
A research question can sound original and still have been answered already. The problem often starts when students ask AI to suggest a topic before checking what researchers have actually studied.
AI can help you explore a subject, compare papers, identify differences between findings, and refine a question. But it cannot establish that a research gap exists simply by producing a convincing explanation. That requires reading real studies, checking their methods and limitations, and searching the wider literature.
The most reliable approach is to use AI after you have gathered relevant sources not instead of gathering them. This guide explains how to build a research question from evidence, with a worked example using published studies on social media use and sleep among university students.
Key Takeaways
Start with a research problem, then use published studies to understand what is already known.
Use AI to organize evidence and suggest possible questions, not to certify originality.
Compare research methods, populations, measures, findings, and limitations before deciding what to investigate.
Treat a research gap as a possibility until you have searched the relevant literature thoroughly.
Keep every important claim connected to a paper, page, or section that you have checked yourself.
Protect unpublished work and participant information when using AI tools.
Brainstorming a Topic Is Not the Same as Building a Research Question
Brainstorming gives you possible directions. A research question needs something more: a defined population, a specific issue or relationship, a researchable scope, and evidence showing why the question is worth investigating.
For example, asking an AI tool to suggest questions about student wellbeing might produce:
How does social media use affect the sleep quality of university students?
This is a starting point, not a finished research question. It leaves several things unanswered. Which students? Which platforms? What does social media use mean total daily time, late-night activity, or repeated checking? How will sleep quality be measured? Does the question assume that social media causes poor sleep?
A useful question becomes more precise as you compare it with existing research.
What makes a question source-grounded?
A source-grounded question has a traceable connection to published evidence. You should be able to explain:
What previous researchers investigated.
Which populations and settings they studied.
How they measured the variables.
What they found and what they did not find.
Which limitations or unresolved issues might justify further research.
A possible gap is not simply a topic that an AI tool failed to mention. It is a distinction supported by your review of the available literature.
Start With Research Papers, Not AI-Generated Answers
Before asking AI to compare studies, decide what counts as credible evidence for your topic.
For empirical research, prioritise original peer-reviewed studies. Systematic reviews and meta-analyses can help you understand the broader evidence base, while official datasets, government reports, and professional guidance may be more appropriate for certain applied questions. Peer review is a useful quality-control process, but it does not guarantee that a study is free from limitations.
For a broader overview of how to assess and report evidence, consult the PRISMA 2020 statement. It provides reporting guidance for systematic reviews; it is not a substitute for evaluating the quality of individual studies.
Choose sources that can answer your actual question
A search result or abstract can help you decide whether a paper is relevant, but it rarely provides enough detail to support a research question on its own.
When you open a paper, record its:
Full title, authors, publication year, DOI, or stable URL.
Research objective and study design.
Population, sample size, and setting.
Variables, measures, and data collection period.
Main findings and statistical results.
Limitations identified by the authors.
For example, a study based on self-reported sleep surveys does not provide the same kind of evidence as one that measures sleep using wearable devices. Both may be relevant, but they should not be treated as interchangeable.
Use research tools to locate and compare papers
Research platforms can make literature discovery more manageable. For example, Elicit offers tools for searching and analyzing academic papers, while Consensus provides academic search and evidence-oriented summaries.
These tools can help you identify candidate papers and organize information. However, their search collections and coverage differ. A result from one platform does not establish that every relevant study has been found.
Use AI-generated summaries as reading aids. Open the original paper to verify the study design, sample, results, and limitations before relying on a claim.
Worked Example: Narrowing a Research Question Using Real Studies
The best way to understand source-grounded research is to follow a question as it becomes more precise. Consider a student interested in the relationship between social media use and sleep among university students.
Step 1: Begin with the broad topic
The initial idea might be:
How does social media affect students' sleep?
This is too broad. It does not specify the student population, the type of social media activity, the sleep outcome, or whether the research is examining an association or a causal effect.
The next step is not to ask AI to declare a research gap. It is to find out what published studies have already investigated.
Step 2: Compare what real studies have examined
Consider these three research sources:
Study 1: Social media use and sleep among undergraduate students
Garett, Liu, and Young's study, The Relationship Between Social Media Use and Sleep Quality among Undergraduate Students, examined 197 first-year students at UCLA during their first academic quarter. The researchers collected Twitter activity and weekly self-reported sleep-quality information over a ten-week period.
