Can you trust AI news summaries?

Summary: No, not as a primary news source. The largest study of its kind found 45% of AI news answers carried a significant issue, with sourcing the most common failure. They hold up for evergreen explainers and background on long-running stories, and fail on breaking news, quotes, dates, and recent personnel changes. Verify every specific against the source.

AI news summaries have a significant problem nearly half the time. That number comes from the largest study of its kind: 22 public service media organizations in 18 countries, 14 languages, more than 3,000 AI responses, four leading AI assistants tested. 45% of answers carried at least one significant issue. An earlier BBC-only study from February 2025 put the figure at 51%. Half of UK adults under 35 say they trust AI to produce accurate summaries, per BBC research published in October 2025. The summaries haven't earned it.

How often do AI news summaries get it wrong?

The BBC's February 2025 study tested ChatGPT, Copilot, Gemini, and Perplexity on 100 news questions, asked them to source from BBC News, and had 45 BBC journalists rate the answers. 51% of responses had significant issues. 19% introduced factual errors when citing BBC content. 13% of direct quotes were altered or never appeared in the cited article.

The October 2025 follow-up, coordinated by the European Broadcasting Union and led by the BBC with 22 public service broadcasters, scaled the test to more than 3,000 responses across 18 countries and 14 languages. The overall rate came in at 45%, and 31% had serious sourcing problems. 20% had major accuracy issues, including hallucinated details and outdated information. Counting smaller errors too, 81% of responses had an issue of some kind. The errors spread across accuracy, attribution, context, and tone, and the researchers call them systemic.

Where do AI summaries fail most?

• Sourcing. The single largest failure mode. About a third of answers attribute claims to outlets that didn't make them, cite syndicated copies instead of originals, or fabricate links entirely.

• Outdated information. AI models train on snapshots. News changes daily. Any story that moves between the training cutoff and the query invites confidently stale facts. ChatGPT told researchers that Pope Francis was still leading the Catholic Church weeks after his death. Gemini called it a misconception that NASA astronauts had been stuck in space, despite two crew members spending nine months stranded on the International Space Station.

• Hallucinated specifics. Names, numbers, dates, and quotes invented from nothing. The BBC found that one in five answers introduced incorrect dates, numbers, or factual statements and attributed them to sources that contained no such claim.

• Missing context. A factually accurate one-sentence answer that omits the qualifier, the counterargument, or the timing detail reads as complete while misleading you.

• Opinion presented as fact. Chatbots routinely flatten the line between editorial framing and reporting. The researchers found multiple examples of chatbots adding adjectives or framing the cited source hadn't used.

A single response can hit several categories at once. The researchers counted any one of them as a significant issue.

Why is AI bad at news specifically?

Most AI tools draft emails, summarize meetings, and explain concepts with reasonable accuracy. The gap on news is structural.

• News changes faster than training data. Most models update their knowledge on a slower cycle than the news moves. When yesterday's headline contradicts last month's training, the model often picks the version it was trained on.

• Confidence calibration is broken. OpenAI conceded in a September 2025 paper that its models are rewarded during training for guessing rather than admitting uncertainty. Independent 2025 research from Columbia's Tow Center for Digital Journalism found premium chatbots produced more confidently incorrect answers than their free counterparts.

• Source attribution isn't enforced. A chatbot can cite a link without reading the article, paraphrase a syndicated copy as the original, or invent a plausible citation. You see a hyperlink and trust it.

• News writing has voice. Reporting carries judgment about what's significant, what's contested, what counts as evidence, and what to leave out. A model that flattens that voice will distort the story even when the literal facts survive.

Which AI assistants are most accurate?

The October 2025 EBU study ranked the four leading assistants by the share of responses with significant issues:

• Gemini. Significant issues in 76% of responses and sourcing errors in 72%, the worst by a wide margin.

• Copilot. Significant issues in roughly a third of responses. Sourcing errors at 15%.

• ChatGPT. Significant issues in roughly a third of responses. Sourcing errors at 24%.

• Perplexity. Best overall at 30% significant issues, tied with Copilot for best on sourcing at 15%.

A 30% error rate wouldn't be tolerated from a news source. The market tolerates it from these tools because the interface feels like a search result rather than a journalist.

When are AI summaries fine?

Where AI summaries compete with reading the underlying material:

• Evergreen explainers. What's inflation, how a parliamentary system works, the history of NATO. Settled topics with broad documentation and slow-moving facts.

• Background on long-covered stories. A war reported on for two years, an ongoing legal case, a public figure with a decade-long record. Plenty of training data, stable framing, lower hallucination risk.

• Concept translation. Explaining a technical story (a Federal Reserve decision, a climate report, a court ruling) to a non-specialist.

Where they're unreliable:

• Breaking news. Anything from the last 48 hours. The model is guessing or pulling thin sourcing.

• Direct quotes. One in eight quotes from cited articles is altered or invented. Check the source.

• Recent personnel changes. Heads of state, CEOs, military leadership. Training-cutoff drift produces confidently wrong answers here.

• Contested framing. Stories where the choice of what to call something matters. The model flattens the distinction.

• Statistics and dates. Often hallucinated.

What does a sane news stack look like in 2026?

1. Treat AI summaries as a starting point. Use them to surface a topic or get fast context, then verify against a primary source before acting on any specific claim.

2. Read the underlying article when accuracy matters. If the topic is recent, sensitive, or would change your behavior, click through to the cited source and confirm the assistant didn't invent or alter the quote.

3. Pick one human-edited input and read it daily. A defined-input source you trust, written by editors who carry the legal and reputational risk of being wrong. Format matters less than the editorial layer.

4. Distrust confident specifics. Numbers, dates, names, quotes: assume them wrong until verified. Confidence in the chatbot's voice isn't evidence of accuracy.

5. Avoid AI for breaking news. Fast-moving stories are where AI summaries struggle most. Read a human report instead.

These rules aren't common practice because the assistant interface is designed to feel like a finished answer rather than a draft to verify. The discipline has to come from you, because the product won't supply it.

The short answer

Not as a primary news source, and yes, with verification, for tasks where the failure modes don't bite. The mistake is using one tool for both jobs and assuming a fluent voice is doing an editor's work, when an editor brings judgment and a chatbot brings probabilities.

How informed.now thinks about AI in the news stack

We built informed.now on the bet that a weekday brief, written and edited by humans and sent to a channel without an algorithm, solves a different problem than the AI assistant does. The assistant is built to sound fluent, and the brief is built to carry the day's signal with as little noise as possible.

The two are complementary. Read the brief at 8 AM. Ask a chatbot for background on a topic you want to go deeper on, then verify its specifics against the underlying coverage. What the data doesn't support is using the chatbot as your primary news input.

The category will improve. The studies show progress between February and October 2025 for the same four assistants. But improving and trustworthy are different things. Your defense in the meantime is the same as for any information source: pick the humans you trust, read them on a schedule, and verify the rest.