Start With a Mess, Not a Masterpiece
A vague interest is not a research question. "I want to study social media and mental health" is a topic — a parking lot the size of three departments. A research question is narrow, answerable with evidence you can actually get, and arguable enough that a smart person could disagree with your conclusion. The gap between those two is where most proposals die.
AI is good at closing that gap if you use it as a sparring partner, not a vending machine. You don't ask it for a question. You hand it your mess and make it interrogate you. The work is yours; the pushback is what you're renting.
Start by dumping everything you actually know and care about — unfiltered.
I'm a [your level/field] student. I'm interested in [topic],
specifically [whatever detail hooks you]. I have access to
[datasets, archives, survey tools, lab equipment, none].
I have [number] weeks and [solo / small team].
Ask me 8 sharp questions to narrow this into a researchable
question. Don't propose questions yet. One question at a time,
and adjust based on my answers.
The one-at-a-time constraint matters. A wall of eight questions gets a lazy wall of eight answers. A conversation forces you to think.
Run the Question Through a Gauntlet
Once you have two or three candidate questions, stop polishing and start attacking. A defensible question survives scrutiny before your supervisor applies it, not during your defense.
The standard screen is FINER — Feasible, Interesting, Novel, Ethical, Relevant. It's a checklist, and checklists are exactly what AI is good at running mercilessly.
Here are 3 candidate research questions: [paste them].
For each, score 1-5 on Feasible, Interesting, Novel, Ethical,
Relevant. Be harsh on Feasible given my [time/resources above].
Flag any that need IRB/ethics approval or data I probably can't
get. End with the single biggest weakness of each.
Watch for the feasibility trap especially. AI will happily bless a question that requires longitudinal data over five years or interviews with 200 executives. You have a semester. When it flags "this needs data you probably can't get," believe it — that's the cheapest failure you'll ever buy.
Then run the opposite test. Make it argue against you.
Pick my strongest question. Now play a skeptical reviewer on a
committee. Give me the three objections most likely to sink this,
and what I'd need to add to survive each one.
If the objections are easy to answer, your question might be too safe — bordering on a question whose answer everyone already knows. If they're impossible to answer, it's too ambitious. You want objections that are real but addressable. That's the sweet spot of "defensible."
Scope It Until It Hurts
Most first-draft questions are still 30% too big. The fix is to add constraints until the question almost feels small — population, timeframe, context, mechanism. "Does social media use affect anxiety?" becomes "Is daily Instagram use associated with self-reported anxiety among first-year university students in a single cohort?" Boring? No. Answerable in a semester? Yes. Boring is finishable.
Use AI to expose where your terms are still mushy.
Rewrite this question 4 ways, each one narrower than the last,
by adding constraints on population, timeframe, and context:
[question]. For each version, tell me what becomes easier to
measure and what I'm giving up.
Then flag any word that's still vague or unmeasurable
("impact", "engagement", "wellbeing") and suggest a concrete
operational definition.
You won't take the narrowest version. But seeing the full ladder tells you exactly which rung your time and data can support — and forces you to define what you actually mean by "engagement" before a reviewer makes you. If you want a deeper walkthrough of the whole research arc, the AI for Academic Research course covers the full pipeline from question to defense.
Draft the One-Pager — Then Make It Defend Itself
A proposal is just a question wearing a suit. The standard skeleton:
- Background — two or three sentences on what's known.
- Gap — the specific thing that isn't known.
- Question / hypothesis — your scoped question.
- Method — how you'll actually answer it.
- Significance — why anyone should care.
- Feasibility — why you can finish it on time.
Write the first draft yourself, badly and fast — bullet points are fine. Then use AI to pressure-test the logic, not to write the prose. This ordering is deliberate: if AI writes the draft, it writes the thinking, and you'll defend ideas you don't actually hold.
Here's my one-page proposal draft: [paste].
Check the logic chain only: does the gap follow from the
background? Does the question address the gap? Does the method
actually answer the question? Point to any link that's broken
or assumed. Don't rewrite my prose — just find the holes.
The most common break is the method-question mismatch: a survey that can't establish the causation your question claims, or a sample that can't support the generalization. Catch that on one page now, not in chapter three of a finished draft.
One integrity note, because it matters from the first draft: the gap, the method, and the significance have to be yours. Using AI to stress-test your reasoning is fine in almost every program. Letting it invent a gap you don't understand — or cite literature you haven't read — is how you end up defending a proposal built on a hallucinated reference. Keep a one-line log of what you asked AI and what you decided; future-you, and possibly your ethics board, will want it.
You now have a question that survived a gauntlet and a one-pager whose logic holds together. Run it past a human supervisor next — but walk in with something already sharp enough to argue about, instead of a topic and an apology.