The study reported that more frequent weekday late-night tweeting was associated with lower reported sleep quality. It also found a different association for weekday evening activity. This is important because it shows that the timing of social media activity may matter not just total use.
Read the full study on PubMed Central.
Study 2: A systematic review of social media, mental health, and sleep
Ahmed and colleagues' 2024 systematic review examined research on social media use, mental health, and sleep among young people. The review included 182 studies, with 98 included in its meta-analyses.
The authors reported considerable variation across study findings. Their results distinguished general social media use from problematic social media use, and the associations differed depending on the outcome being examined. The review also noted that there were not enough studies to examine some platform-specific relationships.
This matters because a broad question about all social media platforms may hide differences between types of use and outcomes.
Read the systematic review in the Journal of Affective Disorders.
Study 3: Measuring social media use and sleep beyond self-report
A more recent study, Beyond Self-Report: The Relationship Between Daily Social Media Use and Sleep in University Students, used a longitudinal repeated-measures design over 14 days. The researchers studied 62 university students, measured social media use using smartphone data, and used accelerometers to assess sleep.
The study did not find overall within-person or between-person effects of daily social media use on sleep duration, sleep-onset latency, sleep efficiency, or bedtime. It did report platform-specific between-person findings involving TikTok and bedtime, and Snapchat and sleep-onset latency.
This is a useful contrast with research based on self-reported social media activity and sleep. It shows why the measurement method and the specific platform should be considered before making broad claims.
Read the full study on PubMed Central.
Step 3: Identify what changes when you compare the papers
The three sources do not provide one simple conclusion. They investigate related questions using different populations, measures, and designs.
What to compare | What the papers show | Why it matters for your question |
|---|---|---|
Population | First-year students in one study; university students in another; young people across studies in the review | Avoid treating all student populations as identical |
Social media measure | Twitter activity, broader social media use, and smartphone-recorded platform use | Define exactly what counts as exposure |
Sleep measure | Weekly self-reports and accelerometer-based measures | State whether sleep is self-reported or objectively measured |
Time frame | Ten weeks, a 14-day period, and studies with varying designs | Specify the observation period |
Findings | Associations vary by timing, platform, outcome, and measurement approach | Do not assume every study supports the same conclusion |
This comparison reveals a more useful direction than the original broad topic: the relationship may look different depending on how social media activity and sleep are measured.
However, these three papers alone do not prove that a new research gap exists. A student would still need to search additional databases, review newer studies, and check related research before making a novelty claim.
Step 4: Narrow the question without overstating the evidence
The question can now be developed in stages:
Initial question:
How does social media affect students' sleep?
More focused question:
How is social media use associated with sleep quality among undergraduate students?
A more researchable direction:
Among undergraduate students, how are daily changes in platform-specific social media use associated with objectively measured sleep outcomes over a 14-day period?
The final version defines the population, exposure, outcome, and observation period. It also uses the word associated rather than affect, because an observational design does not automatically establish causation.
This is a candidate question, not a confirmed original research gap. The 14-day study already examined a closely related relationship. To justify a new project, the student would need to identify a clearly different population, setting, measurement approach, or unresolved result and then verify that distinction through a wider literature search.
The key lesson is simple: reading the papers changed the question. AI may help organize that comparison, but the evidence not the chatbot determines what can reasonably be asked.
Use AI to Compare Evidence, Not Invent Research Gaps
Once you have selected relevant papers, AI can help you organize their information. It can extract study details, group similar findings, highlight differences, and suggest questions for you to investigate further.
The important distinction is between asking a tool to report what the papers say and asking it to invent what the papers supposedly missed.
Build a literature comparison matrix
A literature matrix makes it easier to see similarities and differences across studies.
Source ID | Study design | Population | Main measure | Finding to verify | Limitation or question to investigate |
|---|---|---|---|---|---|
S1 | Longitudinal | First-year undergraduates | Twitter activity and self-reported sleep | Late-night weekday activity was associated with lower sleep quality | Platform and self-report context |
S2 | Systematic review and meta-analysis | Young people across included studies | Social media, mental health, and sleep measures | Findings varied by type of use and outcome | Heterogeneity and limited platform-specific evidence |
S3 | 14-day repeated measures | University students | Smartphone use and accelerometer-based sleep | Overall daily-use effects were not found for several sleep outcomes | Small sample and the need to assess how findings generalise |
This is a simplified example of a comparison matrix, not a substitute for the full extraction of each paper. Add page numbers, exact measures, and the original statistical results when preparing your own research notes.
