Research methodology · Research insights
From Observation to Research Proposal: A Complete Human–AI Tutorial
Follow a complete human–AI research workflow from observation to literature search, hypothesis and proposal. Includes practical prompts, evidence checks and a fictional example.
The Novelty Toolkit · Series 2 · Part 9 of 9
A doctoral student notices that inexpensive water-quality probes disagree more often after weeks of outdoor operation. The student could ask AI for a 'completely novel smart sensor' and receive a polished answer within seconds. That answer might be useful for brainstorming, but it would neither explain the observation nor prove originality.
Instead, we will follow a disciplined researcher's notebook. The scientist owns the question at every stage. AI helps clarify terminology, suggest searches, organize real papers and challenge hypotheses. All scientific papers mentioned in the example matrix are hypothetical; the actual novelty of this proposed study has not been established.
Stage 1. The human records the observation
Researcher note: 'During repeated checks, two low-cost sensors appeared to diverge from reference readings after prolonged outdoor use. I do not yet know whether this reflects temperature, calibration age, fouling, changes in water composition, reference measurement error or sampling conditions.'
This note is appropriately cautious. It reports what was seen and names uncertainty. It does not make the leap from disagreement to a chosen solution.
AI's permitted role: ask about instrument model, reference method, number of readings, operating conditions, calibration records and measurement units. The researcher supplies answers or marks them unknown.
Example prompt
Act as a methodical laboratory colleague. I will provide an observation and its known measurement conditions. Ask me clarifying questions that distinguish observations from interpretations. Do not recommend a solution or invent evidence. End by identifying the three most important missing pieces of information.
Stage 2. The human chooses a provisional problem
After reviewing the questions, the researcher decides: 'The important uncertainty is how temperature and time since calibration jointly affect agreement with a reference method during extended field use.'
This is not yet a research gap. It is a bounded uncertainty the researcher considers worth investigating. AI can offer alternate phrasings, but the researcher's decision should reflect available instruments, expertise and relevance.
Stage 3. AI helps plan—but does not perform—the literature search
The researcher asks AI for synonyms: 'sensor drift', 'measurement stability', 'calibration interval', 'temperature compensation', 'field validation', 'low-cost monitoring'. AI returns candidate Boolean queries. The researcher actually executes these queries in relevant scholarly databases, records the dates and exports real results.
Example prompt
Using this human-defined question [insert], propose three database-ready search strategies from broad to focused. Explain what each is intended to find. Do not fabricate papers. Identify which additional source types—preprints, conference papers, theses or patents—could materially affect the assessment.
The researcher then checks titles, abstracts, methods, relevant full texts and citation chains. They do not accept a confident 'no previous studies found' statement from AI unless the underlying search is adequately documented and the closest papers have been independently examined (the PRISMA-S search-reporting guideline; Cochrane’s study-search guidance).
- 01Observe and defineSeparate the measurement from its possible explanation.
- 02Search and verifyRetrieve real records and read the closest work.
- 03CompareBuild a source-grounded evidence matrix.
- 04ChallengeConsider rival explanations and discriminating tests.
- 05PlanAlign the question, objectives and methods.
- 06AuditCheck sources, limits and the conditional contribution.
Stage 4. Build a source-grounded evidence matrix
For teaching purposes, suppose a matrix contains three fictional papers:
- Study A evaluates short-term temperature response in laboratory water samples, but does not assess prolonged field exposure.
- Study B records long-term field measurements, but does not isolate temperature from calibration ageing.
- Study C tests a correction algorithm, but lacks an independently held-out evaluation period.
The researcher must replace these placeholders with real source-verified papers before drawing any conclusion. AI can draft the structure and compare supplied passages; the human checks every number and limitation.
A cautious provisional gap might read: 'The documents examined so far have not jointly resolved temperature-associated error and time-since-calibration effects under the defined validation conditions.' It still requires targeted searching and expert criticism.
Stage 5. Red-team the explanation
The researcher proposes: 'Temperature contributes measurably to error even after accounting for calibration age.' AI is asked to list rival explanations and discriminating tests. The scientist considers biofouling, sensor-type differences, sampling inconsistencies and variation in the reference method.
Example prompt
Here is my testable hypothesis and my planned comparison [paste both]. Act as a skeptical reviewer. Identify confounders, alternative explanations, possible measurement bias and situations where my hypothesis would be wrong. For every objection, suggest an observable check. Do not score novelty or assert facts that I cannot verify.
This stage may radically change the design. That is success, not failure. Discovery often begins when an attractive explanation meets a better objection.
Stage 6. Convert uncertainty into a study plan
Main question: 'How do ambient temperature and elapsed time since calibration relate to error against a reference measurement during prolonged outdoor monitoring?'
Objective 1: Estimate error and repeatability under a documented field measurement schedule.
Objective 2: Compare error across specified temperature and calibration-age conditions while accounting for device-level repeated measurements.
Objective 3: Evaluate a pre-specified correction approach against a baseline on held-out measurements, if sufficient valid data are available.
