AI Assistant or Reputation Killer: How to Write Biotech Papers Without the Risk of Retraction

Let's start with a number that should give every biotech researcher pause: by 2025, over 2,100 retractions had been attributed specifically to AI-generated content, with retraction notices citing issues such as "tortured phrases," "nonstandard phrases," and text "generated by a large language model." These were not cases of researchers accidentally submitting a first draft. They were cases where AI-generated text made it through peer review, into the published record, and then — expensively and publicly — back out of it.

At the same time, here is another number worth considering: a 2025 survey by Elsevier of more than 3,200 active researchers across 113 countries found that 58% were already using AI tools for work — up from 37% just one year earlier. The tools are not going away. Researchers are not going to stop using them. The question is not whether to use AI in scientific writing. The question is how to use it without building your reputation on a foundation that a journal editor can dismantle in an afternoon.

This article is a practical guide to exactly that question.

Why the Risk Is Real and Specific

Before the checklist, it is worth understanding precisely what is going wrong — because the failures are specific, not generic.

A key issue is the questionable originality of AI-generated text and the fact that AI cannot meet authorship criteria requiring accountability for a manuscript's integrity. This is not a philosophical point about creativity. It is a legal and institutional one. Every major editorial body — ICMJE, COPE, WAME — now makes clear that accountability for a published paper rests with the human authors, and that AI tools cannot share or absorb any of that accountability. When something goes wrong with an AI-assisted paper, the retraction notice names the researcher. The AI tool is not contactable.

The specific failure modes that lead to retraction fall into four categories:

1. Fabricated citations

   AI language models hallucinate references with impressive confidence. A tool asked to support a claim about CRISPR efficacy in oncology will produce citations that look real, have correct formatting, and may even match genuine author names — but point to papers that do not exist. The 2025 ICMJE recommendations explicitly warn that AI systems can produce fabricated citations and incorrect facts, and journals are increasingly using automated reference verification tools that will catch these before — or after — publication.

2. Undisclosed AI use

   Only 24% of top publishers have explicit generative AI guidelines, compared to 87% of top journals. This asymmetry means that researchers submitting to journals with clear disclosure requirements who fail to declare AI assistance are not merely bending a rule. They are creating documented grounds for retraction. Hidden AI use is now explicitly discouraged by ICMJE and may be treated as a form of misrepresentation — which is a category of misconduct that appears on retraction notices and follows a researcher's name in perpetuity.

3. Data confidentiality breaches

   Uploading unpublished experimental data, proprietary sequences, patient data, or pre-submission findings into a commercial AI platform is not merely an ethical concern — it is a confidentiality violation that some institutions and funders treat as grounds for termination of research agreements. ICMJE has specifically recommended that reviewers not upload manuscripts to AI platforms that cannot guarantee confidentiality — the same principle applies to authors uploading their own unpublished data.

4. "Tortured phrases" and detectable AI text

   Post-publication integrity tools — including those used by journals during and after peer review — are becoming increasingly effective at identifying AI-generated text through linguistic pattern analysis. Papers containing the characteristic hedging language, structural repetition, and semantic smoothness of unedited AI output are being flagged both before and after acceptance. One journal noted a large increase in review paper submissions from mid-2023, with referees observing inadequate reference lists, generic language, and overt scientific errors in papers that subsequently passed peer review and were later suspected of being largely AI-generated.

What the Current Rules Actually Say

The regulatory landscape for AI use in scientific writing has moved quickly and is now relatively clear at the top journal level, even if inconsistently enforced across the broader publishing ecosystem.

ICMJE's 2025 recommendations require authors to explicitly disclose any significant use of AI tools in manuscript preparation — including drafting or rewriting sections of text, translating content between languages, generating or editing images and diagrams, and summarising literature or proposing analyses. This disclosure belongs in the Methods or Acknowledgements section and should specify which tools were used and for what purpose.

The January 2026 ICMJE revision strengthens this further, requiring medical writers to actively disclose AI use, rigorously verify AI-assisted content, and confirm they have complete access to and responsibility for the study data.

COPE does not ban the use of AI but requires transparent disclosure so that editors and reviewers can recognise and assess the influence of AI usage, and warns that AI use can pose risks to research integrity including data fabrication, copyright infringement, and ethical violations.

Across all commonly used ethical and editorial recommendations — ICMJE, WAME, COPE, STM, and EASE guidelines — the position is consistent: AI cannot be credited as a co-author of a manuscript.

Where AI Actually Helps (Used Correctly)

Used within these constraints, AI tools offer genuine value at several specific points in the manuscript preparation process:

• Literature search and synthesis

  AI tools are effective at helping researchers identify relevant papers across large bodies of literature — but every citation must be independently verified against the actual source before inclusion. Use AI to find; use your own judgement to assess.

• Grammar, clarity, and language polishing

  For researchers writing in a second language, AI tools that improve sentence clarity and grammatical correctness are a legitimate and significant aid — provided the intellectual content remains entirely the author's own. This use should be disclosed but is widely accepted.

• Structural drafting

  AI can help generate a first structural outline — section headers, logical flow, argument sequencing — that the researcher then populates entirely with their own content and analysis. The scaffold, not the building.

• Abstract drafting

  An AI-generated first draft of an abstract, based entirely on content provided by the author, can be a useful starting point — subject to thorough revision, factual verification, and disclosure.

• Figure and visual design

  AI tools that assist with data visualisation and figure formatting are increasingly used and accepted, provided the underlying data is not uploaded to unsecured platforms and the AI assistance is disclosed in the Methods section.

The Bottom Line

AI is neither the villain nor the saviour of scientific publishing. It is a tool — one that can make the manuscript preparation process faster, clearer, and more accessible, particularly for researchers operating outside their first language or across very large bodies of literature. It is also a tool that, used carelessly, generates fabricated citations with perfect confidence, produces text that journal integrity systems are increasingly trained to detect, and creates disclosure obligations that, if ignored, become retraction grounds.

The researchers whose AI use ends careers are not, in general, fraudsters. They are people who moved too fast, trusted the output too completely, and did not understand that the accountability for a published paper rests entirely with the human name at the top of it — regardless of how many words below it were drafted by a machine.

The checklist above will not write your paper. But it will help ensure that the paper you write, with whatever tools you use, is one you can stand behind when someone looks at it closely. In the current publishing environment, someone always will.