Artificial Intelligence in Academic Writing: The ICMJE Limits

Author or assistant

The role of artificial intelligence in an academic paper begins with this question. The authorship criteria of the ICMJE (International Committee of Medical Journal Editors) are four, and all of them must be met together:

1. Substantial contribution to the conception, design, or data collection/analysis/interpretation of the study
2. Contribution to the writing or critical revision of the paper
3. Approval of the final version to be published
4. Being responsible for the accuracy and integrity of the study, and being able to answer questions

An LLM cannot meet the fourth criterion by definition. A model cannot say "I am responsible for this sentence," cannot be a witness in court, and cannot participate in a retraction process. For this reason, ICMJE and the Vancouver bibliography prohibited listing artificial intelligence as an author (December 2023, updated 2024). This decision was adopted by Nature, Science, JAMA, NEJM, Lancet, Cell, and BMJ in the same period.

The correct category: artificial intelligence is a tool. Whatever Microsoft Word, SPSS, R, and EndNote are, GPT-4 or Claude are the same. The use of the tool is declared, but it is not listed as an author.

Which use is acceptable, which is a red line

Dividing the use of academic artificial intelligence into four levels provides a practical framework:

Level 1: Language and format (green light)

• Translation of a Turkish draft into English
• English language editing, revision of academic tone
• Citation format conversion (Vancouver to APA)
• Revision of table and figure captions

This use is in the same category as Word's spell-check or Grammarly. Most journals ask for a one-sentence declaration even at this level, but it does not touch content integrity.

Level 2: Structural help (yellow light)

• Creating a section draft (for example, "a 5-paragraph skeleton for the discussion")
• Summarizing literature (from papers the author has read)
• Generating counterarguments for the discussion
• Writing code (R, Python statistics script)

At this level, every sentence of the output must be verified by the author, cited, and rewritten if necessary. The "hallucination" risk of an LLM is high at this level; especially in citation generation, artificial intelligence may fabricate papers that do not exist. A detailed declaration in Methods or Acknowledgements is mandatory.

Level 3: Data analysis (orange light)

• Numerical analysis (artificial intelligence generating and running code)
• Visual generation (figure, infographic)
• Table automation (raw data to publication-ready table)

In this use, control of the risk becomes harder. The statistical accuracy of the code produced by artificial intelligence is the author's responsibility. At this level, JAMA and Lancet ask for the archiving of "raw output + verified output" and its presentation as supplementary material. Otherwise there is a retraction risk.

Level 4: Content generation (red light)

• Having the entire text written by artificial intelligence and submitting it with minimal correction
• Fabricating findings without data (an imaginary result instead of a literature summary)
• Not checking artificial intelligence's fabrication of references
• Presenting artificial intelligence output as "my own research" in ethics approval processes

This level is academic fraud. In 2024, Frontiers and Springer Nature retracted hundreds of papers in which they detected AI-generated paper mill content. There were usually the following signs: the meaningless phrase "regenerate response," conceptual contradictions with no English language errors, and imaginary citations.

The Vancouver bibliography and artificial intelligence

The Vancouver style (recommended by ICMJE) is a reference format. The question: if I used artificial intelligence, how do I write it in Vancouver?

Artificial intelligence is not cited as a reference. That is, a reference like the following is wrong:

Bad: GPT-4. Discussion on knee osteoarthritis pathophysiology. 2026.

Artificial intelligence is written as a method. An example in the Methods section:

We used OpenAI GPT-4 (version gpt-4-turbo-2024-04-09) to assist
with English language editing and to generate an initial outline
of the Discussion section. All AI-generated text was reviewed,
revised, and verified by the authors. References were manually
verified against PubMed; no AI-generated citations were used.

A shorter variant in the Acknowledgements section:

Acknowledgement: The authors acknowledge the use of OpenAI GPT-4
for English language editing. No content was generated by AI
beyond grammatical revision.

These two templates are the format accepted in most journal guidelines.

