TL;DR
Topic selection is the worst-shaped part of grad school: you do not yet know what you do not know. Use a general model (GPT-5.5, Claude Opus 4.7, or Gemini 3.1 Pro) to generate 15 directions in one pass — each with a falsifiable research question, an honest feasibility score, and a fit-to-advisor rating. Then verify every suggested paper with a real academic tool (Semantic Scholar is free; Elicit Basic is free) because chat models still fabricate citations. The output is a shortlist of 3 to take to office hours, not a topic to commit to.
The task
You need a thesis topic — undergraduate honours, master’s, or PhD proposal — and every idea you have either feels done, won’t fit your timeline, or doesn’t match your advisor’s interests. AI helps here for one reason: it scans broadly across a field and forces you to compare 15 directions instead of fixating on the first one. It does not help you judge which is correct, novel, or actually feasible — those still need a real literature search and your advisor.
What each AI tool is good for
A chat model brainstorms breadth. A purpose-built research tool verifies whether the papers it cited exist and whether the question is already answered. Use both. Pricing and limits below are as of June 2026.
| Tool | Best for here | Cost (June 2026) | Watch out for |
|---|---|---|---|
| ChatGPT (GPT-5.5) | Generating the 15 directions; clarifying-question style | Free (ads, tight limits) / Plus $20 | Fabricates paper titles and author names |
| Claude (Opus 4.7) | Long, structured reasoning over your constraints | Free (limited) / Pro $20 | Same citation risk; cannot browse on Free |
| Gemini 3.1 Pro | Wider source coverage, structured comparison tables | Free / Google AI Pro $19.99 | Free Deep Research capped at 5 runs/month |
| Semantic Scholar | Confirming a paper is real (200M+ papers) | Free | Discovery only, not idea generation |
| Elicit | Checking whether a question is already answered | Basic free (2 reports/mo) / Plus ~$12/mo | Built for lit review, not topic brainstorm |
Deep Research mode matters for the verification step: ChatGPT Plus includes about 10 Deep Research runs/month, Gemini’s free tier allows 5/month, and Google AI Pro raises that to roughly 20/day. Run the brainstorm on a normal chat (cheap, fast), then spend a Deep Research run only on your top 3.
When AI helps — and when it does not
AI is excellent at breadth: listing 15 directions, suggesting cross-disciplinary angles, and roughing out a research question. It is unreliable at depth. It confidently invents papers, misattributes findings, and overrates feasibility. Treat the feasibility scores as a draft you ignore until verified, and treat every cited title as unverified until you find it in Semantic Scholar or your library database. Nothing the model says about novelty is trustworthy; that is what a literature review and your advisor are for.
What to feed the AI
- Field and sub-area (e.g. Health Informatics, focused on EHR machine learning)
- Your advisor’s known interests and 2-3 recent paper titles
- Method preferences: qualitative, quantitative, mixed, or computational
- Hard constraints: time to deadline, data access, IRB/ethics feasibility, funding
- Your prior work and coursework, so it does not propose topics you have already touched
- Programme requirements: page count, defence format
Copy-ready prompt
Generate 15 thesis topic directions.
Field and sub-area: [area]
Advisor's interests (recent paper titles or themes): [list]
Method preferences: [quant / qual / mixed / computational]
Hard constraints: [time / data / IRB / funding]
My prior work or related coursework: [list]
Programme requirements: [page count, defence format]
For each direction, return a table row with:
- A one-sentence research question (specific, falsifiable, scoped)
- Why it matters (2 sentences: to whom, what changes if answered)
- Feasibility within my constraints, 1-5 (be honest)
- Main risks (data access, ethics, scope creep, novelty)
- Adjacency to advisor interests (1-5)
- A starting paper — name a real paper ONLY if you can verify it;
otherwise write [VERIFY: type of paper to find]
The 15 must span at least 3 sub-areas. If two are too similar,
replace one. Do not invent citations.Follow-up pass for your shortlist:
For directions #[x], #[y], #[z], write a half-page proposal each:
problem, hypothesis, method sketch, expected contribution, top 2 risks.How to check the output is usable
- Each research question is falsifiable and scoped: not “study X in Y context” but “does X change Y under condition Z?”
- Feasibility scores match your constraints honestly. A 5/5 with no data access is wrong.
- Every starting paper is real — search the exact title in Semantic Scholar before trusting it.
- The 15 are genuinely distinct, not minor variations of one idea.
- At least 3 clearly align with your advisor’s interests.
Common mistakes
- Trusting AI feasibility scores. Models overestimate almost everything; a “doable in 6 months” claim usually isn’t.
- Letting it invent citations. Even a confidently named paper with plausible authors is often fabricated. Verify every one.
- Skipping advisor adjacency. A brilliant topic your advisor cannot supervise is a dead topic.
- Picking the most exciting idea. Pick the most finishable one inside your timeline.
- Staying wide. Once you have 15, narrow fast — the goal is a shortlist, not a longer list.
FAQ
- When should I do this? 6-12 months before your proposal deadline for a thesis; later is fine for an undergrad project.
- Which model should I use? Any general model works for the brainstorm. Running it twice across two models (e.g. GPT-5.5 and Gemini 3.1 Pro) gives better variety because they have different reading biases. Spend a Deep Research run only on verifying your top 3.
- Do I still need Elicit or Semantic Scholar? Yes, for the verification step. Chat models cannot reliably tell you whether a question is already answered or whether a paper exists; that is exactly what those tools are built for, and both have a free tier.
- What about novelty? AI is bad at novelty assessment. A real literature search plus your advisor are the source of truth.
Related
- Study plan prompts: schedule the topic-search process
- Explain difficult concept: for tough background reading
- Study notes cleanup: clean notes from lit reviews
- AI paper reading workflow: read recent literature efficiently
- Claude long doc workflow: when papers are long
- Exam study plan: same scheduling pattern