RESEARCH / BLOG

Research,
Explained.

Read the research question, follow the authors’ methods and explore the evidence behind the article.

Research Blog

Research methodology · Research insights

Use AI to Clarify a Research Problem Without Outsourcing Your Thinking

Use AI to clarify a research problem while retaining scientific judgment. Follow a human-led workflow for definitions, assumptions, evidence and PhD research questions.

By NPT Research6 min read
The novelty toolkit · S205 / 09
Ask better questions of your idea.The researcher owns the observations and the final decision.
NPT Research

The Novelty Toolkit · Series 2 · Part 5 of 9

Picture a library with an exceptionally fast reference desk. You bring the puzzle; the librarian helps locate explanations, unfamiliar terms and useful sources. But the librarian never visited your laboratory, never observed your sensor fail and cannot determine what you actually saw. AI is useful in a similar way: it can organize possibilities and challenge language, but it should not replace your observations or scientific judgment.

The question is not 'Can AI give me a research problem?' It is 'Can I use AI to sharpen a problem that I can explain and defend myself?'

Write the human-only starting note

Before opening any AI tool, write five lines in your own words: what you observed, how often it occurred, what you already know, what you suspect, and what you do not yet understand. For example, 'Two low-cost water probes disagreed after several weeks outdoors; I have not ruled out temperature, contamination or maintenance effects.'

Do not embellish a single anecdote into a pattern. If you have no observations yet, state your scientific interest and the preliminary evidence you are trying to understand. Original thinking is not limited to personally collected data; it also includes careful questioning of theory and published results.

Ask AI to clarify, not to invent

A poor prompt says: 'Generate a groundbreaking, never-before-seen PhD problem in water monitoring.' It rewards confidence and novelty-flavoured language. A better prompt asks for specific uncertainties and exposes what is missing.

Example prompt

I am investigating a research problem. My observation is: [your observation]. My current explanations are: [your own ideas]. My available evidence is: [data or cited papers]. Ask me five clarifying questions, one at a time, about definitions, measurement conditions, plausible alternatives, feasibility and significance. Do not generate a novelty claim or research title. Separate my observations from interpretations.

This prompt uses AI as a questioning tool. You remain responsible for each answer.

The novelty toolkit · Series 2Keep the decisions with the researcher.AI questions can expose missing information; they do not supply observations.
  1. 01
    Human noteRecord observations, assumptions and uncertainty.
  2. 02
    AI questionsRequest clarification rather than a novelty verdict.
  3. 03
    Evidence checkVerify proposed facts and alternative explanations.
  4. 04
    Human revisionWrite the problem statement and explain each decision.
A learning framework, not a measured result. Verify evidence before making a research claim.
Figure 1. The researcher supplies the evidence and makes decisions; AI asks and challenges between those decisions.

Refuse the easy answer

An AI system may reply, 'This is likely sensor drift; build a deep-learning correction model.' That is an attractive story, not an established explanation. Respond: 'What other mechanisms could produce the same readings? Which observation would distinguish them? Which parts of your response are general knowledge rather than verified facts?'

This step matters because fluent language can make uncertainty look smaller than it is. Research on AI in science has warned about 'illusions of understanding': producing more apparent explanation without necessarily increasing genuine understanding (Messeri and Crockett’s Nature Perspective).

Have AI generate a confusion map

Ask the system to separate definitions (what does 'accuracy' mean here?), assumptions (same water sample), measurements (reference method), alternative explanations (temperature, fouling, ageing), and decisions that only you can make (which sensors, budget, sampling schedule).

Then inspect the map and mark each item: known, plausible but unchecked, unsupported, or needs a new measurement. If AI confidently adds a sensor property or citation that you did not provide, do not adopt it until verified.

The novelty toolkit · Series 2Label the status of every suggestion.Fluent language does not change the quality of the evidence.
  1. 01KnownSupported by a specific record or source.
  2. 02PlausibleA possibility requiring a check.
  3. 03UnsupportedA claim without adequate evidence.
  4. 04DecisionA choice the researcher must justify.
A learning framework, not a measured result. Verify evidence before making a research claim.
Figure 2. A useful AI conversation distinguishes decisions, possibilities and evidence.

