How to Start Your First Research Paper: A Practical Guide for University Students in the AI Era

Many students want to begin research but remain stuck at the same question: Where do I start? They may read published papers, see complicated methods and polished results, and assume that researchers begin with a perfect topic and a complete plan. In reality, most good papers begin with uncertainty. The difference is that experienced researchers know how to turn that uncertainty into a sequence of small, testable decisions.

This guide presents the approach I wish more students were shown at the beginning. It is especially useful for students in software engineering, data science, machine learning and health-related fields, but the main principles apply much more widely. You do not need to discover a revolutionary idea on your first day. You need a meaningful problem, an answerable question, an honest method and the patience to improve the work through feedback.

1. Start with a problem, not a paper title

A common beginner mistake is trying to invent an impressive title before understanding the problem. Titles feel productive because they make an idea look complete, but a title cannot rescue a weak research question. Begin instead with something that is unclear, inefficient, unfair, poorly measured or not well understood. Look around your courses, community, workplace, public datasets and recent papers. What decisions are people making with limited evidence? What process takes too much time? Which population is missing from existing studies? Where do current systems perform well in a laboratory but struggle in the real world? These observations are more valuable than a list of fashionable algorithms.

For example, “artificial intelligence in healthcare” is a field, not a research problem. “Can routinely collected clinical variables identify patients at high risk of readmission?” is closer to a problem. It identifies a decision, a population and an outcome. The question can then be narrowed further after examining available data and earlier work.

2. Turn the problem into one answerable question

A strong first research question is specific enough to answer within your time, data and skill constraints. It does not try to solve an entire field. A useful template is:

Among [population or setting], can [method, exposure or intervention] help explain or predict [outcome], compared with [baseline or current approach]?

This template is not a rule; it is a test of clarity. If you cannot identify the population, the input and the outcome, the project may still be too broad. For qualitative or descriptive research, replace prediction language with questions such as “How do…?”, “What barriers…?” or “What patterns…?” The important point is that the question determines the data and method, not the other way around.

3. Choose the kind of paper before choosing the tools

Not every research paper needs a new dataset or a complex model. Your first major decision is the type of contribution you can realistically make.

·     An empirical study analyzes collected or existing data to answer a defined question.

·     A systematic or scoping review organizes existing evidence using a transparent search and selection process.

·     A benchmark study compares methods under one consistent and fair evaluation protocol.

·     A software or system paper develops a tool and evaluates whether it solves a real user problem.

·     A qualitative study explores experiences, perceptions or barriers through interviews, observations or textual data.

The best choice is not the one that sounds most advanced. It is the one that matches your question, available evidence, ethical requirements and timeline. Reporting resources such as the EQUATOR Network can help you identify the guideline commonly used for a particular study design. Systematic reviews, for example, commonly use PRISMA; prediction-model studies may use TRIPOD or TRIPOD+AI. Reading the relevant checklist before beginning can prevent major omissions later.

Beginners often download dozens of papers and then feel more confused. Reading becomes useful when each paper has a purpose. Start with five to ten recent papers that are directly connected to your question. For each one, record the problem, dataset or participants, method, evaluation, main result and limitation.

A simple spreadsheet is enough. Add one row per paper and include a final column titled “What this means for my study.” This forces you to move beyond summarizing. You may notice that several papers use the same dataset, ignore a particular population, report only accuracy, or evaluate with random splits that do not reflect real use. Those patterns can lead to a defensible research gap.

Use a reference manager such as Zotero or Mendeley from the beginning. Save the PDF, citation and your notes together. Do not wait until the final week to reconstruct references from browser tabs. Good organization is not administrative work separate from research; it protects your reasoning and saves time when the manuscript changes.

5. Find a gap without claiming that nobody has done it

A research gap does not always mean that no previous paper exists. That claim is difficult to prove and is often unnecessary. A gap may be a limitation in population, geography, data quality, evaluation, comparison, interpretation or practical use.

·     Population gap: a method was tested mainly on one demographic or clinical group.

·     Setting gap: results come from one hospital, institution, country or laboratory environment.

·     Evaluation gap: studies report strong performance but do not test time, subject, site or geographic generalization.

·     Method gap: an appropriate baseline, ablation or interpretable analysis is missing.

·     Application gap: a technically strong model has not been translated into a usable decision-support workflow.

A mature gap statement sounds like this: “Existing studies report promising results on randomly divided samples, but evidence about performance across institutions remains limited.” This is more credible than “No one has researched this topic.” It also points directly toward a study design.

6. Check feasibility before falling in love with the idea

Before committing, answer four questions. Can you obtain the data? Can you complete the work ethically? Do you have the required skills or supervision? Can the project be finished within the available time? A smaller completed study is more valuable than an ambitious project that never reaches evaluation.

For data-driven projects, inspect the dataset before finalizing the research question. Confirm the unit of observation, target variable, missingness, class balance, repeated subjects, measurement dates and usage license. A dataset with one hundred thousand rows may contain only a few hundred independent patients. A target label may have been created using information that would not exist at prediction time. These details can completely change what the project can honestly claim.

