Reference checker
Check every reference in a thesis before you submit it
A reference list can look right and still be wrong. A reference checker has to answer three questions about each entry: does the work exist, are its authors and year right, and does the sentence citing it say what the work says. A citation-style checker only answers a fourth, whether the commas are in the right places.
hedwigai answers the three. Give it the whole draft, from Word, Google Docs, LaTeX or a PDF turned into text. It looks up every entry under References or Bibliography in Crossref and on the web, then reads each sentence that cites one, whether as [3], (Smith et al., 2020) or a footnote, against the work found. Your words come back exactly as written, with a verdict and its source beside each.
Why an AI-written draft needs it
More theses, literature reviews and grant proposals are now drafted with an AI, and AI writing tools produce references that look exactly like real ones. Walters and Wilder checked all 636 works cited in 84 papers ChatGPT wrote. More than half of GPT-3.5’s references did not exist. GPT-4 did better, but still invented almost one in five, and a quarter of its real references had the wrong year, authors or pages.
| GPT-3.5 | GPT-4 | |
|---|---|---|
| References that don't exist | 55% | 18% |
| Real references with an error in authors, title, year, journal or pages | 43% | 24% |
Newer models invent fewer, but none invents none. And a sentence can cite a real paper that says something else. Your examiner checks by hand, so check first.
What it says about each reference
| Verdict | What it means |
|---|---|
| Found | A record with this title whose first author and year agree, the year within one. The record is linked. |
| Found, but | Crossref has the work, but its year or authors differ from the entry, and the note says how. |
| Not found | Neither Crossref nor a scholarly site on the web has it. The note asks you to check it exists as written. |
It tries the DOI first, when the entry gives one. Then it searches Crossref for the title and reads every record with that title, not just the top hit, because titles get reused. When we built it, the top Crossref hit for “Language models are few-shot learners” was a 2026 record with the same title, not the 2020 paper. Only when Crossref has nothing does it search scholarly sites on the web for the exact title.
It never calls a reference fake. Books, theses, reports and very new or non-English work are often missing from Crossref, so “not found” means look it up yourself, not delete it.
Each sentence that cites a reference gets its own verdict: cited when the work states it, contradicted when a source says otherwise, with the passage quoted, and no source when nothing read states it. A sentence whose reference wasn’t found says so.
One draft, checked
Three sentences and three references: one correct, one with the wrong year, and one we made up for the test. The check took 53 seconds.
Multi-layer neural networks learn layered representations of data [1]. A 2023 study reported that GPT-3.5 fabricated more than half of the works it cited [2]. Spectral methods on citation graphs flag invented references [3]. ## References 1. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521(7553):436-44. 2. Walters WH, Wilder EI. Fabrication and errors in the bibliographic citations generated by ChatGPT. Sci Rep. 2020;13:14045. 3. Halvorsen K, Delacroix M. Spectral sparsification of citation graphs for detecting fabricated references. J Sch Integr. 2021;14(2):88-104.
| Reference | Verdict | Note |
|---|---|---|
| [1] LeCun, Bengio, Hinton (2015) | Found | Nature, 2015, doi:10.1038/nature14539. The sentence citing it was read against the paper and marked cited. |
| [2] Walters, Wilder (2020) | Found, but | “Crossref has it, but it was published in 2023.” The paper is real; the year is wrong. |
| [3] Halvorsen, Delacroix (2021) | Not found | “Not found in Crossref or on the web — check that this reference exists as written.” The sentence citing it is marked as citing a reference that could not be found. |
Run it from the agent you already use
In Claude, connect hedwigai from the connector directory, attach your draft and ask it to check the references. In Claude Code, Codex, Cursor or any agent that reads skills, add the reference-checker skill, which tells the agent how to turn your file into markdown, start the check and report what it found:
npx skills add un-seen/hedwigai-skills --skill reference-checker
A check is billed per reference. A thesis with 80 references costs a few dollars, and your first workbook is covered by sign-up credit.
- Study figures: W. H. Walters and E. I. Wilder, “Fabrication and errors in the bibliographic citations generated by ChatGPT”, Scientific Reports 13, 14045 (2023).
- Reference [3] in the example does not exist: we wrote it to test the check. References [1] and [2] are real papers.
- Reference styles read: APA, IEEE and Vancouver, numbered or author–year.
More head-to-head evals
| Measure | SubQ | jev | Decider | Nimble |
|---|---|---|---|---|
| Found, mean of five columns at section length | 64% | 65% | 66% | 68% |
| Mechanism of action, found at section length | 57% | 51% | 61% | 54% |
| Measure | ask | Apollo | treg |
|---|---|---|---|
| Usable leads | 13 of 40 | 7 of 40 | 6 of 28 |
| Cost per usable lead | $0.037 | $0.27 | $0.23 |
| Measure | ask with jev | ask with Decider |
|---|---|---|
| Questions right, of 92 | 89 | 89 |
| Wrong answers | 1 | 2 |
| Measure | jev | ask |
|---|---|---|
| Our 370 clinical decisions, right | 94.6% | 93.0% |
| PubMedQA, right (best EdgeEvals scored: 62.2%) | 77.0% | 74.4% |
| Measure | ask | Claude Opus 5.5 |
|---|---|---|
| Questions right, of 95 | 59 | 63 |
| Small companies right, of 20 | 13 | 9 |
| Measure | realestate-us | Best of four assistants |
|---|---|---|
| Floor-plan rents right, of 26 | 26 | 23 |
| Amenities claimed the listing doesn't state | 12 | 129 |
| Measure | ask | jev |
|---|---|---|
| Right, of the questions answered | 98% | 79% |
| Wrong answers, of 92 | 2 | 19 |
| Measure | Trained localiser | Best of four frontier models |
|---|---|---|
| Mean overlap with the defect (IoU, 0 to 1) | 0.421 | 0.109 |