By team
Check the facts in your paper before you submit
Catch the claims an AI assistant got wrong before your supervisor, reviewer or reader does.
It's late, the chapter is due with your supervisor in the morning, and the paragraph an AI assistant helped you tighten now reads beautifully. It also says, with total confidence, that intermittent fasting beats calorie restriction for fat loss when calories are matched. You aren't sure where that came from. This page shows how to check the factual claims in an essay, thesis chapter or paper against sources, what a result means for your next step, and what CiteJury doesn't do.
Preprint servers and conferences are cracking down
In January 2026 GPTZero reported scanning 4,841 papers accepted at NeurIPS 2025, one of the most selective AI conferences, and confirming 100 hallucinated citations across 51 of them. Every one of those papers had been through several reviewers.
In May 2026 arXiv announced that authors whose submissions show "incontrovertible evidence" of unchecked AI output, such as hallucinated references or leftover chatbot text, face a one-year ban, after which their next submission must first be accepted at a reputable peer-reviewed venue. Thomas Dietterich, who chairs arXiv's computing section, gave the reasoning: "If a submission contains incontrovertible evidence that the authors did not check the results of LLM generation, this means we can't trust anything in the paper."
A fake reference is the easiest failure to spot, because a paper either exists or it doesn't. The quieter failure is a sentence that sounds like the literature but isn't what the literature says: an effect stated more strongly than any study found, a finding moved to the wrong population, a myth repeated because it's familiar. That's the part of the problem CiteJury works on.
What CiteJury checks, and what it doesn't
Be clear about this before you start. CiteJury checks what a sentence claims against sources it finds. It is not a citation checker: it won't tell you whether the references in your bibliography exist, and it doesn't format or validate them. It isn't a plagiarism detector, and it doesn't detect AI-written text.
Here's what it does do. You paste a claim, or upload a chapter as a PDF or Word file up to 20 MB and pick up to 25 of the checkable claims it finds. Each claim is split into parts. CiteJury searches several independent search indexes for evidence for and against each part, favoring primary sources such as government statistics, regulators and official pages, and it reads PDFs as well as web pages. The best passages are sealed into one evidence pack. Claude, ChatGPT and Grok each read that pack and answer separately, every quote they cite is checked word for word, and a final review gives the verdict: Correct, Incorrect, Sources disagree or Can't confirm.
Here's a simple one, a claim that turns up in more student essays than anyone would like.
“Humans only use 10% of their brains.”
Incorrect
Claude, ChatGPT and Grok agree12 sources usedTook 86 seconds
It's a myth. Neuroscience sources say the whole brain is used.
Humans use only 10% of their brains
MIT, Johns Hopkins and UW call it a myth.
The other 90% is inactive
Imaging shows activity across the whole brain.
Safe way to say it
Imaging shows the whole brain is active, though not every region is busy at the same moment.
All three AIs agreed it's Incorrect, and the sources that contradicted it included pages from MIT, Johns Hopkins and the Association for Psychological Science. The safe wording says what imaging shows instead: the whole brain is active, though not every region is busy at the same moment. It took 86 seconds.
When the answer is Can't confirm
Now the fasting sentence: "Intermittent fasting beats calorie restriction for fat loss when calories are matched." Plenty has been written about fasting, so you might expect a quick Correct or Incorrect.
“Intermittent fasting beats calorie restriction for fat loss when calories are matched.”
Can't confirm
Claude, ChatGPT and Grok agree9 sources usedTook 2 minutes
The evidence found doesn't test this. It measures weight, not fat, with unmatched calories.
Fasting produces more fat loss at matched calories
No source measured it.
Fasting preserves lean mass better
Not covered by the sources.
Safe way to say it
A 2025 CU Anschutz trial found more one-year weight loss with a 4:3 fasting plan (7.6%) than daily calorie restriction (5%), but not that fasting beats it for fat loss at matched calories.
The verdict was Can't confirm, and all three AIs agreed. The headline explains why: the evidence found doesn't test this claim. It measures weight, not fat, and the calories weren't matched. No source measured fat loss at matched calories, and none covered whether fasting preserves lean mass better.