Test with a deliberately narrow follow-up

Example prompt

Here is my revised problem statement: [text]. Critique it only for ambiguity, hidden assumptions, alternative causal explanations and missing operational definitions. Suggest no new technology unless a specific unresolved problem justifies it. For each critique, give me a question to answer or an observable test to consider.

The desired output is not a longer problem statement. It is a clearer one. If the AI introduces new facts, label them for literature checking rather than treating them as evidence.

Use a human sign-off rule

After every AI response, write one sentence: 'I accept/reject this suggestion because…' Record any important definition or assumption that changed, and why. This produces an intellectual audit trail. It also teaches you to disagree constructively with AI when its suggestions are premature.

Be careful with confidential or patentable ideas: do not send undisclosed technical details or sensitive participant data to external systems without suitable permissions and safeguards. Relevant publication venues may require disclosure of AI assistance, and the ICMJE states that AI is not an author and that authors remain responsible for their work (ICMJE’s AI publishing recommendations).

Reader exercise: ten prompts are not needed

Select one messy problem statement. Write the five-line human-only note. Run the clarification prompt once, respond to the questions, and use the critique prompt to identify three weaknesses. Stop. Produce a revised statement yourself and explain precisely which AI suggestion you adopted. This exercise values good judgment over prompt volume.

A worked mini-dialogue

Researcher: 'I saw inconsistent sensor readings after prolonged field exposure. I suspect temperature but lack a stable explanation.'

AI (useful response): 'Were measurements made from the same water sample? What reference method was used? Were time of day and calibration interval recorded?'

Researcher: 'No consistent reference measurements were taken. I should not call this inaccuracy yet; it is disagreement between devices.'

AI (useful follow-up): 'Then distinguish inter-device disagreement from error against a validated reference. You may first need a measurement protocol before testing a causal explanation.'

The scientific improvement came from the researcher correcting the claim after recognizing missing evidence. The AI did not contribute a miraculous invention; it helped expose a hidden weakness in the original wording.

Three decisions AI must not quietly make

Which observation is real? Only reliable records, instruments and observations can answer this. Which uncertainty matters? The researcher, collaborators, stakeholders and field evidence must judge significance. Which hypothesis is worth committing resources to? The research team must evaluate feasibility, ethics and competing explanations.

AI may provide options for each decision, but producing options is not the same as assuming responsibility for the choice. Require a written human rationale before moving to the next stage.

A stopping rule for AI conversations

Stop brainstorming when the system repeats similar suggestions or begins inventing specificity. Move to reading, measurement, expert consultation or analysis. More prompts are not always more progress. A productive research conversation changes what you can define, test or verify; an unproductive one merely accumulates plausible prose.

What to remember

AI can help you understand what you are asking; it cannot substitute for deciding why the question is worth asking. In the next article, we use AI for a narrower and more auditable job: improving the literature-search strategy.

Methodological context and limits

Messeri and Crockett’s Nature Perspective is a conceptual Perspective on scientific understanding, not a trial of these prompts. Use Stanford’s research-question guidance to consider scope, resources and ethics alongside the assistant’s suggestions.

Continue the research novelty series

Series 1: originality, evidence and scientific judgment

Begin with what makes research novel, then find the closest prior research and build a novelty matrix. Before using an assistant, examine the limits of AI novelty judgments and why the researcher owns the question. Develop the conversation with scientific prompting and six-round brainstorming, then challenge the idea against evidence. The complete worked research dialogue connects those decisions.

Series 2: from a problem to a defensible proposal

First separate a research gap from a research problem and identify a genuine problem worth investigating. Next search published literature for the gap and align the question, objectives and hypothesis. For AI-assisted work, clarify the problem without outsourcing judgment, verify the literature search and its references and compare papers in an evidence matrix. Then stress-test the research hypothesis before following the complete observation-to-proposal workflow.

Sources and further reading

  1. Messeri & Crockett. Artificial intelligence and illusions of understanding in scientific research. Nature 627, 49–58 (2024), Perspective. Published 6 Mar 2024
  2. ICMJE. Use of Artificial Intelligence in Publishing. Current recommendations, accessed 22 November 2026.
  3. Stanford University. Selecting Your Research Question: FINER and study planning. Institutional guidance.