AI news research: multiple answers still need source verification

What BBC and EBU actually measured

BBC–EBU collected four consumer assistants’ responses on 24 May–10 June 2025. Of 2,709 core-question responses, 45% had at least one significant issue, 31% a significant sourcing issue and 20% a significant accuracy issue. The study covered 22 media organizations, 18 countries and 14 languages. Categories overlap; these are not current error rates for all AI. BBC–EBU, News Integrity in AI Assistants, 2025

Tow Center’s CJR report, published on 6 March 2025, tested eight AI search tools with 1,600 requests to identify news excerpts. More than 60% of answers were incorrect. The checks concerned the article, publisher and URL: a specific attribution task, not an evaluation of every kind of news summary. Tow Center, Columbia Journalism Review, 2025

Do not add these percentages or turn them into a universal ranking. Our editorial recommendation is to check the response and its citations separately. A plausible account of an event still needs evidence for its particular number, quotation or causal explanation. A respected publication’s name beside a paragraph is a reason to inspect the article, not a substitute for doing so.

Inspect the chain behind each claim

Start with the statement you intend to publish. Open its link and find where the source actually supports it. Then establish whether the journalist reports an original observation, cites a document or repeats another outlet. If several links lead back to one press release, record that shared origin. Different domains do not automatically represent independent observations.

Record the event date, publication date and page update date separately. A page updated today may still contain a figure from a previous quarter. The summary should preserve the figure’s original period. An evaluative statement needs a speaker too: who considers the outcome successful, and what evidence supports that assessment?

Worked example: a 30% increase changes meaning

Imagine a fictional company announcement reporting a 30% monthly increase in registrations. A summary describes it as an increase in paying customers. Another outlet repeats the summary, and three AI answers repeat that account. The chain looks persuasive, but the metric has changed. All figures and participants in this teaching example are hypothetical.

Verification must recover what was measured, the period and the comparison base. If registrations rose from 1,000 to 1,300, the increase is indeed 30%. Those values do not establish how many people paid. The final briefing can report registrations while marking paid conversion as unknown. There is no need to fill that gap with a plausible invented conversion rate.

A useful second perspective checks whether the metric keeps the same definition along the chain. Finding another similar article is insufficient. A third pass can verify the arithmetic, but correct arithmetic will not repair a substituted concept. Assign source, meaning and calculation checks separately before asking for a unified account.

A concrete Consensus brief

Give Colay Consensus the original material you can access and a list of claims to check. Ask participants to compare definitions, chronology and quotation fidelity. Request a coordinator’s table linking each claim to its source and limitation, with unsupported claims listed separately. Do not assume that the mode opened links or verified freshness; establish what material was actually available.

A person should open every consequential link before publication. If the full text is unavailable, do not ask a model to reconstruct it from the headline. Find an accessible primary document, narrow the statement to what can be established or omit it. Consensus can organize comparisons, while the evidence determines which wording is justified.

Quotation requires literal checking

For a direct quotation, save the original passage and identify the speaker. Check whether removing a surrounding condition changes its meaning. Mark translations as translations and keep the original wording available. When you only need the substance, a brief attributed paraphrase is preferable to presenting a model’s polished reconstruction as somebody’s exact words.

Do not extend the credibility of one verified paragraph to an entire briefing. A short news item can combine an established fact, an interested party’s claim and a prediction. Those categories should remain distinct after compression. Ask reviewers to look for transformations such as prediction becoming fact, registration becoming purchase or one writer’s judgment becoming an organization’s position.

A practical publication threshold

A briefing is ready when a reader can trace its important conclusions to specific source passages and see the remaining limitations. A team can keep a simple verification record containing the wording, URL, access date, reviewer and status. This is a proposed editorial practice that can be maintained manually, not a claim about a separate Colay feature.

Take a current draft and inspect its five most consequential claims. Use Consensus to locate differences among the supplied documents, then confirm those differences yourself. If no problem is found, retain the record of what was actually checked. A quiet discussion does not establish that every statement in a document is correct.