7. Write the evaluation plan before running experiments

This step is particularly important in machine learning research. If you repeatedly inspect test results and adjust the pipeline, the test set quietly becomes part of model development. The final score may look excellent while giving an unrealistic picture of future performance.

Decide early how data will be separated, which baseline methods will be compared, which primary metric answers the question, and what subgroup or robustness checks matter. If a patient has multiple records, keep that patient within one split. For future prediction, consider a time-based split. If the intended use is a new hospital or region, a site or geographic holdout may be more informative than a random split.

Accuracy alone is rarely enough, particularly when outcomes are imbalanced. Depending on the problem, include precision, recall, F1 score, ROC-AUC, PR-AUC, calibration or error measures, and explain why each matters. The objective is not to create the largest table. It is to make the evaluation match the real decision.

8. Create the paper outline before writing the introduction

A research paper becomes easier when it is treated as an argument rather than a long writing assignment. Before drafting full paragraphs, write one or two sentences under each section:

·     Introduction: What is the problem, why does it matter, what is missing and what does this study do?

·     Related work: How has the problem been approached, and where does your study differ?

·     Methods: What data and procedure would allow another researcher to understand or reproduce the study?

·     Results: What did the planned analyses show, without exaggeration?

·     Discussion: What do the findings mean, how do they compare with earlier work, and what are the limitations?

·     Conclusion: What is the most defensible takeaway and the next reasonable step?

Many students try to make the introduction perfect before the experiments are stable. It is often more efficient to draft the methods first because you know what you actually did. Results can follow when analysis is complete. The introduction and discussion become clearer once the contribution and limitations are visible.

9. Use AI as an assistant, not as the researcher

Generative AI has changed how students search, plan, code and write. Used carefully, it can help brainstorm keywords, explain an unfamiliar concept, improve grammar, generate a checklist or identify questions to discuss with a supervisor. It can reduce friction, but it cannot accept responsibility for your paper. Never trust an AI-generated citation without opening the original source. Do not upload confidential, identifiable or restricted research data to a public tool. Verify every factual claim, equation, code output and reference. Keep notes describing where AI was used, and check the target journal's policy before submission.

Current ICMJE guidance says authors should disclose the use of AI-assisted technologies and remain responsible for the submitted work. COPE states that AI tools cannot be listed as authors because they cannot take responsibility, manage conflicts of interest or agree to publication terms. The practical rule is simple: use AI to support your thinking, never to replace verification, judgment or accountability.

10. Ask for feedback before the paper feels ready

Waiting until a manuscript is “finished” makes feedback more expensive. Share the research question early. Ask whether the contribution is clear before building the full pipeline. Show your proposed evaluation before running every experiment. A supervisor can correct a weak design in one conversation; after weeks of analysis, the same correction may require starting again.

Ask specific questions. “Is this good?” is difficult to answer. Better questions include: Is the scope manageable? Does the split reflect the intended use? Is the baseline fair? Which claim is not supported? What would a reviewer question first? Specific questions produce specific improvements.

Day 1: Write three problems you genuinely care about. For each, explain who is affected and why the problem matters.

Day 2: Search for recent review papers and five directly relevant studies. Record their data, methods, results and limitations.

Day 3: Draft one research question using a clear population, input or intervention, and outcome.

Day 4: Identify the paper type, possible data source, ethical requirements and reporting guideline.

Day 5: Write a one-page concept note containing the problem, gap, question, proposed method and expected contribution.

Day 6: Draft the evaluation plan and list the main risks, including missing data, bias, leakage, limited sample size or unavailable labels.

Day 7: Discuss the concept note with a supervisor or experienced researcher and revise the scope before implementation.

·     Choosing a fashionable algorithm before defining the research problem.

·     Calling the model application itself a contribution without explaining what new evidence it provides.

·     Reading widely but keeping no structured notes or citation library.

·     Using the test set repeatedly during model development.

·     Reporting only the strongest metric and hiding failed experiments or limitations.

·     Selecting a journal only after the manuscript is complete and then forcing the paper into an unsuitable format.

·     Using AI-generated text or references without verification and disclosure.

Your first paper does not need to transform an entire discipline. It should make one clear, honest and reproducible contribution. A careful benchmark can be valuable. A well-designed negative result can be valuable. Testing whether a model generalizes to a new population can be more useful than adding another small improvement on the same dataset.

As a student researcher, I have learned that progress becomes much faster when the focus moves away from appearing advanced and toward making each decision defensible. Why this problem? Why this dataset? Why this split? Why this metric? What would change if the result were used outside the experiment? These questions improve both the research and the researcher.

Start small enough to finish, but carefully enough that the result deserves trust. Write the question before the code, plan the evaluation before the score, and ask for feedback before certainty. That is a much stronger beginning than waiting for a perfect idea.

·     EQUATOR Network: Search for reporting guidelines

·     PRISMA 2020 checklist for systematic reviews

·     TRIPOD+AI reporting guidance for prediction-model studies

·     ICMJE guidance on AI use by authors

·     COPE position on authorship and AI tools