期刊如何识别 AI 代写论文？审稿人关注的 7 个危险信号

AI slop is not simply research that used AI.

It is what happens when AI-generated material enters a paper without enough reading, checking, subject knowledge or author judgment. The sentences may sound academic. The citations may look real. The document may even be neatly formatted. But beneath the surface, the paper is generic, unreliable or no longer clearly owned by the researcher.

This guide follows Dr. Lyndon Walker’s SciSpace webinar, “AI in Academic Writing: Red Flags Journal Reviewers Notice.” Walker approaches the subject as a journal reviewer, grant reviewer, statistician and researcher with more than 25 years of experience across higher education and research. He is currently a Senior Education Researcher at the Royal Australian College of General Practitioners.

His position is not that AI-assisted research is automatically bad. It is that AI should help researchers produce better work, not simply more work.

The difference comes down to whether the researcher still controls the question, checks the evidence, understands the analysis, protects their voice and accepts responsibility for the final paper.

7 red flags at a glance

1. The research looks mass-produced rather than researched

More papers do not automatically mean more contribution.

Walker begins with an example of a researcher associated with 257 papers in 9 months. His concern is not productivity by itself. It is the increasing volume of papers that appear repetitive, low-effort and disconnected from a credible programme of research.

He says reviewers, editors, grant panels and hiring managers are seeing more material while its average quality is falling. A long publication list can therefore create doubt rather than authority when the papers look formulaic or appear in low-quality outlets.

A reviewer may not know how many papers you produced that year. They can still recognise the same underlying problem inside one manuscript:

• The research question is vague
• The contribution is difficult to state
• The introduction repeats familiar background without identifying a real problem
• The method appears to have been chosen because software could run it
• The discussion restates generic implications
• The conclusion could fit dozens of unrelated studies

The test is not whether the paper contains enough words, tables or citations. It is whether it adds something specific and defensible.

Check the paper’s value

Before submitting, complete these three lines:

1. The unresolved problem is: [specific uncertainty or limitation]
2. This study addresses it by: [design, population, comparison or evidence]
3. The contribution is: [what can now be understood, tested or done differently]

If the answers remain broad, the manuscript may be adding volume without adding enough value.

2. The manuscript contains obvious copy-paste traces

The easiest AI problems to catch are often the most damaging to the author’s credibility.

Walker describes searching Google Scholar for URLs containing source=chatgpt.com and finding thousands of results. He has also received manuscripts containing the same tracking information inside their references.

A tracking parameter is not, by itself, proof that a claim is wrong. But leaving it in a submitted manuscript tells the reviewer that the authors may not have checked what they copied.

He also shows published examples where authors accidentally retained model-generated text explaining that the system could not complete a task. The passage had apparently passed through the authors, reviewers and editors without anyone removing it.

Search your manuscript for these traces

Before submission, search the full document—including footnotes, references, tables and supplementary files—for:

• source=chatgpt
• utm_source
• “As an AI language model”
• “I cannot complete”
• “Insert citation”
• “Add reference here”
• “Rewrite this section”
• Prompt instructions
• Markdown symbols or headings
• Unresolved comments
• Placeholder text
• Broken hyperlinks
• Abrupt changes in spelling, tone or terminology

Do not stop with a keyword search. Read the final exported document from beginning to end.

A manuscript can change during conversion to PDF, reference formatting or collaboration. Check the exact file the journal will receive, not only the working draft.

3. The references look real but are not

AI-generated references are becoming harder to identify by sight.

Walker shows an example in which four references contained incorrect details that did not resemble ordinary typing errors. In one case, the author list included the name of the model Kate Moss in a paper from Nature Geoscience.

The reference still looked structurally plausible. It had a title, journal, year and author list. The problem only became clear when someone checked it against the real publication record.

AI can get any part of a reference wrong:

• Paper title
• Author names
• Author order
• Journal
• Publication year
• Volume or issue
• DOI
• URL
• The connection between the reference and the claim

A DOI resolving successfully is not enough. The DOI may lead to a real paper that is completely different from the one named in the reference.

Google Scholar does not make a citation automatically reliable

Walker describes a case in which Google Scholar appeared to recognise a paper that did not exist because many other documents had cited the fabricated reference. The false citation had begun to reinforce itself through the scholarly record.

That means a result appearing in Google Scholar is useful evidence to investigate, but not final proof that the bibliographic record is correct.