Ask AI to extract information from selected papers
An AI document chat tool can help you work across a collection of PDFs. For example, NotebookLama allows users to upload documents and ask questions about their contents.
You can ask it to organize information from papers you have selected. But a citation or page reference returned by an AI tool is only a route back to the source. Open that page and confirm that it supports the summary.
For a related workflow, see AI Open-Book Exam Practice With Your Course PDFs.
A Practical Workflow for Building Source-Grounded AI Research Questions
A repeatable process makes it easier to move from an initial idea to a question you can defend.
Describe the problem. Write a short explanation of the topic, who it concerns, and why it matters. Avoid writing a final research question immediately.
Search for relevant literature. Use suitable academic databases and search platforms. Include synonyms and related terms rather than relying on one exact phrase.
Select and record sources. Give each paper a source ID. Record its citation, research design, sample, measures, findings, and limitations.
Use AI for structured extraction. Ask the tool to report information from your selected sources. Require source IDs and page references, and tell it to mark missing details as "Not reported."
Compare the evidence yourself. Look for differences in populations, methods, outcomes, settings, and findings. Do not interpret every difference as a research gap.
Draft several candidate questions. Make each question specific enough to investigate with your available time, data, and methods.
Search again using the candidate wording. Try alternative terms, related concepts, older terminology, and different databases. Check recent reviews, cited references, and newer papers.
Finalize the question. Confirm that the question is feasible, ethically appropriate, and consistent with the research design you intend to use.
Keep a record of the databases searched, search dates, terms, filters, and relevant results. This makes your process easier to explain and repeat.
Check Whether the Research Question Is Novel, Feasible, and Ethical
A promising question is not automatically a suitable research project. Before committing to it, check three things.
Verify that the research gap is real
Search for the exact concepts in your question, along with synonyms and related terminology. Look across more than one relevant database, especially when the topic overlaps different disciplines.
Read recent systematic reviews where available. Follow their reference lists and check newer papers that cite the studies you have found. Depending on your field, you may also need to examine registered protocols, dissertations, conference papers, preprints, and ongoing studies.
A gap in your first search results may simply mean that your search was too narrow. Do not claim that no research exists until you have conducted a sufficiently broad and documented search.
Check whether the question is feasible
A question can be interesting but impractical to investigate. Consider whether you can access the participants or data, recruit an adequate sample, use suitable measures, and complete the work within your timeline.
The wording must also match the research design. For example, a cross-sectional survey may identify an association, but it generally cannot establish that one variable caused another. Use terms such as associated with, related to, or predicts only when they accurately reflect the proposed analysis.
Consider ethics and data access
If your project involves interviews, surveys, health information, or identifiable records, check your institution's ethics requirements before collecting or uploading material.
Do not assume that removing names makes a dataset anonymous. Locations, job titles, unusual experiences, and combinations of personal details may still identify participants.
Protect Research Data When Using AI Tools
Research documents can contain sensitive information even when the main text appears harmless. Interview transcripts, unpublished manuscripts, survey exports, and working notes may include names, contact details, participant identifiers, or confidential findings.
Before uploading any material, check whether the tool is permitted under your institution's policies and the project's consent and ethics conditions.
Use a clean working copy containing only the information needed for the task. Remove unnecessary direct identifiers and inspect document comments, properties, version history, speaker notes, embedded files, and hidden spreadsheet rows.
Also review the active account, workspace permissions, retention terms, deletion process, data-sharing settings, and any connected services. A university account or paid subscription does not automatically make every research upload appropriate.
Keep original research records in the approved repository. Do not use a public AI tool to process restricted participant data or confidential material unless the relevant institutional and project requirements explicitly allow it.
Maintain Academic Integrity and a Clear Research Trail
AI can help with search-term variations, literature matrices, early outlines, and question refinement. It cannot take responsibility for your research decisions, source interpretation, citations, or conclusions.
Keep dated search logs, reading notes, draft questions, comparison matrices, and revisions. These records show how your research question developed as you engaged with the literature.
It also helps to distinguish three types of notes:
Source evidence: What the original paper reports, with a citation and page reference.
Your interpretation: Your explanation of similarities, differences, and possible implications across sources.
