03 · What You Need to Know
What Counts as an Assumption in a Research Idea?
An assumption is not automatically a mistake
Research cannot proceed without assumptions.
Researchers assume that constructs can be defined meaningfully, that measurements capture something relevant, that observations provide information about the phenomenon, that participants interpret questions in sufficiently comparable ways, and that particular methodological procedures permit particular inferences.
The objective is therefore not to eliminate assumptions. It is to know which assumptions your study depends on, determine whether they are reasonable, and recognize what happens if they fail.
Assumption
A premise the research reasoning depends on, whether explicitly stated or not.
Unsupported assumption
A premise treated as established even though the evidence, theory, design, or reasoning does not adequately justify it.
AI can interrogate the words you use to frame the problem
Terminology is rarely innocent.
Consider a proposal about “students' overreliance on generative AI.” Before any data are collected, the phrase already suggests that students are using AI more than they should.
What threshold distinguishes reliance from overreliance? Relative to what standard? Does asking AI for feedback count? What about generating a complete response? Is reliance inherently undesirable? Could external cognitive support sometimes be adaptive?
Generative AI can help surface these questions because it can rewrite the phenomenon from different conceptual perspectives.
The result may reveal that what appeared to be a neutral research topic already contained the conclusion.
AI can challenge causal assumptions
Research questions frequently imply causality before the design has earned it.
“How does social media use reduce academic performance?” presumes a negative causal effect. “Why does generative AI decrease creativity?” does the same.
An AI critique can ask whether the relationship might operate in reverse, whether another variable could explain both, whether the effect is conditional, or whether the evidence establishes association rather than causation.
This connects directly with using AI to generate alternative explanations for the research problem. Alternative explanations are one way of exposing causal assumptions that have become invisible in the original framing.
AI can challenge construct assumptions
Researchers may use familiar concepts as though their meaning were obvious.
What exactly is “AI literacy”? Is it knowledge, skill, critical evaluation, ethical judgment, practical capability, or some combination? Is “AI use” frequency, duration, purpose, sophistication, or dependency? Does “academic performance” refer to grades, standardized assessments, task performance, or learning gains?
AI can ask whether the construct has been defined too broadly, whether several distinct phenomena have been collapsed into one label, or whether the operational definition matches the conceptual definition.
These questions become especially important when researchers adopt fashionable terminology faster than measurement traditions can stabilize around it.
AI can question why your chosen population is the relevant one
Researchers often select populations partly because they are accessible. Convenience is not automatically a methodological sin, but it should not be disguised as theory.
If you propose studying first-year university students, AI can ask why year level matters. If the answer is simply that those are the students you can recruit, then the sampling choice is logistical rather than theoretical.
That may still be acceptable. What changes is the claim you can make about why the population was chosen and how broadly the findings should be interpreted.
AI can challenge assumptions about what counts as the desirable outcome
Research questions can contain normative premises.
Studies of technology adoption may implicitly treat adoption as success. Research on engagement may assume more engagement is always better. Work on productivity may privilege output over other values. Studies of AI acceptance may frame reluctance as a deficit requiring correction.
You can ask AI:
“What values are embedded in the way I have framed the desirable outcome?”
This can expose questions that conventional methodological checking may miss. Perhaps non-adoption is rational under certain conditions. Perhaps efficiency comes at a cost to learning. Perhaps lower engagement with one activity reflects strategic allocation of effort rather than disengagement.
AI can challenge assumptions about the literature
A research idea may depend on statements such as “few studies have examined this,” “existing research focuses mainly on students,” or “previous findings are inconsistent.”
Those are not merely background sentences. They are assumptions about the scholarly record until verified.
Ask AI to identify every claim in your rationale that would need literature evidence. Then verify those claims yourself.
This is particularly useful because an AI-suggested research gap is not itself evidence that a gap exists. AI can challenge an unsupported gap claim, but it can just as easily generate one.
AI can challenge methodological assumptions before data collection
Researchers may assume that their chosen method measures what they need simply because the method is familiar.
A survey does not automatically measure behavior accurately. An interview does not provide direct access to unconscious processes. A cross-sectional association does not establish temporal order. A statistically significant difference does not automatically indicate a practically important effect.
You can provide AI with the research question and proposed method and ask:
“What must be true for this design to answer this question?”
The resulting list can reveal assumptions about measurement validity, participant recall, sampling, missing data, causal identification, intervention fidelity, researcher interpretation, or statistical modeling.
This overlaps with evaluating whether the question is feasible, but the emphasis here is different. Feasibility asks whether the study can realistically be conducted. Assumption testing asks what must be true for its reasoning to hold.
AI can expose assumptions about novelty and importance
Researchers can become attached to an idea before they have established what it contributes.
You might assume that studying a new technology creates novelty, that studying an underrepresented population creates significance, or that an unanswered question must be important.
AI can be instructed to attack those premises directly:
“Assume this study is less novel than I think. What existing lines of research might already cover the underlying phenomenon?”
Or:
“Assume this question is genuinely unanswered but scientifically unimportant. Why might that be?”
These prompts force novelty and importance apart, which can make the rationale considerably more honest.
AI can challenge the theory you prefer
A favorite theoretical framework can become a lens through which every observation conveniently confirms the theory.
You can ask AI what the phenomenon would look like under competing frameworks, what evidence your preferred theory struggles to explain, which predictions would distinguish the theories, or what assumptions the theory makes about actors, context, cognition, institutions, or causality.
Again, the model may misunderstand the theory. Verify its characterization against authoritative theoretical literature.
AI can also reinforce your assumptions if you prompt it badly
Generative AI is not automatically adversarial. If you ask, “Explain why my theory is appropriate,” the system will usually attempt to comply.
If you provide a strongly framed premise, it may elaborate that premise rather than question it.
Watch Out
AI can become an extraordinarily articulate confirmation-bias machine when every prompt asks it to justify what you already believe. Challenge requires prompts that explicitly permit rejection, reframing, and disagreement.
Research on scientific hypothesis generation also suggests caution about treating generated reasoning as inherently truthful. LLMs can produce useful hypotheses, but studies continue to document challenges involving hallucination, truthfulness, evaluation, and generalization.
Challenge the assumption, then challenge the challenge
A useful AI criticism should not automatically replace your original view.
Suppose AI argues that your construct is culturally biased. That is a claim requiring examination. Suppose it suggests your causal direction may be reversed. Ask what evidence supports the reverse relationship. Suppose it says another theory fits better. Read the theory.
A strong workflow therefore has two rounds of skepticism.
State the research idea
Describe the problem, question, rationale, theory, and proposed design as clearly as possible.
Expose assumptions
Ask AI to identify what must be true for your reasoning to hold.
Generate challenges
Ask which assumptions are contestable, unsupported, value-laden, or vulnerable to alternative explanations.
Interrogate the AI critique
Ask what evidence would support each criticism and verify those claims independently.
Revise selectively
Change the research idea where the critique survives scrutiny; retain assumptions that are necessary and defensible.
Not every assumption needs to disappear
Research always has boundaries. You may deliberately assume that a validated scale functions adequately in your target population based on existing evidence. You may define a particular construct narrowly because that definition matches your theoretical purpose. You may restrict the population because the question concerns that population specifically.
The goal is intellectual visibility.
You should know what the study assumes, why those assumptions are reasonable, which are empirically testable, and how conclusions would change if important assumptions failed.