Verify every reference as one record

For each citation, confirm:

Walker’s practical rule is simple: return to the original source. If the source cannot be opened or traced, treat the citation as unresolved rather than assuming it is correct.

A real source may still fail to support the claim attached to it. For that second check, use the SciSpace Citation Verifier, which separates reference verification from claim verification.

For the full citation-to-evidence workflow, see AI Hallucinations in Research: How to Verify Citations, Claims and Sources. That guide covers citation records, full-text retrieval, supporting passages, evidence comparison and final reference auditing.

4. The writing is fluent but says very little

Generic writing is harder to catch than a fake DOI because every sentence may be grammatically correct.

Walker says that large language models often produce vague, bland prose. The wording is not necessarily incorrect. The problem is that it could appear in almost any article, grant or research proposal.

When reviewers read many submissions together, the difference becomes more visible. One paper sounds like it was written by someone who understands the field. Another contains polished sentences but little precision, evidence or disciplinary meaning.

Generic wording often hides one of these gaps

• The population is not named
• The measured outcome is replaced by a broader concept
• The sentence does not identify the study design
• The method is described as “robust” without explaining why
• The finding is called “important” without establishing its size or relevance
• The paragraph lists benefits without evidence
• The implication is not connected to the result
• Qualifications and uncertainty disappear
• Every paragraph follows the same sentence pattern

Test each paragraph for specificity

Ask:

1. What exact point does this paragraph establish?
2. Which evidence supports it?
3. Does it name the relevant population, method or outcome?
4. Could the same wording appear in a paper from another discipline?
5. Does the paragraph explain why the finding matters?
6. Is the researcher’s own interpretation visible?
7. Are uncertainty and limitations preserved?

Consider the difference:

Generic

These findings highlight the transformative potential of innovative approaches to improve outcomes across diverse contexts.

Specific

In this sample of first-year nursing students, weekly retrieval-practice quizzes were associated with higher final-exam scores than optional revision worksheets. The study did not assess whether the effect persisted in later courses.

The second version says what was studied, what was found and what remains unknown.

Protect the author’s voice

Walker argues that researchers, especially doctoral researchers, should be more specific than a general-purpose model because they are developing subject expertise. He recommends working from your own notes and outline, keeping AI away from the main intellectual work and reading the manuscript aloud to catch clumsy or unnatural phrasing.

For a researcher-led approach to drafting and editing, see How to Write a Research Paper Step by Step with Prof. Iain Jackson. The guide recommends revising the argument before polishing the language and rejecting edits that change the strength of a claim or erase the author’s voice.

5. The claims do not match the data

A sentence can sound reasonable and still be unsupported by the study.

Walker identifies this as part of the subtler problem with generic AI writing: a model may produce statements that fit the topic but do not fit the actual data.

The resulting paragraph may be coherent. The error becomes visible only when someone compares it with the Methods, Results, tables or cited source.

Common claim–data mismatches

Run four consistency checks

Check the sample

• Do the participant numbers agree across the abstract, Methods, Results and tables?
• Are exclusions explained?
• Do percentages have the correct denominators?

Check the outcome

• Is the conclusion about the outcome that was actually measured?
• Was the result primary, secondary or exploratory?

Check the design

• Does the wording distinguish association, prediction and causation?
• Does the design support the claim being made?

Check the uncertainty

• Are effect sizes and confidence intervals reported?
• Are null, conflicting or incomplete findings visible?
• Have important limitations disappeared from the summary?

6. The manuscript tries to manipulate an AI reviewer

Hidden instructions designed to influence a review system are an integrity problem, not a clever shortcut.

Walker discusses prompt injection: adding concealed text intended to make an AI system approve, praise or rank a document more favourably.

He says the tactic is both unethical and unlikely to work reliably. Systems increasingly detect it, and the possible reputational harm is far greater than any unlikely benefit.

The same principle applies when hidden instructions are placed in:

• Manuscripts
• CVs
• Grant applications
• Student assignments
• Job applications
• Web pages intended for machine review

Do not add hidden text, white-on-white instructions, manipulated metadata or other content designed to change an automated assessment.

7. The researcher has given away intellectual ownership

The deepest problem is not that AI helped. It is that the researcher can no longer explain what the paper says or why.

Walker draws a line between using a tool to improve communication and asking it to produce the entire intellectual product.

He considers limited help with phrasing reasonable when the research question, data, interpretation and final decisions still belong to the author. He becomes concerned when the model writes the paper, makes the interpretation or replaces the author’s judgment.

