Sourcing Failures Outnumber Accuracy in AI News Answers

You ask ChatGPT what happened in the news today, and it gets something wrong. Your first thought is probably “hallucination.” The model made something up, right?

Turns out that’s usually not it. The biggest study on this question found something more specific, and more useful if you’re the one building the retrieval layer: most failures aren’t invented facts. They’re bad sourcing.

The 31% vs. 20% Breakdown

Forty-five percent of AI answers to news questions had at least one significant issue. That’s the headline from the EBU and BBC’s 2025 study, the biggest test of its kind. More than 3,000 responses, run through ChatGPT, Copilot, Gemini, and Perplexity, graded one by one by working journalists across 18 countries.

Split those flawed answers by type, and this is what shows up:

• Sourcing problems made up 31% of flawed answers. Missing, misleading, or just plain wrong attribution, sometimes even when the actual content was fine.
• Accuracy problems made up 20%. Outdated facts, fabricated details, that kind of thing.
• Gemini’s sourcing issue rate hit 72%, a real outlier, roughly triple everyone else.

So sourcing beat accuracy as the bigger failure by a wide margin. Which means “AI gets the news wrong” isn’t really one bug. It’s two, and you fix them differently.

Reuters Institute’s 2026 report. Chatbots do worse, just 20%. And more people keep leaning on it anyway.

Why Sourcing and Accuracy Break Differently

Most AI assistants run on retrieval-augmented generation, RAG for short. The system pulls source material first, then writes an answer from it. Two separate steps, two separate ways to fail.

Evaluation frameworks like RAGAS split this apart on purpose. One check, faithfulness, asks if the answer matches whatever got retrieved. A different check, retrieval quality, asks if the system found the right source to begin with.

A model can write a perfectly faithful summary of the wrong article, or find the right article and still botch the attribution. Either way, that’s a sourcing failure, no hallucination required.

News happens to be a nightmare domain for this kind of retrieval, for reasons that pile up fast:

• Huge volume, constantly refreshing.
• The same story gets syndicated across a dozen outlets within the hour.
• Facts that are true in the morning and wrong by lunch.
• Byline and publish date matter a lot here, and generic web retrieval mostly ignores both.

Benchmarks like BEIR test how well retrieval holds up outside whatever it was trained on, and news keeps coming up as one of the tougher categories. A system built around static documents just doesn’t know how to rank and attribute a breaking story on its own.

The Industry Tried a Legal Fix First

When AI labs looked at the sourcing problem, they mostly saw an access problem. So the first move was licensing deals.

OpenAI’s deal with News Corp is reportedly worth something like $250 million over five years, covering current and archived content from the Wall Street Journal, the New York Post, and other News Corp papers. They’ve also signed with Axel Springer, the Associated Press, Le Monde, and a dozen-plus other publishers.

Made sense on paper. Feed the model a licensed, structured source instead of an open crawl, and attribution should get cleaner.

But what the deal actually guarantees almost never shows up in the press release. Is attribution contractually required? How visible does it have to be? Does that survive the next redesign? Those clauses would actually tell you whether licensing fixes sourcing, and they’re almost never public.

Licensing Didn’t Fix Sourcing

Two separate studies went and checked. Columbia’s Tow Center ran an eight-tool test across 1,600 queries and found wrong citations more than 60% of the time, and it didn’t matter whether the publisher had a deal in place or not.

A 2026 analysis of 31 million AI citations shows exactly why. The Associated Press blocks every single OpenAI crawler in its robots.txt. And yet 80% of its citations still come from ChatGPT, because licensed content reaches the model through the contract, not the crawl.

Flip it around, and you get The New York Times. It blocks everyone and gets zero citations from ChatGPT, Gemini, and Claude. It still picked up 37,642 citations from Grok, a model with no opt-out mechanism at all.

Paying for legal access buys you the right to use the content. But it tells you nothing about whether the system can find, rank, and attribute that content correctly at the moment someone asks it a question.

What It Takes to Get Sourcing Right

If you’re building anything that answers questions using news content, this is the part that decides whether your citations hold up.

Four things separate reliable sourcing from guesswork:

1. Real-time ingestion, so you’re working off current articles instead of a stale index.
2. Consistent metadata, byline, timestamp, source domain, available across every source, not just the handful with clean feeds.
3. Deduplication, so credit lands on the original story instead of the fifth outlet that reran it.
4. Structured fields a model can cite directly, instead of guessing at attribution from a wall of unstructured text.

Building the model or the app around it is a different kind of specialized engineering work, closer to retrieval architecture than to fine-tuning. Which is probably why it keeps falling to firms building production-grade generative AI systems instead of getting bolted on at the end.

Smaller Teams Hit the Same Wall

The EBU/BBC and Tow Center research both focused on the big consumer assistants, but if you’re building something smaller (a research tool, a market-monitoring feature, a media analysis product), you hit the exact same tradeoff.

Crawl the open web broadly and coverage is great, but attribution is a mess. Lean on a narrow, licensed feed and attribution gets easier, except whole publishers just aren’t there.

Structured, real-time news data doesn’t make the tradeoff disappear. It gives you a head start instead. A feed that already has byline, timestamp, and duplicate detection handled means less retrieval work on your end than starting from raw HTML.

This is also just a different problem than what most retrieval benchmarks were built to test. That’s why BEIR-style evaluation keeps finding weak spots in domains like news, where stories move fast and get duplicated everywhere.

Frequently Asked Questions

1. What’s the difference between a sourcing failure and an accuracy failure?

A sourcing failure means the attribution is wrong, missing, or doesn’t back up the claim, regardless of whether the facts check out. An accuracy failure means the facts themselves are wrong.

2. Which one is the bigger problem, according to the EBU/BBC study?

Sourcing, by a wide margin. 31% of flawed responses had serious sourcing problems, compared to 20% with accuracy issues.

3. Do licensing deals fix citation accuracy?

Not on their own. Tow Center’s testing found licensing made no measurable difference to attribution accuracy, and separate citation data shows blocked publishers still get cited by crawlers that simply ignore opt-outs.

4. Why is news harder to retrieve accurately than typical web content?

It’s high volume, constantly changing, heavily duplicated through syndication, and full of attribution details that generic retrieval tends to skip past. Hallucination gets the headlines. Sourcing is the bigger, more stubborn problem, and no amount of licensing was ever going to be the whole fix.