you have to open it.
What should you check, and in what order?
- Is there a source for the claim? Look for a link or footnote at the sentence itself. A list at the bottom shows what the author read, not which claim came from where.
- Does the source exist and open? A dead or invented link is a warning about the whole article.
- Does the source say it? Find the passage. A real source attached to a sentence it does not support is the failure this step exists to catch.
- Is the source any good for this claim? A company is a fair source for its own price list and a poor one for how good its product is.
- Is it current? Check the date on anything that can change: prices, records, office holders, laws.
- Do other sources agree independently? Five articles repeating one press release count as one source.
You do not need to do this for every sentence. Check the claims you plan to rely on, plus two or three chosen at random. If those fail, distrust the rest.
Why check claim by claim instead of judging the whole article?
Because one article mixes true and false statements, and a single verdict hides that. Researchers made this point formally with FActScore, a measure that breaks generated text into atomic facts and computes the percentage supported by a reliable knowledge source.
Human reference works reach for the same principle. Wikipedia's verifiability policy requires an inline citation, to a reliable source that directly supports the material, for quotations and for anything challenged or likely to be challenged.
Is an AI-written article worse than one written by people?
There is no general answer we can stand behind. Human reference works make mistakes as well: when Nature compared science entries in 2005, it reported that Wikipedia came close to Britannica for accuracy, and the same page notes that Britannica disputed the results. The practical point holds for both: judge the sourcing of the claim in front of you, not the kind of author.
How does Heartipedia use this?
We built this approach into how a page is made, and it has limits we would rather state than hide:
- A page's draft is written from a sheet of facts taken from its sources, and each claim carries a numbered citation you can follow.
- AI graders then read the sentences against the cited sources. Sentences they find unsupported are rewritten or removed.
- The check is still AI checking AI. Today the checking models come from the same provider as the drafting model, and each page's "About this page" box names the models that ran. As our about page says, no page has been reviewed by a named person yet.
So please use the checklist on us too. Pick a page, open a citation and see whether it holds. If it does not, suggest an edit from that page: that is how it gets better.
