Faster fact-checking for reporters and editors

A number is moving fast: Oracle has laid off 30,000 people. It's in aggregator headlines and reshared posts, and now it's in the second paragraph of the draft on your screen. Nobody has linked a filing, and the desk wants the story within the hour. This page shows how to run a number like that through CiteJury, what comes back, and how to hand your editor a trail they can follow without redoing your work.

Why the checking load keeps growing

More copy now passes through an AI system somewhere on its way to you: a press office that drafts with a chatbot, a syndicated section, a summary tool. The errors that come out of that process are rarely wild. They're plausible, specific and slightly wrong, which is the hardest kind to catch on a read-through.

The evidence is piling up. In a study coordinated by the European Broadcasting Union and led by the BBC, 22 public service media organizations in 18 countries reviewed more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity. The EBU reported in October 2025 that 45% of the answers had at least one significant issue, and 31% had serious sourcing problems: missing, misleading or incorrect attributions. Audiences are moving toward those same tools anyway. The Reuters Institute's 2026 Digital News Report found weekly use of AI chatbots for news rose from 7% to 10% in a year.

It reaches print, too. In May 2025 a syndicated summer section in the Chicago Sun-Times recommended books that do not exist, in the words of the paper's own chief executive. It came from a content partner, not the newsroom, and it ran anyway.

What a check does with a claim

You paste the sentence as it stands in your draft. CiteJury splits it into the separate things it asserts, then searches several independent search indexes for each part, looking for evidence against the claim as well as for it. It favors primary sources: filings, regulators, government statistics, court records, official pages and wire reports. Social media and video pages are skipped, because they have no quotable text.

The best passages are locked into one sealed evidence pack with an ID, and nothing can be added to it later. Claude, ChatGPT and Grok each read that same pack and answer separately, with no web access at that stage, so none of them can lean on a source you can't see. Every quote an AI cites is checked word for word against its source. A final review settles any disagreement and writes the verdict, a one-line headline, a short explanation and a safe way to say it.

Here is the Oracle check running, stage by stage.

A Standard check with all three AIs usually takes one to three minutes and costs about 8 credits ($0.80). This one ran on Deep, where the AIs also run their own searches while sources are gathered, because a fast-moving layoff story rewards a wider net.

The Oracle number, traced

The claim was one sentence: "Oracle laid off 30,000 employees in September 2026." CiteJury searched 233 results, set aside pages it couldn't read, such as Reddit and Facebook posts, and locked 28 sources into the pack before the AIs read anything. Claude, ChatGPT and Grok all said Incorrect. The claim broke into four parts, and only one held:

• Oracle made layoffs in September 2026. Correct, reported by the BBC and CIO.
• About 30,000 people were laid off. Incorrect. The figure traces to a TD Cowen estimate of 20,000 to 30,000 for the full fiscal year.
• It was one announced action. Incorrect. No announcement of that size exists.
• Oracle disclosed the figure. Incorrect. No filing or statement gives it.

The headline on the result read: "Wrong source for the number: 30,000 is an analyst's range for the whole year." It's a common shape of error in fast news. The event is real, the number is real, and the two have been stuck together. The top of an analyst's range for a whole restructuring became a count for a single month.

The safe way to say it was usable almost as written: "Oracle began a further round of layoffs of undisclosed size in mid-September 2026. The widely quoted 20,000 to 30,000 figure is an analyst estimate for its whole fiscal-2026 restructuring." The check took about eight minutes.

Most errors in copy are smaller than that. We also checked the claim that Deloitte refunded its full A$440,000 fee to the Australian government over a report with AI-generated errors. The story is real. One detail isn't.

Three of the four parts held. The sources said Deloitte repaid only the final installment, so the full-refund framing was wrong, and all three AIs agreed. A correction like that is cheap before publication and awkward after it.

Checking a draft, step by step

1. Copy the sentence exactly as you'd print it. Keep the hedges and the attribution. "Up to 30,000, according to one analyst" is a different claim from "30,000", and CiteJury checks the words you give it.
2. Pick a tier. Quick uses one AI with no final review and suits a sanity check. Standard runs all three AIs plus the final review. Deep casts the widest net, for breaking or contested numbers.
3. For a whole draft, upload it. A Word file, a PDF or pasted text goes through a document check: CiteJury lists every checkable claim and you choose up to 25.
4. Read the sources marked Contradicts it first, then the parts. The headline says where to look; the sources say whether to believe it.
5. Ask follow-ups. A question such as "Which source first used the 30,000 figure?" is answered only from the same sealed sources, for about 1 to 2 credits.
6. Share the link with your editor. Anyone with the link sees the check and every source behind it, but not your follow-up questions.

That last step saves the most time. Instead of "trust me, I checked", your editor gets the verdict, the reasoning for each part and every source with a ruling: Backs it up, Contradicts it, Background, Off-topic or Unreliable. Sources set aside are listed with the reason, such as "Couldn't open". You can also download the sources as Excel or CSV for the story file.

Habits that make it work on a desk

• Check the attachment, not just the number. Ask what period, what population and who measured it. Numbers often fail because they've drifted from their context, as the Oracle figure did. Our guide to tracing a statistic back to its source goes further.
• Go to the document the story is about. For layoffs that means company filings and state WARN notices; the Oracle run turned up edd.ca.gov, where California publishes them. For a court story it's the docket. For an economic number it's the statistics office.
• Treat Can't confirm as information. If a claim that's all over social media comes back Can't confirm, the missing primary source may be the story. It can also just mean the event is hours old and coverage hasn't caught up.
• Look at what couldn't be opened. Paywalled outlets and sites that block automated access show up in the set-aside list with a reason. Your newsroom's own subscriptions can often fill that gap.
• Use the safe wording as a draft, not a quote. It's written to match the sources, but the sentence you publish, and the source you credit, are still yours.
• Notice planted text. CiteJury flags text hidden in a page and aimed at AI, which tells you something about the page.

The same approach works on a chatbot answer a colleague pasted into the shared doc; see fact-checking AI answers.

What it won't do

CiteJury doesn't replace reporting. It can't call a source or tell you whether a spokesperson is being straight with you. A verdict is only as good as the sources it found, and it can be wrong or incomplete, which is why every sentence links to its source.

• It checks text, not images, charts or video. It won't tell you whether a photo was altered or a clip is genuine.
• It skips social media and video pages, so it can't confirm what a post or clip itself said. It can check what reliable outlets reported about it.
• Pages behind paywalls or logins, and some sites that block automated access, can't be read. You'll see which ones.
• Scanned PDFs without a text layer can't be read, because there's no OCR.
• Very recent events may have little coverage yet, which often means Can't confirm.
• It isn't legal advice, so it's no substitute for a lawyer's read on a risky story.

Used well, it takes on the slow half of the job: finding the primary sources and showing exactly where a claim bends. The call on what to publish stays with you and your editor, who can now see what you saw.

Questions people ask