A cancer paper does not become trustworthy because it sounds polished. It becomes trustworthy when its claims can be traced to real data, its sources are credited, its methods can be examined, and named researchers accept responsibility for every important sentence.
That standard is under pressure. Generative AI can produce fluent prose, plausible references, tidy summaries, and even convincing explanations of results it did not generate. Text-reuse software creates a different kind of signal: it can reveal copied passages, recycled language, or suspicious overlap that a busy editor might miss. Neither technology replaces scientific judgment. Both can make that judgment faster and better targeted.
The cancer-publishing problem is bigger than copied sentences
In cancer research, a misleading sentence is not merely a writing flaw. It can distort how clinicians understand a treatment, how investigators design the next experiment, or how patients interpret an early finding. The risks extend beyond word-for-word copying:
- An AI-generated reference may point to a real paper but misstate what it found.
- A polished paragraph may combine findings from different tumor types, endpoints, or patient populations.
- A methods section may reuse language without crediting the source of a protocol or conceal a meaningful difference in how the work was performed.
- A similarity score may be high because of standard technical language, while genuine plagiarism may involve an idea, figure, dataset, or interpretation that software cannot recognize.
The U.S. Office of Research Integrity defines research misconduct as fabrication, falsification, or plagiarism in proposing, performing, reviewing, or reporting research. Its definition of plagiarism includes taking another person’s ideas, processes, results, or words without appropriate credit, not just copying a paragraph. That distinction matters in oncology, where a novel assay, classification scheme, cohort design, or interpretation may be the intellectual contribution.
The scale of the issue is not theoretical. A 2025 study of oncology journals examined 241 titles and found that 209 of the 224 journals whose websites mentioned plagiarism also said they check submissions for it. Crossref Similarity Check was the most frequently named method. The finding shows that screening is now normal editorial infrastructure, not an accusation aimed at a few careless authors.
What the tools can see, and what they cannot
Plagiarism software is good at locating textual overlap. It is not good at deciding whether overlap is unethical. A shared description of a standard staining procedure, a clinical definition, or a statistical method may be unavoidable. Conversely, a heavily paraphrased passage can hide borrowed reasoning while producing a modest similarity score.
COPE therefore advises editors not to use a fixed percentage as a plagiarism threshold. The relevant questions are more specific: Is the source cited? Is the matching text concentrated in a distinctive passage? Is the overlap in the methods, results, or discussion? Are ideas or data being presented as original? A report should point a reviewer toward those questions, not answer them automatically.
AI-content detectors have a similar limitation. A flagged passage may reflect machine-generated text, but it may also reflect formal scientific prose, language editing, or a detector’s uncertainty. COPE’s guidance on suspected AI-generated manuscripts says detection results must be interpreted by experts and validated through other evidence. In other words, an AI detector can be part of a screening process, but it should never be the sole basis for rejecting a paper, accusing an author, or retracting a study.
My view is straightforward: a detector should create a conversation, not a conviction. The more serious the possible consequence, the more important it is to inspect the manuscript, ask the authors how the text was produced, verify the citations, and assess whether the scientific claims remain sound.
Why disclosure is part of the evidence
Disclosure does not automatically make AI use acceptable, and nondisclosure does not automatically prove misconduct. It does, however, give editors a way to assess the tool’s role and the authors’ oversight.
The January 2026 recommendations from the International Committee of Medical Journal Editors say authors should disclose AI-assisted technologies used in producing a manuscript. They also state that AI tools cannot be authors because they cannot be accountable for accuracy, integrity, originality, conflicts of interest, or final approval. Humans remain responsible for checking AI-produced text and images for plagiarism and for giving proper attribution.
That approach is already visible in cancer publishing. The American Cancer Society journal Cancer tells authors that use of AI-generated content in any portion of a manuscript must be described transparently and in detail in the Methods section. The practical message is useful: do not force editors to guess whether AI helped with copyediting, drafted new scientific content, generated code, summarized literature, or analyzed data. Those are different uses with different risks.
A defensible workflow for authors and editors
Tools matter most when they sit inside a repeatable process. For a cancer manuscript, a sensible workflow has five stages:
- Screen for signals. Run a similarity check and, where journal policy permits, an AI-content check before submission or during initial editorial assessment. Preserve the reports rather than treating a score as a conclusion.
- Read the matches. Exclude or interpret expected overlap in references, standard methods, and specialist terminology. Examine blocks of distinctive text, uncited ideas, copied discussion, figures, tables, and reused results.
- Verify the science. Check every important citation against the original paper. Confirm that the cited population, endpoint, intervention, and limitation match the manuscript’s wording. Test whether the results and conclusions are supported by the reported data.
- Ask neutral questions. If a concern remains, ask authors what tools were used, for which tasks, and how they reviewed the output. Give them a chance to explain legitimate text reuse, language assistance, or a shared protocol.
- Escalate proportionately. A correctable attribution problem is not the same as fabricated data or copied results. Serious concerns should move through the journal’s research-integrity procedure and, where appropriate, the authors’ institution.
This workflow protects authors as well as journals. A false AI flag can unfairly stigmatize a researcher, especially someone writing in an additional language. A high similarity score can punish a precise methods description while missing conceptual plagiarism. Human review is not a soft alternative to technology; it is the safeguard that makes the technology usable.
The confidentiality trap reviewers must not ignore
There is a second risk in using automated tools: the manuscript itself may be confidential. ICMJE warns that uploading a submitted paper or peer-review materials to an AI system can breach confidentiality when the system’s data handling cannot be assured. Nature likewise says manuscripts, peer-review reports, and sensitive data must not be shared with unsecured or public AI systems.
That warning applies to detection services too. Before submitting a manuscript to an external checker, an editorial office should know what is retained, whether the text enters a training dataset, who can access it, and whether the journal has permission to process it that way. Reviewers should not upload a confidential paper on their own initiative simply because a tool is convenient.
What responsible use looks like in practice
Authors should keep a record of substantial AI assistance, verify references against the original literature, review every generated sentence for changes in meaning, and disclose use according to the target journal’s instructions. They should never assume that an AI system’s confident wording is evidence, or that a generated citation is real merely because it has a familiar title.
Editors should publish clear policies, train staff to interpret reports, and avoid a single numerical cutoff. They should distinguish language support from autonomous drafting, and textual overlap from plagiarism. They should also document why a concern led to revision, rejection, correction, or referral.
Readers and institutions have a role as well. A paper that survives screening is not guaranteed to be correct. Post-publication questions, data availability, reproducible methods, and willingness to correct the record remain essential. Detection tools are an early warning system, not a substitute for peer review or scientific skepticism.
The standard worth defending
Cancer research deserves more than prose that passes as human or text that happens to produce a low similarity percentage. It deserves a record in which credit is visible, methods are honestly described, conclusions stay within the evidence, and someone accountable has checked the work.
AI detection and plagiarism tools matter because they help editors and researchers find questions that deserve attention. They matter just as much for what they cannot do: they cannot establish intent, evaluate a clinical claim, authenticate data, or replace an expert reading closely. Used as prompts for investigation rather than verdict machines, they strengthen research integrity without turning publishing into a contest between algorithms.