Comparison of journal policies (2025-2026)

Examples of active policies:

• Nature: Model and use detail mandatory in the Methods section. Requires a separate paragraph for visual generation. Main visuals produced with generative AI are not accepted.
• JAMA: Writing assistant accepted, full transparency + code sharing for data analysis. Scanning is done for AI content detection.
• NEJM: Language and format use with free declaration. AI use in generating clinical decision support or diagnostic suggestions is rejected (clinical responsibility must be on the physician).
• Lancet: Applies the ICMJE framework exactly. If AI is listed as an author, the manuscript is desk-rejected.
• Cell: Detailed declaration in Methods or Supplementary. AI-assistance disclosure to the reviewer is requested.
• Frontiers, MDPI, Springer Nature: An "AI usage" check-box added to the author declarations form. If left blank, the manuscript is returned.
• PLOS: The official policy is compliant with ICMJE; additionally, with the open data requirement, the AI usage code must also be deposited.

For most of the TR Dizin and ULAKBIM journals in Turkey there is not yet an official policy. In this case Vancouver can be used as a practical standard, and it is declared in the Methods.

Author disclosure and the ethics committee

Should artificial intelligence use be stated in an ethics committee application? The answer depends on the type of study.

In clinical research (TUBITAK 1001, clinical trial protocols) it should be stated that AI will be used in data analysis. Because the ethics committee asks for a pre-specified analysis under "data analysis method." The analysis steps of the AI use, which model, and which parameter values are written down.

In a systematic review protocol (PROSPERO registration) AI use is stated at the following steps:

• For screening: which LLM, which prompt, how much is manually checked
• For data extraction: which model, dual extractor (human + AI) or AI primary + human check
• For risk of bias: whether the AI suggestion is final with human approval

Cochrane's 2025 methodological update says that for AI-assisted SR a dual-extractor (at least one human) and 100% manual verification of the LLM output are required. AI does not perform extraction on its own.

Preparation for reviewer scrutiny

While reading the paper, the reviewer looks for signs of AI content. Three things to be prepared:

1. Run the AI detection tools: GPTZero, Originality.ai, Turnitin AI detector. If you used artificial intelligence beyond language editing, these tools give a high score. If the score is high: either strengthen the disclosure or rewrite the text more aggressively.
2. Verify citations manually: Check every reference through PubMed and Google Scholar. Artificial intelligence hallucination occurs most at the citation level.
3. Share code and data: Share your AI prompts, the raw output, and the manually edited version as supplementary material. The reviewer likes transparency; the road to retraction starts with opacity.

Practical core checklist

The minimum to be done for a complete manuscript in AI-assisted academic writing:

• Which model, which version, which date
• In which sections it was used (Introduction / Methods / Results / Discussion)
• Data analysis, language, or structural
• Whether it was declared in the Methods section
• Whether it was declared in Acknowledgements (if the journal requires)
• Whether citations were manually verified
• Whether the AI output was revised by a human
• Code and prompts in supplementary
• All authors read and approved the final text
• AI use recorded in the journal disclosure form

Three common mistakes

Mistake 1: The "nobody will notice" assumption. GPT-4 has a stylistic signature (long sentences, a passion for three-item lists, phrases like "in conclusion"). Reviewers and the editor recognize these traces. Week by week, journals share a common retraction list.

Mistake 2: Making AI fabricate references. When you say "suggest 10 references on this topic," the model produces good-looking but nonexistent citations. The reviewer catches it with a single click, and the manuscript goes into desk-reject or editorial misconduct.

Mistake 3: Listing as an author. If the editor sees "ChatGPT" in the author list, they reject the submission. In 2023 this happened frequently; in 2026 there is now automatic scanning.

Serteser Danismanlik support for academic artificial intelligence integration

For the use of academic artificial intelligence to be ethical, efficient, and journal-compliant, an outside eye may be valuable. Serteser Danismanlik for academics:

• ICMJE/Vancouver compliant AI declaration templates
• Disclosure guide by target journal
• AI-assisted screening protocol in a systematic review (PROSPERO compliant)
• Design of an AI vs manual hybrid pipeline in data analysis
• Preparation of the AI methodology section in an ethics committee application
• Defense of AI usage against reviewer comments

In a first 15-minute free introductory meeting, we evaluate the risk level of your use scenario and produce a journal-specific roadmap. With clinical research experience published in an international peer-reviewed journal and active management of PROSPERO-registered systematic reviews, we support you in preserving both practical and ethical standards at the same time in academic artificial intelligence use.

For support in the design of your first artificial intelligence study, you can review the study design service within the scope of academic consulting.