Research methodology · Research insights
You Are the Researcher: Why AI Cannot Certify Your Research Idea
Turn your observations into a testable research question. Learn how PhD scholars can use AI without outsourcing scientific judgment or mistaking a suggestion for novelty.
The Novelty Toolkit · Series 1 · Part 5 of 9
A useful research question starts with something worth explaining. For PhD scholars and academic researchers, AI can help organise a literature review, compare explanations and refine a hypothesis. It cannot establish the originality of a research proposal merely by calling the idea innovative.
Imagine walking into a library and asking the librarian to invent a discovery for you. You may receive a fascinating list of topics. But you have not yet observed a problem, established a gap or tested a claim. An AI assistant can make that experience feel instantaneous—and that speed makes it easy to confuse an appealing suggestion with a breakthrough.
The guiding rule of this series is simple: the scholar supplies the scientific problem and makes the final judgments. AI can retrieve, compare, explain, challenge and suggest possibilities. None of those abilities, by itself, certifies that an idea is original.
Your observation is the beginning of research
Think of two students looking at a water-quality sensor installed in a pond. One asks AI, “Suggest a novel smart-water project.” The other records that the sensor measures accurately in the laboratory but drifts after several weeks in the field. The second student has a stronger starting point: a specific, observable problem whose explanation is uncertain.
A useful research question grows from a mismatch: what you expected versus what you actually see, or what the literature claims versus what the evidence appears to show. The observation might come from experimental failure, fieldwork, clinical experience, a mathematical contradiction or careful reading.
- 01Observe and defineRecord the mismatch between what you expected and what happened.
- 02Explore with assistanceAsk for explanations, prior studies and competing hypotheses.
- 03Judge and testVerify the sources and decide which measurement could change your mind.
The library analogy: powerful, but not literal
It is useful to imagine AI as a library you can question. Unlike a static library, however, a generative model can reason over inputs, combine concepts and propose hypotheses. Indeed, empirical work has found AI suggestions that human reviewers judged creative or novel in particular settings. That does not prove they were genuinely unprecedented or viable after execution.
The practical rule is therefore stronger than the statement “AI ideas already exist”: any AI-generated innovation claim is unverified until checked against prior work and evidence. An AI suggestion might duplicate known work, combine familiar methods in an interesting way, or propose something no one has yet tested. The scholar must find out which.
Core Principle
AI-generated is neither automatically old nor automatically novel. Novelty is a claim to investigate, not a label a chatbot can award.
Divide responsibilities correctly
| Researcher owns | AI may assist |
|---|---|
| Observation and problem choice | Generate alternative problem formulations |
| Scientific assumptions and ethical boundaries | List hidden assumptions and possible confounders |
| Choice of meaningful outcome | Suggest outcome measures and comparisons |
| Final reading of primary sources | Find and summarize candidate sources |
| Experimental design and interpretation | Critique possible methods and identify weaknesses |
| Publication and intellectual contribution | Help organize notes and draft accessible explanations |
A better first conversation
Weak opening: “Give me five breakthrough PhD topics in environmental engineering.” The question supplies no observations, evidence or constraints with which to evaluate the answer.
Example prompt
My observation is [specific phenomenon]. I have measured or observed [evidence]. Existing methods I know are [methods]. My constraints are [time, cost, data, equipment]. Do not invent a project for me yet. First identify five possible explanations, the assumptions behind each, and the evidence needed to distinguish them. Label guesses as hypotheses.
This request makes AI an information and reasoning aid rather than an oracle. More importantly, it keeps the researcher’s field knowledge visible in the conversation.
From doubt to scientific contribution
- 01Name the suggestionDescribe the actual mechanism, not a fashionable label.
- 02Compare prior workRead the nearest methods, findings and operating conditions.
- 03Design a discriminating testChoose a comparison that could support or refute the difference.
- 04Bound the claimState what the evidence establishes and what remains unknown.
Suppose AI recommends “AI-based drift compensation.” Before calling that innovation, ask what compensation methods already exist, whether the proposed mechanism is new, and what measurable difference matters. A serious outcome might be a new explanation of drift, a more reliable validation method or a demonstrable advance under a meaningful constraint.
An exercise before you open any chatbot
- Write three things you directly observed or measured.
- Write what existing theory predicts about each observation.
- Circle one mismatch between prediction and reality.
- Write one sentence describing what evidence would settle that mismatch.
- Only then ask AI to identify related methods and alternative explanations.
Three traps that make a suggestion look original
| Misleading signal | What to do instead |
|---|---|
| The phrase has no search results | Search the underlying concept using several terminologies |
| Two known technologies are combined | Compare the mechanism, outcome and conditions with prior implementations |
| AI calls the idea a breakthrough | Request closest prior studies and independent experimental validation |
A new label can hide an old method. A new combination can be important—or can repeat a known design without changing what is learned. A confident answer often hides an incomplete search. Treat each signal as an invitation to investigate, not a reason to stop. The originality of a scientific contribution is established by the closest defensible comparison and the evidence for the difference.
The two-minute input test
Before your next AI conversation, finish three sentences: “I observed …”; “I do not yet understand …”; and “I would change my explanation if …”. The third sentence is particularly useful because it tells you what kind of result could overturn the preferred hypothesis. Put these three sentences into the prompt, then ask for competing explanations and source-backed ways to check them.
A researcher does not outsource curiosity. Use AI to expand the space of questions, then return to evidence to decide which questions deserve experiments.
What the evidence does—and does not—show
Si, Yang and Hashimoto’s ICLR study examined expert judgments of research ideas in a defined setting; an idea judged novel is not automatically a successful experiment. The accessible 2025 preprint of The Ideation-Execution Gap follows ideas into execution and reports that judgments can change when outcomes are available. It is cited here as a preprint, separately from the published ICLR idea-evaluation study.
Messeri and Crockett’s Nature Perspective discusses the risk of mistaking fluent AI assistance for understanding. It is a conceptual analysis, not a trial of this tutorial. Meincke and colleagues’ Nature Human Behaviour analysis concerns idea diversity in specific brainstorming tasks; it does not establish that all AI-assisted scientific work is unoriginal. Together, these sources justify checking the evidence rather than accepting or rejecting an idea simply because AI suggested it.
The examples and prompt frameworks are educational illustrations. Verify sources, use domain expertise and follow research ethics. Do not upload confidential or patentable material to an AI service without authorisation.
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
- Messeri & Crockett. Artificial intelligence and illusions of understanding in scientific research. Nature (2024), Perspective. Published 6 Mar 2024
- Si, Yang & Hashimoto. Can LLMs Generate Novel Research Ideas? ICLR (2025).
- Si, Hashimoto & Yang. The Ideation-Execution Gap: Execution Outcomes of LLM-Generated versus Human Research Ideas. arXiv preprint (2025). Published 25 Jun 2025
- Meincke, Nave & Terwiesch. ChatGPT decreases idea diversity in brainstorming. Nature Human Behaviour (2025), Matters Arising. Published 14 May 2025