The method must make these comparisons feasible. Define sensor inclusion criteria, reference method, calibration protocol, sampling intervals, important covariates, missing-data rules and analysis approach. Consider sample-size or precision planning appropriate to the design, rather than choosing an arbitrary number of readings. Use FINER to test feasibility, ethics, potential novelty and relevance (Stanford’s research-question guidance).
- 01Own the questionExplain why this uncertainty matters.
- 02Verify the evidenceRead and check the essential sources.
- 03Design the testChoose valid measurements and comparisons.
- 04Report honestlyState limitations, revisions and the role of AI.
Stage 7. Make an evidence audit before claiming novelty
Ask four uncomfortable questions. Have I found the closest previous method, including work with different terminology? Could a paper or patent already disclose the proposed correction? Does the new design resolve a meaningful uncertainty or merely change a setting? Can the study demonstrate the claimed difference using valid evidence?
At this stage the appropriate output may be a conditional novelty statement, not a declaration: 'The contribution is provisionally framed as an independently validated analysis of two potential sources of field measurement error, subject to completion of the documented prior-work audit.'
Stage 8. Record the division of labour
A transparent research log should show human-originated observations, assumptions and decisions; AI-assisted question clarification, search-term generation, evidence organization or methodological criticism; and independently verified references and findings. Do not upload confidential or sensitive information without authorization. Follow relevant institutional and publisher rules on documenting AI use; the ICMJE advises that authors remain responsible for content and should disclose AI assistance where applicable (ICMJE’s AI publishing recommendations).
The reader's reproducible practice challenge
Use the same stages with your own topic. Produce exactly seven artefacts: an observation note; a provisional problem statement; a search strategy with log; a verified three-paper matrix; a set of rival explanations; a question-objective-method alignment sheet; and a conditional novelty statement with limitations. Give those seven artefacts to a supervisor or peer for criticism.
If the reviewer finds that your idea already exists, the process has worked: you discovered the relevant prior art before investing years. You can now refine the question, identify a genuine unresolved feature, or choose another investigation. If the idea remains distinctive after rigorous comparison, you have a defensible reason to proceed—not an AI-issued certificate.
Stage 9. The proposal abstract the researcher finally writes
Here is a restrained example of a proposal synopsis derived from the hypothetical work:
Inexpensive field sensors are useful in distributed environmental monitoring, but field measurements may be affected by both environmental conditions and calibration history. This study will estimate the association of ambient temperature and time since calibration with measurement error relative to a predefined reference method during a longitudinal deployment. It will compare baseline measurements with a prespecified adjustment strategy using held-out observations and report uncertainty, device variability and important limitations. The contribution will be framed relative to the closest verified prior studies identified during the literature audit.
The paragraph does not claim to be the first ever. It says what the study plans to examine and how the research contribution will be established. That restraint makes it more credible.
Stage 10. A decision gate before collecting data
Ask the researcher to sign off five questions. Is the problem meaningful independent of the proposed technology? Is the gap supported by real, close prior work? Does the design distinguish important alternatives? Are measurements and resources adequate? Is the novelty claim appropriately conditional? A 'no' answer means revise the plan, not ask AI to invent a more flattering justification.
After the design is stable, consider registering a study plan where relevant. A transparent record can preserve which decisions were made before results were available, but not every exploratory project benefits from the same registration approach (OSF’s preregistration guidance).
Stage 11. The change-of-mind test
Suppose a newly retrieved paper demonstrates exactly the analysis the researcher intended. This is not an intellectual defeat. The scholar can compare weaknesses and strengths, identify genuinely untested conditions, investigate a competing mechanism or redirect the project. The key requirement is that the revised contribution be scientifically substantial and not simply a new label for an existing method.
Final lesson: the scientist owns the uncertainty
AI is an information instrument, not the principal investigator. It can widen the search, challenge comfortable assumptions and organize complex material. But it cannot replace the researcher's responsibility to observe carefully, ask an important question, verify sources, design valid tests and change course when evidence demands it. Use AI to become a more critical researcher, not to avoid becoming one.
Methodological context and limits
Messeri and Crockett’s Nature Perspective frames the limits of apparent AI understanding. The worked workflow illustrates those boundaries; its sensor studies, predictions and proposed contribution remain hypothetical until replaced by verified evidence.
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.
Related articles
Sources and further reading
- Rethlefsen et al. PRISMA-S: an extension to the PRISMA Statement for Reporting Literature Searches in Systematic Reviews. Systematic Reviews (2021).
- Cochrane Handbook for Systematic Reviews of Interventions. Chapter 4: Searching for and selecting studies.
- Stanford University. Selecting Your Research Question: FINER and study planning. Institutional guidance.
- ICMJE. Use of Artificial Intelligence in Publishing. Current recommendations, accessed 22 November 2026.
- Open Science Framework. Registrations and preregistrations. Official guidance.
- Messeri & Crockett. Artificial intelligence and illusions of understanding in scientific research. Nature 627, 49–58 (2024), Perspective. Published 6 Mar 2024