A polished paragraph is not proof that a research paper was generated by AI. Neither is a factual mistake, a vague citation, or a high detector score. Each can justify a closer look, but the conclusion must come from the paper, its sources, the writing process, and the applicable policy.
The seven checks below work for coursework, journal screening, and editorial review. Adjust the process to the authority you actually have. An instructor may be permitted to review drafts that a journal editor or outside reader cannot access.
1. Read the policy before the prose
Start with the rule the writer was expected to follow. Some courses permit AI for brainstorming or grammar feedback but prohibit generated final prose. Some journals require disclosure of any generative assistance. A review cannot be fair if the rule is invented after the paper is submitted.
Write down the exact concern. Undisclosed drafting, fabricated sources, copied language, and poor research are different issues. They may overlap, but each needs different evidence.
2. Separate the writer's prose from other text
Assignment instructions, quotations, tables, reference lists, formulas, and publisher templates can distort an authorship check. Keep an untouched copy of the paper, then prepare a second copy containing the prose you intend to examine.
Short abstracts and captions provide little text. A detector result from them should carry less weight than a result from several connected paragraphs. Do not combine unrelated sections just to produce a larger score.
3. Verify a sample of claims and sources
Choose claims from the introduction, evidence sections, and conclusion. Open the cited source and ask whether it supports the sentence. Check author names, titles, dates, page numbers, and identifiers. Generated text can invent plausible references, but people also mistype citations and misunderstand sources.
Record what is actually wrong. "The cited paper does not exist" is evidence. "The citation looks like AI" is an impression. If a source exists but does not support the claim, treat that as a research problem first.
4. Look for a traceable argument
A sound paper should show how evidence leads to the conclusion. Mark places where the argument jumps, changes its definition, introduces a new claim without support, or repeats a general point instead of analyzing the evidence. These weaknesses deserve revision regardless of who drafted the sentence.
Abrupt changes in vocabulary or tone can identify a section worth discussing. They do not identify the tool or person responsible. Topic difficulty, editing help, fatigue, and collaboration can also change a writer's voice.
5. Use an AI detector as a locator
Submit the prepared prose to the AI detector only if policy and privacy terms allow it. Save the date, input, result, and passages that were highlighted. Read the explanation of the score. A document likelihood is not necessarily the percentage of words written by AI.
Do not keep trying detectors until one supports a preferred answer. Different systems can disagree because they use different training data, features, and thresholds. Our guide to detector disagreement explains what that variation means.
6. Check the writing process
Where the rules permit it, review outlines, reading notes, source annotations, drafts, version history, lab records, or feedback. Look for a plausible development of the research question and argument. Process evidence can explain a detector flag more directly than another automated score.
Missing drafts are not proof if drafts were never required. A clean version history is also not automatic proof of authorship because text can be pasted or edited elsewhere. Evaluate the record as a whole.
7. Ask specific, neutral questions
Give the writer an opportunity to explain the paper. Ask why a source was selected, how a disputed paragraph supports the thesis, what changed between drafts, or how a calculation was produced. Questions tied to the submitted work are more useful than a surprise memory test.
Document the response and follow the established decision process. Stress, disability, language background, and the time since writing can affect an oral explanation. A serious finding should rely on multiple pieces of relevant evidence and include a path to review or appeal.
What this process avoids
- accusing a writer because the prose is formal or unusually smooth;
- confusing source overlap with generated authorship;
- treating a percentage as the probability that a person cheated;
- uploading confidential research to an unapproved service;
- penalizing a writer for assistance that the policy allowed.
The aim is not to make every case certain. It is to make the review explainable, proportional, and open to evidence that challenges the first suspicion.