A manuscript reads smoothly, its abstract has the right shape and the discussion sounds confident. Is that evidence that a language model wrote it? Not really. Fluent language has never been a reliable test of whether the underlying research is genuine.
As AI writing tools become commonplace, journals face two separate questions: was AI used to prepare the manuscript, and can the evidence reported in that manuscript be trusted? Confusing those questions can lead editors to punish legitimate writing assistance while overlooking fabricated research.
Writing assistance is not the same as invented science
An author may use AI to improve grammar, reorganise a paragraph or explain a complex sentence. Another may ask it to generate an entire discussion section. A third may use it to fabricate references or produce tables without data. These activities have very different integrity implications.
The International Committee of Medical Journal Editors advises that AI use should be disclosed where relevant, that tools should not be listed as authors, and that humans remain responsible for accuracy, attribution and originality. Its guidance reflects a simple principle: the author must be able to defend what has been submitted.
A journal policy should tell authors what is permitted and how to disclose it. Without that clarity, editors are left interpreting behaviour after the fact.
The limitations of AI-text detectors
Tools that estimate whether text was AI-generated can produce false positives and false negatives. They may be influenced by writing style, formulaic scientific language or deliberate rewriting. A high score should not be treated as proof of misconduct.
Consider a researcher whose manuscript has been heavily copyedited into concise English. A detector may find the language unusually predictable. That says little about the validity of the experiment. Conversely, a paper containing fabricated results may have been written entirely by a human.
The better first question for an editor is: what part of the scientific record can we verify? Request underlying information when appropriate, check that citations support claims, compare tables with the stated methods, and examine anomalies with qualified reviewers.
Where the stronger checks belong
Reference verification is a practical starting point. Does each cited paper exist? Do its authors, year and DOI match? Does it support the statement for which it is cited? A list of genuine references can still be misleading if the cited work does not say what the manuscript claims.
Methodological checks matter more. If a study reports a randomised trial, are the allocation, outcomes, registration and analysis described coherently? If it presents images, are the preparation and processing procedures plausible? If a dataset is declared available, can the researchers explain how it was collected and accessed?
No single software product can resolve all of those questions. AI can draw attention to contradictions, but editors need a way to check the underlying evidence and give authors a fair opportunity to respond.
Protect unpublished work while screening
A confidential manuscript should not be pasted into an unapproved AI service. Journals need to know where content is processed, what is retained and which contractual protections apply. ICMJE guidance specifically warns about the confidentiality risks of uploading manuscripts to AI systems without adequate safeguards or appropriate permission.
Staff and reviewers should be given a simple approved workflow rather than told merely to 'use AI responsibly'. The workflow should specify permitted tools, disclosure expectations and who reviews a concerning result.
What publishers should measure
A successful integrity programme should not be judged by the number of 'AI-written' manuscripts flagged. Measure the accuracy of checks, proportion of concerns resolved with evidence, time taken to investigate, false accusations avoided and corrections issued when genuinely needed.
That approach creates a fairer standard for authors and a stronger defence of the scholarly record. The objective is not to keep AI-generated sentences out of science at all costs. It is to keep unsupported claims, hidden manipulation and unreliable research from passing into the literature unnoticed.
