The Source Taster

ChatGPT and other LLMs are known to generate academic references that look real but don't exist — up to 40% of AI-generated citations are fabricated. Traditional plagiarism checkers can't catch them. The Source Taster is an open‑source browser extension that extracts, searches, and validates academic references in seconds. It checks each citation against five academic databases simultaneously and provides a transparent, field‑by‑field confidence score using fuzzy‑string matching (Levenshtein‑Damerau). ── Features ── ✓ Extract references from any text or PDF — paste them or drag‑and‑drop a PDF; AnyStyle or an LLM parses them into structured data. ✓ Search 5 databases at once — OpenAlex, Crossref, Semantic Scholar, Europe PMC, and arXiv. Smart early termination stops searching once a high‑confidence match is found. ✓ Field‑by‑field scoring — every decision is transparent. See exactly which fields matched (title, authors, year, journal, DOI) and by how much. ✓ Classify results instantly — Exact Match, Strong Match, Possible Match, or No Match. Each classification includes the evidence behind it. ✓ Batch‑verify entire bibliographies — check a full reference list in one pass. Average time: under 3 seconds per source. ✓ Privacy‑first — the extension is stateless. API keys are AES‑256‑GCM encrypted and stored in the backend keystore; no data persists on your machine. ── Validated Accuracy ── Tested in a controlled study (Master's thesis, University of Siegen, 2025): • 93% exact‑match rate on genuine APA references (avg score 99.11%) • 100% hallucination detection — every fabricated reference was correctly flagged • 96.22% robustness score on perturbed references • <3 seconds average processing time per reference Tested against 10 major citation styles with consistently high accuracy. ── Open Source ── The Source Taster is MIT‑licensed and available on GitHub. Self‑host the backend with Docker and bring your own AI key to use it completely free and without limits. → GitHub: https://github.com/jeffnawroth/source-taster → Documentation: https://sourcetaster.com