Research paper detector with citation-aware scoring.

About the Research paper detector

Research papers break generic AI detectors in specific, predictable ways. Academic prose is highly formal by genre, that reduces burstiness and perplexity, which are two of the three main signals most detectors use. Heavy quotation and paraphrase of other sources add another layer of "not the author's voice" text that a naïve detector misreads as AI.

This detector is tuned for the research-paper genre. It excludes citation blocks and reference lists from scoring, handles quoted passages separately from body prose, and reports scores at the section level (introduction, methods, results, discussion) so that the formal-register-heavy sections don't pull the score for the whole paper.

What this detector knows

• Citation-aware. Inline citations (APA, MLA, Chicago, Vancouver), footnote markers, and reference list entries are excluded from the AI-likelihood score.
• Block-quoted passages. Passages longer than 40 words in block-quote form are scored separately. The body text's score is the primary output.
• Section-level scoring. We attempt to detect section headings (Abstract, Introduction, Methods, Results, Discussion, Conclusion) and report per-section scores. The discussion section is usually the most AI-prone in cases we've reviewed.
• Methodological register. The classifier is calibrated so that passive-voice-heavy methods sections don't score as AI by default.

Known failure modes for research papers

• Non-native English authors. This is a well-documented failure mode of all AI detectors. We publish stratified false-positive rates on /transparency. A high score alone is not evidence against a non-native author.
• Highly technical methods text. Template-style methods descriptions ("Samples were collected…", "Data were analyzed using…") read as AI-like to many detectors because they are standardized.
• Heavily workshopped papers. A paper that's been through multiple rounds of revision with co-authors often has voice-smoothing that reads higher than a single-author draft.

How to interpret the result

A research paper with a moderate overall score and one section (typically discussion or introduction) scoring high is worth a conversation. A paper with a uniformly moderate score and no clear hot-spot is likely the genre, not AI. For publication integrity work, pair this detector with COPE guidelines and an authorship-contribution declaration, no detector should be the sole evidence for a misconduct finding.