# How accurate is ChatGPT on today's news?

> "What happened today?" The question people want ChatGPT to answer most is the one it answers worst. Where it fails, where it holds up, and a test to run before you trust it.

Source: https://informed.now/blog/chatgpt-current-events-accuracy

By Kira Shishkin · May 23, 2026 · AI News Quality

**Summary:** ChatGPT is reliable on background and settled events and unreliable on what happened today. In a 2026 Stanford test it answered 85% of same-day news questions correctly, 19% when the question carried a misremembered detail, and most of its errors came from fetching the wrong source or none. Read the reporting first, use the model for context, turn on browsing, and click every citation.

Ask ChatGPT why the Federal Reserve exists and you get a decent tutor. Ask it what the Fed did this morning and you're trusting a search it ran, a page it picked, and a sentence it pulled out, with no way to see which of the three failed. The question people bring to it most often is the one it handles worst.

That gap comes from how the system knows anything. ChatGPT has two sources: The training data baked into the model, which ends months before the day you ask, and live browsing, which searches the web, opens a few results, and writes from whatever it opened. Closed book, the model has nothing on last week and writes fluently anyway. With browsing on, each step has its own failure rate, and in Stanford's February 2026 evaluation, retrieval failures (the wrong page, or no page) drove over 70% of all errors. A wrong answer about today and a right one arrive in the same confident tone, and the model rarely says "I don't know."

### How accurate is ChatGPT on same-day news?

Stanford researchers [**tested**](https://arxiv.org/abs/2605.22785) six chatbots from February 9 to 22, 2026, on 2,100 factual questions written from same-day BBC reporting across six regional services. The best system answered 95.6% correctly. GPT-5, the OpenAI model in the test, scored 85.0%, behind every non-OpenAI model in the study. GPT-4o mini, an older and smaller model, scored 69.0%. On the same questions, a GPT-5 answer was roughly three times as likely to be wrong as an answer from the top system.

The harder test was a set of questions with a subtly false premise, the kind of detail a reader misremembers: A wrong date, a swapped name, or an outcome that hadn't happened. GPT-5 answered those correctly 19.0% of the time, which means it mostly went along with the false detail instead of catching it. The researchers single that number out because of how many people ask ChatGPT about the news.

### Where does ChatGPT get the news wrong?

- **Sources.** In a March 2025 Columbia Journalism Review [**test**](https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php) of eight AI search tools on 200 articles from 20 publishers, ChatGPT wrongly identified the source of 134 of the 200 excerpts, signaled doubt 15 times, and never declined to answer. Across all eight tools, more than 60% of answers were wrong.
- **False claims.** NewsGuard's January 2026 [**audit**](https://www.newsguardtech.com/ai-monitor/january-2026/) of the 11 leading chatbots found them repeating false claims on controversial news topics more than 28% of the time.
- **Guessing instead of passing.** Across the 6,000 knowledge questions in a November 2025 [**benchmark**](https://arxiv.org/abs/2511.13029), GPT-5 took a guess on 81% of the ones it couldn't get right instead of saying it didn't know.
- **Stale facts in the present tense.** With browsing off, the model describes officials who've left office, prices that have moved, and products that have shut down as if nothing changed since its training data ended.

All four come from the same design. The model is built to produce fluent text, the retrieval layer returns whatever is indexable, and no step in between checks a claim against a source.

### When is ChatGPT accurate on news?

ChatGPT is accurate on news whenever the facts already sit in a fixed text:

- **Background and history.** A primer on a regulator, a country, or a person you don't know is what the model does best, because the training data on settled subjects is deep and consolidated.
- **Explainers on settled events.** Last year's court ruling, a policy change from before the cutoff, or a scientific finding already in the literature.
- **Summarizing an article you paste in.** On Vectara's hallucination [**leaderboard**](https://github.com/vectara/hallucination-leaderboard), updated September 22, 2026, the GPT-5 family invents material in roughly 7% to 11% of document summaries. That's far from zero, and far better than asking the model to find the article itself.
- **Drafting the questions to check.** Angles, sources to look up, and people to ask. The model scaffolds the reporting without doing it.

The failures begin when you ask the model to be the first witness to something that happened after its training data ended.

### What about the times it gets today right?

It often does, and 85% on same-day questions is a high number for a tool that costs nothing. The biggest events get the most coverage, which is what retrieval finds first, so for a mainstream story a few hours old with wire coverage behind it, ChatGPT with browsing will usually get the basics right. We can't tell you the odds on your particular question.

The other 15% doesn't announce itself. Nothing in the answer marks it as one of the misses, and the way people ask about the news, half remembering a name or a date, is exactly the case where accuracy fell to 19%. The costs are lopsided: A right answer saves you a minute of reading, and a wrong one you repeat in a meeting.

### How do you use ChatGPT for news without getting burned?

1. **Read the reporting first, then ask for context.** Open the story from a newsroom that did its own reporting, then ask ChatGPT what the agency in paragraph three does or what happened the last time this law came up.
2. **Turn on browsing for anything recent.** The closed-book answer reads identically to the grounded one and is wrong far more often.
3. **Ask for sources and click every one.** A model that can't produce a real page for a claim is bluffing.
4. **Give it the text.** Paste the article and ask for a summary, the strongest counterargument, or the three claims worth checking. The text in the prompt is the most reliable thing the model has.
5. **Keep it off the last two days.** For anything that happened since yesterday, go to a newsroom first.

Run the test yourself. Pick a verifiable event from the last three days (a vote, a court filing, or a product launch) and ask ChatGPT to summarize it with sources. Then read the same story on a news site. You'll get one of three results: The model got it right, it missed detail without inventing any, or it wrote something confident and wrong. The first means the tool works for that kind of question, the second means you fill in the gaps, and the third means you stop asking it there. The same research across every major tool, ChatGPT included, is in our piece on AI news [**summaries**](https://informed.now/blog/can-you-trust-ai-news-summaries).

### A human brief answers the question the model fails

The question ChatGPT handles worst, what happened today in a few accurate minutes, is the question a human-made brief exists to answer. That's what we write at [**informed.now**](https://informed.now/): A short news research briefing each weekday morning, built from primary sources about that day's events, with follow-up questions answered from that brief instead of from a model's months-old memory. A good morning paper does the same job on newsprint. Our [**curated**](https://informed.now/blog/curated-news-vs-algorithmic-feeds) news piece covers what a human editor adds that a model can't.

ChatGPT is a strong reader and a weak witness. Give it a text and it explains that text well. Ask what happened while you slept and you get a confident sentence with no way to tell which step behind it failed. The one news question everyone asks it is still the one to ask a reporter first.