An AI Workflow Guide for Academics

Academic AI use is still a missing discipline

Most academics have discovered artificial intelligence but do not know how to integrate it. Two extremes are observed. On one side, ad hoc use: asking ChatGPT a question and copying the answer that comes back, sending it off without catching the hallucinations and fake references. On the other side, moral panic: rejecting AI entirely, viewing all forms of use as "forbidden" in the name of academic ethics.

The right approach lies in the middle. When used at the right points, the AI tool does not lower research quality; on the contrary, it adds a layer of time and error management. When used at the wrong points, it can create risks that go as far as paper retraction.

This article covers where AI enters and where it should not enter across the five stages of the academic research process, which tool is most productive with which prompt pattern, and where the line is drawn in terms of the ICMJE Vancouver authorship rules.

Stage 1: Literature Searching

Classic knowledge of searching PubMed, Scopus, and Web of Science is still fundamental. AI does not replace this; it adds a lateral tool on top.

Recommended division of labor:

• Traditional search (PubMed, Scopus): Boolean operators, MeSH terms, systematic. Replicable. Mandatory for systematic reviews.
• Perplexity, Elicit, SciSpace: Semantic search, paper summarization, quick scanning of "what does this article show." For ad hoc pre-screening.
• Connected Papers, ResearchRabbit: Visual citation network, branching out to related papers. Rapidly mapping the literature from a seed paper.

The right usage pattern:

1. First, run a replicable search in PubMed/Scopus and save it.
2. Send the results to Zotero or Mendeley.
3. Give the interesting but uncertain paper to Perplexity or SciSpace, get a summary, and make your decision.
4. Use Connected Papers to look around and check whether there is a classic reference you missed.

Do not:

• Do not tell ChatGPT to "list the 10 most important papers on X." The hallucination risk is high and it produces fake DOIs. One of the cases that still occurs: in 2025 Claude 3.5 Sonnet produced a combination of a non-original paper title plus fake authors, an academic put this into the article, and it was caught during peer review.
• Do not have AI perform the literature search. A non-replicable search is unacceptable for a systematic review.

Stage 2: Critical Reading and Synthesis

At this stage AI genuinely produces value. Models with long context windows (Claude 3.5 Sonnet, Gemini 1.5 Pro, GPT-4.1) can digest 50 to 100 pages of PDF in a single pass.

Prompt patterns:

I have read this article. Summarize it in the following structure:
1. The study's research question (in PICO format)
2. Method (design, sample, measurement instrument, statistics)
3. Findings (with numerical values, including p-value and effect size)
4. Limitations (those stated by the authors + those you see)
5. This article's contribution to the field (1 sentence)
6. 3 points on which this article could be criticized

For synthesis:

I am giving you 5 articles. They are all RCTs on topic Y.
Produce the following table:
| Study | n | Intervention | Comparison | Primary outcome | Effect size | p |
Then answer these questions:
- Which studies' results contradict each other?
- What might the methodological reason for this contradiction be?
- For the next systematic review, how would you narrow the PICO question?

Benefit: Where a human reads and summarizes a paper in 45 minutes, AI does it in 5 minutes. But the value of human critical reading does not end here. Reviewing the AI's summary and accepting it, correcting the misinterpreted part, and catching the extra fine detail are still the human's job.

Stage 3: Methods and Statistics Discussion

The least controversial use of AI for academics: writing code, discussing statistics, clarifying methods.

R / Python / SPSS assistant:

• "What is the syntax for a mixed effects model in the lme4 package in R? How do you write a random intercept + random slope?"
• "Which test is appropriate for this data set? The normality test results and variance homogeneity are such and such."
• "For a forest plot, with which arguments is the forest() function in the metafor package run?"

Method clarification:

• "What is the conceptual difference between propensity score matching and inverse probability weighting? In which situation would I prefer which?"
• "How are the 95% limits of agreement interpreted in a Bland-Altman analysis? What is the difference between clinical meaning and statistical meaning?"