What the sources did support is in the safe wording: a 2025 CU Anschutz trial found more one-year weight loss with a 4:3 fasting plan (7.6%) than with daily calorie restriction (5%). That's a real finding, but it's a different claim from the one in the draft. Weight isn't fat, and groups that ate different amounts can't tell you what happens when they eat the same.
Read Can't confirm as a direction, not a dead end. It tells you the claim needs the primary literature, and it tells you what to look for: trials that measured body composition rather than weight alone, with energy intake matched between groups. If you find them, cite them. If you don't, soften the sentence to what the evidence supports. And notice that one source was a university news story. Cite the trial itself, not the press release.
Checking a chapter before you submit
- Find the load-bearing claims. Numbers, dates, "studies show", "the first", "the only", and anything your argument depends on. Start with the sentences an AI assistant touched.
- Paste each claim exactly as written, or upload the chapter and pick from the list CiteJury finds.
- Pick a tier. Quick uses one AI and suits background facts. Standard, with three AIs and a final review, costs about 8 credits and usually takes one to three minutes. Deep searches wider and suits contested research questions.
- Open the sources. The verdict tells you where to look. The sources, and the ruling on each (Backs it up, Contradicts it, Background, Off-topic or Unreliable), are what you actually learn from.
- Ask a follow-up. A question such as "Which source measured fat mass?" is answered only from the check's own sources, for about 1 to 2 credits.
- Go to the primary literature for anything that comes back Can't confirm or Sources disagree, using your library's databases.
- Cite what you read. CiteJury helps you find and test sources. It isn't a source to cite.
Habits your reviewers will notice
- Check each reference yourself. Open the DOI, confirm the authors, title, year and venue, then confirm the paper says what you cite it for. No sentence checker does the first half of that for you.
- Match the verb to the evidence. "Is associated with", "predicts" and "causes" are three different claims.
- Check the population. A finding in mice, or in one country's adults, isn't a finding about everyone.
- Prefer the paper to the coverage. News stories round numbers and drop caveats.
- Label preprints as preprints. Readers may assume peer review unless you say otherwise.
- Follow your institution's AI policy, and keep a note of what you used an assistant for.
If a chatbot is part of your research routine, fact-checking its answers and asking questions that come back with sources will save you the most rework. For a first look at an unfamiliar area, a sourced research brief separates what's established from what's contested.
What it won't do
- It doesn't validate citations or bibliographies. It won't catch a made-up entry in your reference list or fix your APA formatting.
- It isn't a plagiarism checker or an AI-writing detector.
- It can't read paywalled journals. Pages behind paywalls or logins are out of reach, and the result lists what it couldn't open. Scanned PDFs without a text layer, common for older papers, can't be read because there's no OCR.
- It checks text, not figures, charts or images.
- Very new work may be thin. A preprint from last week may have little coverage yet, which often means Can't confirm.
- It isn't medical advice, whatever the fasting example suggests, and a verdict is only as good as the sources found.
None of this replaces reading the literature. It makes your reading better aimed: you find out which sentences are solid, which are wrong, and which need you to go back to the papers before your supervisor, reviewer or examiner does it for you.
Questions people ask
Does CiteJury check whether my citations are real?
No. It checks what a sentence claims against sources it finds. It doesn't validate or format a bibliography, so check each reference yourself by opening its DOI or library record.
Is CiteJury a plagiarism checker or an AI detector?
No. It doesn't detect plagiarism or AI-written text. It tells you whether the factual claims in your writing hold up against sources.
What does Can't confirm mean for my paper?
It means the sources found don't settle the claim, so CiteJury won't guess. Treat it as a signal to go to the primary literature, or to soften the sentence to what the evidence supports.
Can it read journal articles?
It reads publicly available web pages and PDFs. Articles behind paywalls or logins can't be read, and neither can scanned PDFs without a text layer.
Can I cite CiteJury in my paper?
Cite the sources you read, not CiteJury. Every sentence in a result links to its source, so open it, confirm it says what you need, and cite it directly.
Check the claim before you cite it.
New accounts get 100 free credits. Failed checks cost nothing.