✅ verifAI — catch hallucinated citations & AI-text leftovers before reviewers do

(formerly RemoveAI)

▶️ Open the checker — free, no signup · then hit "Try the demo"

arXiv (and a growing list of venues) now rejects papers containing obvious AI-generated text — chat artifacts like "Certainly! Here is your revised introduction" — or AI-fabricated citations. Reviewers screenshot these on social media. Careers get hurt by a single leftover "I hope this helps!" in a PDF.

verifAI checks your paper before anyone else sees it, and fixes what it finds. Built by Prof. Zangir Iklassov and Kyriakos Tsourekas (student).

What it catches

	Examples	What happens
🗣 Chat artifacts	"As an AI language model…", "I hope this helps!", "You're absolutely right", "Here is your revised…", leftover "Copy code"	sentence removed
🕳 Placeholders	"[insert citation here]", "(Author, Year)", "[Your Name]", lorem ipsum	removed
📋 Paste artifacts	markdown **bold** / ``` inside LaTeX, invisible Unicode, curly quotes	converted / removed
❌ Fabricated citations	title unknown to Semantic Scholar + CrossRef + OpenAlex + DBLP, DOI that doesn't resolve, real title with hallucinated authors	entry removed from .bib and its \cite{…} pruned from the .tex
🔧 Broken citations	wrong year / DOI / venue / misspelled authors	corrected from the matched database record
✍️ Style tells	em-dash density, "delve into", "ever-evolving landscape"	flagged for review only — never auto-removed

Also reported: keys cited but missing from the .bib, entries never cited, duplicate keys.

How it works

• 100% client-side. Your files never leave the browser. Only citation titles/DOIs are sent to the four public scholarly APIs. Nothing is stored anywhere.
• Zero generative AI. Detection is 45+ deterministic key-phrase/regex rules; verification is database lookups with fuzzy title matching + author/year cross-checks. Every finding is reproducible and explainable.
• Minimal diffs. Fixes are spliced into your original source — no reformatting. Download fixed files individually or as a zip, plus a report.md problem table.
• Supports .tex (natbib and biblatex commands, embedded thebibliography), .bib (@string, concatenation, all the weird stuff), .bbl, .txt, and .pdf (text + reference extraction; report-only — recompile from fixed sources).

Engineered to not cry wolf

False positives destroy trust, so the checker is deliberately paranoid about itself:

• A paper about LLMs quoting "As an AI language model" in verbatim/lstlisting/\verb is skipped; the same phrase inside quotation marks is downgraded to "needs review".
• "you are right-censored data" ≠ "you are right" — technical compounds are guarded.
• Books, theses, websites, software: absence from scholarly APIs proves nothing → marked unverifiable, never auto-removed.
• Off-by-one years are optional fixes (arXiv preprint vs. published version differ routinely).
• Same-title collisions (reprints, follow-ups) are disambiguated by author overlap + year.
• Phrases split across hard-wrapped LaTeX lines are still caught.
• Every action is a pre-selected suggestion you can override before files are generated.

Tested against real arXiv papers (clean papers come out clean — the Attention Is All You Need and BERT sources produce zero high-severity findings) plus a planted-problem corpus: 216 unit tests + Playwright end-to-end suites, run in CI on every push.

Honest limitations

Regex catches artifacts of AI generation, not AI-written prose in general — a carefully edited AI text will pass, and absence of findings is not proof of human authorship. Google Scholar has no API; arXiv's API, ACM DL, and ResearchGate block browser requests — coverage comes from Semantic Scholar + CrossRef (carries the full ACM catalog) + OpenAlex + DBLP (carries the ACL Anthology), plus direct arXiv-ID resolution, with one-click manual-search links for the rest. Scanned PDFs have no extractable text. Every automated finding is a lead, not a verdict — you are responsible for your paper.

Contributing

Found a false positive or something that slipped through? Open an issue with the snippet — each report becomes a regression test. New detection rules are one-line PRs in js/core/detect.js.

node --test tests/*.test.mjs   # zero dependencies

If this saved your submission, ⭐ star the repo so others find it before their deadline.

MIT License.