AI-generated suggestions: Ideas or draft wording produced by a tool that still require independent checking.
Do not treat AI-detection scores as definitive evidence of authorship or misconduct. Turnitin itself cautions that its AI Writing Report may be inaccurate and should not be the sole basis for adverse action against a student. See its guidance on using the AI Writing Report.
Follow your institution's and publisher's rules
Your instructor, university, research office, funder, and target journal may have different requirements for AI use. Check the applicable rules before using AI with unpublished manuscripts, peer-review material, participant data, or research findings.
For example, Nature Portfolio's editorial policies explain its expectations for AI use and author accountability. Elsevier's journal AI policy sets out disclosure requirements for AI use in manuscript preparation.
Where disclosure is required, a statement might read:
"I used an AI tool to generate search-term variations and organise notes from researcher-selected sources. I independently verified the citations, interpretations, and final wording against the original materials."
Adapt this wording to your institution's or publisher's actual requirements.
Reusable AI Prompts for Source-Grounded Research
A useful prompt tells the tool which sources it can use, what information to return, and what to do when the evidence is missing.
Prompt 1: Extract evidence from research papers
Using only the attached research papers, create a comparison table with one row per paper. Include a source ID, full citation, research objective, study design, population or dataset, sample size, setting, measures, key findings, limitations, and relevant page numbers.
Do not add information from outside the documents. Do not infer missing sample details, methods, or results. If a detail is unavailable, write "Not reported." Keep the authors' findings separate from your interpretation. Include a short supporting quotation only when necessary and provide its page number.
Prompt 2: Compare findings and possible limitations
Compare the attached papers using only the information in those documents. Identify where their populations, research designs, measures, results, and limitations are similar or different.
For each comparison, include the source IDs and page references. Separate what the authors explicitly state from your own interpretation. Do not describe a limitation as a research gap unless the evidence supports that conclusion. List any claims that require a wider literature search.
Prompt 3: Develop candidate research questions
Based only on the evidence table below, propose five possible research questions.
For each question, specify the population, setting, main variables or concepts, possible research design, and the source IDs that make the question relevant. Explain what previous studies have already examined and what remains uncertain.
Do not claim that any question is original. Mark every possible gap as "requires further literature search." Avoid causal wording unless the cited research design and proposed method can support causal inference.
Prompt 4: Refine the wording without changing the research meaning
Refine the following research question for clarity and precision: [insert question].
Return three versions. Do not change the population, setting, variables, outcome, time frame, or implied level of causation. Explain briefly how the wording differs in scope. Do not add a new research gap or claim that the question is novel.
Always review the output against your original papers. A well-written prompt can reduce unsupported additions, but it cannot guarantee that an AI system will extract every detail correctly.
Frequently Asked Questions
Can AI tell me whether my research question is original?
Not reliably. AI can help you locate related papers, compare findings, and identify possible distinctions. Originality requires a documented search across relevant literature, including recent reviews and primary studies. A missing result in one tool does not prove that nobody has investigated the topic.
Can I use AI to generate research questions for my thesis?
Yes, as a brainstorming and refinement aid, provided your institution permits it. Use your selected papers as the evidence base, verify every important claim, and make the final research decisions yourself.
Are AI-generated research citations reliable?
They should be treated as unverified until checked. Confirm the title, authors, publication details, and DOI through the publisher, an academic database, or the original paper. Then check whether the cited source actually supports the claim you intend to make.
What is the difference between a research topic and a research question?
A topic describes a broad area of interest, such as social media and student sleep. A research question defines the specific population, relationship, outcome, or issue you intend to investigate. The question should be narrow enough to guide your methods and analysis.
Can I upload interview transcripts to an AI tool?
Only if your consent terms, ethics approval, institutional policy, and the platform's data controls allow it. Remove unnecessary identifiers, use an approved workspace, and consider indirect identification risks before uploading any research material.
Build Research Questions That Can Withstand Scrutiny
A strong research question does not come from an AI-generated list of impressive-sounding ideas. It develops through reading, comparing, questioning, and verifying published evidence.
Use AI to reduce repetitive work and make it easier to organize research papers. Let it suggest possible directions, but do not let it decide what the literature proves or whether a gap exists.
The final question should be one you can explain with confidence: what previous researchers have found, what remains uncertain, why the question matters, and how you plan to investigate it. That is what makes an AI-assisted research question genuinely source-grounded.