What sources does PlagLab check against?

Open scholarly indexes (OpenAlex, Crossref, arXiv, Semantic Scholar), open-access full texts from CORE, Europe PMC, DOAJ and Unpaywall, books and reference text on a deeper search, and an archive of previously checked papers.

Does ResearchGuru.AI use Turnitin's database?

No. It has no access to Turnitin's or iThenticate's databases, so its scores are its own and can differ from theirs.

How accurate is the ResearchGuru.AI AI detector?

On 160 held-out documents it reached AUC 0.94, flagging 91% of AI-written documents and 11% of human-written ones at the default threshold.

Are my papers stored?

Checked papers are kept in the comparison vault so resubmitted text can be found later. You can ask for a paper to be erased.

Guide

How ResearchGuru.AI checks plagiarism and AI-written text

By Aditya Sikarwar and Sanidhya Dhangar · ResearchGuru.AI · 20 October 2026

Short answer: PlagLab, the ResearchGuru.AI plagiarism checker, compares a manuscript with open scholarly literature — OpenAlex, Crossref, arXiv, Semantic Scholar, CORE, Europe PMC and the open-access full texts they point to — and with its own archive of previously checked papers. It aligns matching passages word by word and by meaning, so it finds paraphrase as well as copying, and it scores AI-written text section by section with calibrated transformer classifiers. On the public PAN-13 text-alignment benchmark it scores a plagdet of 0.86. This page explains the sources, the method, the measured accuracy and the limits.

What a paper is compared against

Open scholarly indexes: OpenAlex (about 250 million works), Crossref, arXiv and Semantic Scholar are searched with queries built from the manuscript's own distinctive passages.
Open-access full texts: CORE, Europe PMC and DOAJ are searched too, and when a candidate source has a legal open-access copy (found through Unpaywall), the full text is downloaded and compared, not just its abstract.
Books and reference text: a deeper search adds Wikipedia, Project Gutenberg and the Internet Archive.
The comparison vault: papers checked earlier and sources retrieved earlier, so text that was submitted before is still found after its web copy disappears. Your own earlier checks are excluded, so a revised draft is never matched against itself.

PlagLab does not have the subscription full texts that publishers license to commercial checkers, nor any university's private store of student papers. A passage that exists only behind a paywall can therefore go unmatched, and a similarity score can come out lower than an iThenticate or Turnitin score for the same paper.

How passages are matched

The manuscript is split into sentences. Neutral text is set aside first: references and in-text citations, URLs and DOIs, contact details, funding, ethics, data-availability and competing-interest statements, and the front matter before the abstract.
Candidate sources are found by searching the indexes above.
Each candidate is aligned with the manuscript: runs of shared four-word sequences, allowing small gaps, find copied and lightly edited text; sentence embeddings and weighted word overlap find passages that were reworded.
Seeds are extended into whole passages and must hold up as a passage — a single same-topic sentence is not enough to count as a match.
Every matched word is credited to one source only, so overlapping sources do not inflate the total, and each sentence records whether it was quoted and cited.

The similarity percentage is the share of the manuscript's words that fall inside matched passages. The report highlights each passage next to its source, so you can judge it rather than trust a single number — see how much similarity is acceptable.

How AI-written text is scored

The AI score comes from transformer classifiers trained to tell human from machine-written prose, calibrated on held-out human and AI text so that a score means the same thing across subjects. Long documents are scored in sections of about 600 words, and the AI percentage is the share of the text in sections judged AI-written — so one generated section in a human paper shows up instead of being averaged away. A separate style screen shows which sentences read as formulaic.

Measured accuracy

These figures come from the team's own benchmark runs in 2026, on test data the engines were not tuned on. They are not an independent audit, and they move when the engines change.

Plagiarism, PAN-13 text-alignment test corpus (the standard public benchmark for this task): plagdet 0.86, with false detections on 2 of 1,000 clean document pairs.
Rewritten text, internal academic test set: 77% of passages paraphrased by a large language model were found, and all of the passages lightly edited by a model; on pairs of unrelated papers about the same topic, 1.4% of the text was flagged.
AI-written text, 160 held-out documents (HC3, MAGE and Gemini-written text): AUC 0.94. At the default threshold, 91% of AI-written documents were flagged and 11% of human-written ones were flagged in error. A second check on 110 labelled texts gave AUC 0.94 with 15% false positives.
Against Turnitin on 14 real papers: all 9 papers Turnitin treated as human were also scored human, and 2 of the 4 papers Turnitin flagged for AI were caught.

Known limits

Paywalled sources without an open-access copy can be missed.
Human writing is sometimes flagged as AI, most often dense, formal academic prose. AI text that a person has heavily rewritten is often scored human. Read are AI detectors accurate? before acting on any AI score.
Very short texts give unstable AI scores; whole sections are more reliable than single paragraphs.
Scores will differ from other checkers, because each one compares against a different collection of sources and uses different thresholds.

What happens to the papers you check

Checked papers are kept in the comparison vault so that later checks can find resubmitted text. You can delete a check from your history, and ask us to erase a paper from the vault — the privacy policy explains what is stored, where, and for how long.