Important: At this point AI is not an expert, it is a fast dictionary. Using the answer without verifying it is a mistake. Especially with newer methodological approaches (for example, Bayesian network meta-analysis) the answer may be incomplete or wrong. The Cochrane Handbook, classic statistics textbooks, and methodologist consulting are still the verification layer.

Stage 4: Manuscript Draft

Here is the zone that requires care and discipline. Using text written by artificial intelligence as-is is problematic in terms of academic ethics and will be caught during peer review.

Correct use:

• Outline writing: "Based on these findings, what would the outline of a Discussion section look like? Which subheadings?"
• Sentence revision: "Could you write this sentence more fluently? Do not distort the meaning."
• Translation layer: thinking in Turkish and using the AI as a language revision assistant during the process of writing in English.
• Counter-argument generation: "What could a reviewer object to in this Discussion paragraph? Missing points?"

Incorrect use:

• Saying "write me an introduction" and copying the 4 paragraphs that come back. Typical AI language (overgeneralization, vague claims, missing citations) is a red flag for a peer reviewer.
• Using a reference produced by AI without verifying it. It can fabricate DOIs.
• Signing a manuscript entirely drafted by AI as sole author. This violates Item 1 of the ICMJE Vancouver criteria: there is no "substantial contribution to drafting."

ICMJE's view on AI (2024 update):

"AI tools (including LLMs) cannot be listed as authors. AI cannot take responsibility for the work. If AI is used during drafting, this must be disclosed in the Methods or Acknowledgments section."

In other words: AI use is not forbidden, but it must be disclosed. It is not written into the author list. All major journals, including JAMA, Lancet, Nature, and NEJM, have required disclosure since 2024.

Stage 5: Peer Review Preparation

Once the manuscript is ready, using AI as a "hostile reviewer" before submitting is extremely valuable.

Prompt pattern:

You are a strict reviewer. You are reading this manuscript, which will be sent to a given journal (for example: BMC Medical Research Methodology).
Write a critical report under the following headings:
1. Methodological weaknesses that would lead to major revision
2. Statistical analysis choice: is it correct or wrong? Is there an alternative?
3. Is there overstatement in the interpretation of the results?
4. Is the limitations section adequate?
5. Is the reporting standard (for example: PRISMA 2020, STROBE, CONSORT, TRIPOD) fully met?
6. 5 minor points that need to be corrected (sentence, terminology, format).

The report that comes out generally predicts 50 to 70% of the real reviewer report. One or two rounds of self-review before submission seriously reduces the probability of major revision.

Additional ad hoc use:

• Producing a cover letter template (personalization is required).
• Editor response draft (especially after major revision).
• Conflict of interest statement template.

Cross-disciplinary examples

Health sciences: CONSORT/STROBE checklist self-review for clinical study protocols, PRISMA checklist for systematic reviews, TRIPOD-AI 2024 compliance for AI studies.

Social sciences: Survey development (generating an item pool, followed by human revision), qualitative data coding (theme suggestion, followed by human validation), APA style revision.

Engineering: Discussion of algorithm alternatives, code refactoring (again followed by human testing), benchmark table template, IEEE format revision.

Law: Case law synthesis (high risk of fake case law, always verify), counter-argument generation, academic article structure.

Do-not list

• Do not have AI generate sources. The hallucination risk is high.
• Do not use AI output without disclosure. Disclosure is mandatory at major journals.
• Do not add AI to the author list. ICMJE forbids it.
• Do not report a p-value or effect size produced by AI without verifying it.
• Do not use novice models (pre-3.5, small free-tier models) for serious academic work. The error rate is high.

Do list

• Choose a model with a long context window (Claude 3.5 Sonnet+, Gemini 1.5 Pro+, GPT-4.1+).
• Put every AI output through a verification layer.
• Save your prompts, refine them, share them with the team.
• Before submission, run AI as a hostile reviewer.
• Clearly disclose AI use in the manuscript.

A properly integrated AI workflow can increase your annual paper productivity by 30 to 50%. Incorrect integration, on the other hand, can create career damage with a single retraction. The line is use enriched by discipline.

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