Artificial intelligence can help at nearly every stage of scientific research — from shaping an idea to the final edit of the text. But the scientific community has set clear boundaries here: AI can be a powerful assistant, but it can never be an author or a responsible party. In this guide, we look at how to use AI correctly in research and which lines must not be crossed.
Step 1: literature search and analysis
The most time-consuming part of research is studying the existing literature. AI plays three roles here: search, summarization, and comparison.
Ready-made prompt (for a literature map):
I am researching "water resources management in Central Asia". Break the topic into 5 sub-areas, and for each sub-area provide:
1) Key terms (in English and Russian), 2) 3 academic databases to search (and how to phrase the query in Scopus, Web of Science, Google Scholar),
3) A list of the main questions expected for each sub-area.
Abstracts of found articles can be uploaded to AI for summarization:
Read the abstracts of the 8 articles below and build a table with columns: author/year,
method, main conclusion, relevance to my topic (high/medium/low).
Finally: are there contradictory conclusions among these articles — and if so, which ones?
[Abstracts text]
An important limitation: AI can "read" the full text of articles and draw conclusions, but it can also invent sources — citing a nonexistent article ("hallucination") is the most dangerous mistake in scientific writing. Verify every reference in the original source.
Step 2: research design and methodology
AI acts as an "interlocutor" in refining the research question and choosing a method.
Ready-made prompt:
My research question: "The impact of remote work on employee productivity in Tashkent IT companies." Break this question into 3 measurable sub-questions.
For each, suggest a suitable method (survey/interview/observation) and justify the sample size. Critically point out 2 weak points in the methodology.
The last sentence matters — asking AI to criticize its own proposal prevents blind trust. AI also helps in drafting survey questions, but testing questions on real respondents (a pilot) is an irreplaceable step.
Step 3: data analysis
In statistical analysis, AI helps write code (Python/R) and interpret results — but responsibility for the correctness of the analysis lies entirely with the researcher.
Ready-made prompt:
Write code in Python for the following analysis: survey data from 200 respondents (CSV: age, work_experience, remote_days, productivity_score).
1) Data cleaning and descriptive statistics, 2) correlation between remote_days and
productivity_score, 3) a 3-sentence summary interpreting the result. Comment each step in the code.
The golden rule: do not use AI-written analysis code without understanding it. You must know what each line does — "AI did it this way" is not an acceptable answer at a defense or in peer review.
Step 4: scientific writing — the permitted and forbidden boundaries
This is the most delicate topic. The position of major publishers and ethics committees is as follows:
Permitted (usually no disclosure required): fixing grammar and spelling, improving style, translation — i.e., polishing the text you wrote.
Permitted (disclosure required): help in shaping ideas, generating parts of the text, literature search — these must be indicated in the "Methods" or "Acknowledgements" section of the article. For example, Springer Nature journals require documenting LLM use in the Methods section.
Forbidden: listing AI as an author (COPE and ICMJE are unanimous: AI cannot take responsibility, so it cannot be an author); uploading a confidential manuscript to AI during peer review; creating fake data or sources.
Ready-made prompt (for legitimate editing):
Edit the English of the following academic text: improve grammar, academic style,
and logical coherence. Do not change scientific terms or meaning.
Provide a separate list of the places you changed.
[Text]
"Responsibility for every sentence written with AI remains with the author. A journal editor will not accept 'AI wrote it' as an excuse — the mistake counts as yours. So check AI text even more strictly than your own."
Step 5: ethics checklist — before submission
Before sending your manuscript to a journal, answer the following questions:
- Have I disclosed? Have I read the target journal's AI policy and indicated the use in the required section?
- Is authorship correct? Does the list include only real contributors? Is AI absent from the author list?
- Are the sources real? Has every reference been verified in the original source? Is there no AI-"invented" source?
- Is the data protected? Have participants' personal data not been uploaded to AI services?
- Is review confidentiality intact? When reviewing others' manuscripts, did I refrain from uploading them to AI? (ICMJE considers this a breach of confidentiality.)
- Are the facts verified? Has every AI-generated claim been confirmed by an independent source?
An additional point for the Uzbekistan context: in dissertations meeting OAK (Higher Attestation Commission) requirements, the use of AI is not yet fully regulated — agree on it with your academic advisor and the council in advance.
Which tool for which stage: an AI map for researchers
Beyond general chatbots (ChatGPT, Claude, Gemini), there are specialized tools for specific research stages:
Literature search: Elicit and Consensus — find evidence from scholarly articles for your question; Scite.ai — shows how an article was later cited (supported or criticized); ResearchRabbit and Connected Papers — draw a "map" of related works starting from one article.
Data analysis: ChatGPT's code-writing ability (Python/R) or Claude's large-table analysis function — for statistical calculations and visualization.
Writing and editing: Grammarly or Paperpal — for academic English; DeepL Write — for rephrasing sentences.
Plagiarism and originality: Turnitin, iThenticate — standard tools for checking text before submission.
An important warning: these tools can also make mistakes. For example, literature search tools sometimes "find" a nonexistent article. Checking every find in the original database (Scopus, Web of Science, Google Scholar) is an irreplaceable habit.
Additional advice for Uzbek researchers: if access to international databases is limited, use the access provided through your university library — many universities have institutional subscriptions to Scopus and Springer.
Grant applications and academic presentations
Beyond the article, a researcher has two more important written works: grant applications and conference presentations.
For a grant application: AI is strong at checking the structure of an application. Ready-made prompt:
Evaluate the following grant application draft (criteria: clarity of the problem statement,
soundness of the methodology, realism of the budget, measurability of expected results).
Score each criterion 1–5 and concretely suggest how to improve the 2 weakest areas.
[Application text]
For a presentation: turning a 15-minute talk text into a slide plan:
From the following talk text, build a plan for 10 slides: for each slide —
a title, 3 main points, a description of the chart/table to show.
Mark the places where complex jargon needs simplifying.
[Text]
Note: grant organizations are also starting to adopt AI policies — if AI was used in the application, disclose it as the guidelines require. Some foundations do not accept fully AI-generated applications.
Agreement with the academic advisor
Hiding AI use from your academic advisor is the worst strategy. Instead, talk openly at the start of the research:
- Say at which stages you want to use AI (literature search, language editing, code writing).
- Ask about the advisor's boundaries — some advisors oppose text generation but accept editing.
- Record the agreement in writing (at least by email).
This conversation takes 15 minutes, but it lets you answer "Did you use AI?" with confidence at the defense. In many Uzbek universities this issue is still under discussion — your openness can also serve as an example for the department.
Final reminder: AI tools are updated every month — a method that is correct today may become obsolete tomorrow. Re-check journal policies before each submission and exchange experience with colleagues in your field. Scientific integrity is an unchanging value; the ways it is applied are updated with the times.