Intellectual ownership means you can explain:

• Why the research question matters
• How the method answers it
• Why each analytical decision was made
• What the results establish
• What the results do not establish
• Why each source was included
• How AI affected the workflow
• What was checked by a human
• Why the final argument is yours

The required level of human work also depends on the task.

Walker distinguishes between a rapid internal review, a journal article and a PhD. A quick evidence briefing may use more AI assistance if the process is disclosed and the output is checked. A journal article requires deeper rewriting and conceptual precision. A thesis requires even greater intellectual ownership because the work is meant to demonstrate the researcher’s capability.

Check policy, privacy and permission

Before using AI, ask:

• Does my institution permit this use?
• Does the target journal permit it?
• Am I allowed to upload this manuscript or dataset?
• Does it contain confidential or identifiable material?
• Does my ethics approval allow external processing?
• Am I sharing intellectual property that belongs to someone else?
• Do I need to disclose the tool and task?

Walker warns that reviewers should not upload confidential manuscripts to AI systems without permission. A document received for peer review was provided under confidentiality, and sharing it with an external system may violate that obligation.

For disclosure wording and placement, see How to Disclose AI Use in Research Papers: Examples for Every Stage.

Why an AI-detector score does not solve the problem

A detector cannot tell you whether the research is good.

Walker’s position is direct: AI detectors are unreliable and can produce false positives. Human readers may recognise careless AI use, but they cannot consistently distinguish well-edited AI-assisted writing from human writing.

He therefore recommends returning to the question that matters:

Is this work good enough?

That requires checking:

• The contribution
• The source record
• The evidence behind each claim
• The method
• The analysis
• The interpretation
• The author’s ownership
• Compliance with the relevant policy

A detector score cannot complete those checks.

A high score should not be treated as proof of misconduct. A low score should not be treated as proof that the manuscript is accurate, original or responsibly produced.

The goal is not to “beat” a detector. It is to submit work you can explain, verify and defend.

Replace detection with a reviewer-style quality audit

Walker demonstrates a more useful workflow near the end of the webinar.

He uploads a published manuscript to SciSpace and asks it to check the references, DOIs, journal formatting, statements and whether the paper’s claims have citations. SciSpace returns a reference-verification table, cross-checks the records and flags claims that need supporting evidence.

Walker describes using the result as his own “personal peer reviewer.”

Run the audit in this order

1. Check the document

• Remove placeholders and prompt traces
• Standardise terminology
• Resolve broken links
• Read the final exported file

2. Check the sources

• Confirm that every reference exists
• Verify titles, authors and DOIs
• Retrieve the original source
• Check whether it supports the claim

3. Check the argument

• State the problem and contribution clearly
• Remove generic implications
• Connect each major claim to evidence
• Preserve uncertainty

4. Check the results

• Match sample sizes and percentages
• Compare text with tables and figures
• Distinguish statistical from practical importance
• Keep conclusions within the study design

5. Check ownership and policy

• Confirm what AI did
• Record what the authors checked
• Protect confidential material
• Add the required disclosure
• Ensure every author approves the final paper

Use AI without producing AI slop

Walker’s webinar is not an argument against using AI.

He demonstrates using SciSpace to:

• Search the literature
• Follow backward and forward citations
• Open the original source
• Read and question a paper
• Explain technical passages, equations and graphs
• Compare papers in a structured table
• Check references and DOIs
• Review claims in a manuscript
• Convert a paper from one citation style to another
• Track edits before accepting them

The safe division of work is:

The researcher remains accountable even when the tool performs part of the work. Walker demonstrates this after using SciSpace to convert a manuscript from APA to Vancouver style: the changes were tracked, but he still had to inspect and approve them.

The standard is not “no AI.”

It is:

Use AI for speed and support. Keep the expertise, verification and final judgment human.

The reviewer perspective behind this guide

Dr. Lyndon Walker

Dr. Lyndon Walker is a PhD-qualified statistician and Senior Education Researcher at the Royal Australian College of General Practitioners. His work spans statistics, data science, medical education and research methodology. He also reviews journal manuscripts and grant applications across multiple fields.

In the webinar behind this guide, he moves from obvious signs of careless AI use to subtler problems involving references, generic prose, unsupported claims, integrity and intellectual ownership.

His central point is not that reviewers need a perfect way to identify AI-generated text.

It is that researchers need to produce work that is accurate, specific, transparent and worth reading—regardless of which tools assisted them.

Frequently asked questions

注意转载授权 · 阅读原文