Thesis topics fail at two stages: too broad to finish, or too narrow to matter. Both leave you stuck three months in with no real question to answer. These 12 prompts force calibration before you commit — scope, novelty, methodology, and the conversation with your advisor. Replace every [bracketed placeholder] with your own details, then paste into a reasoning model. For a structured run through brainstorm, narrow, and pressure-test in one sitting, see the thesis topic brainstorm workflow.
TL;DR
- Brainstorm with a fast model (GPT-5.5 Instant, Gemini 3.1 Pro), but narrow and pressure-test with a reasoning model (GPT-5.5 Thinking, Claude Opus 4.7 with extended thinking) — scope calibration needs deliberation, not speed.
- Never let the chatbot invent the literature. Use a grounded tool — Semantic Scholar (free, 200M+ papers), Elicit, or Consensus — to pull real abstracts, then paste them into prompts 4 and 7.
- The best topic is finishable and defensible, not impressive-sounding. Prompts 3, 9, and 11 exist to kill topics early.
Best for
- Undergrad final project / capstone
- Masters thesis
- PhD proposal
- Independent research
Which AI to use for topic work (June 2026)
Topic selection is a judgment task, so model choice matters more than for routine drafting. As of June 2026:
| Step | Best tool | Why |
|---|---|---|
| Wide brainstorm (prompts 1, 6) | GPT-5.5 Instant or Gemini 3.1 Pro | Speed + breadth; you want 10 options fast |
| Narrowing + critique (prompts 2, 3, 5, 9, 11) | GPT-5.5 Thinking or Claude Opus 4.7 (extended thinking on) | Scope and novelty calls need real reasoning |
| Reading real abstracts (prompts 4, 7) | Semantic Scholar / Elicit / Consensus | A chatbot will hallucinate citations; these are grounded in indexed papers |
| Synthesizing your own PDFs (prompt 8) | NotebookLM | Answers cite the exact source line; won’t invent references |
Free-tier note: NotebookLM’s free plan covers 100 notebooks of up to 50 sources each and 50 chat queries per day, which is enough for one thesis. Semantic Scholar is fully free. Elicit and Consensus gate heavy use (systematic screening, full-text extraction) behind paid tiers.
1. Brainstorm from interests
My interests: [3-5 interests]. My field: [field]. My constraints: [timeline, resources, advisor expertise]. Brainstorm 10 candidate thesis topics. For each output: (a) 1-line topic, (b) 1-line why it's interesting beyond me, (c) 1-line biggest risk. No vague "AI in X" topics.2. Narrow a broad topic
My current topic is too broad: "[topic]". Suggest 5 narrower versions. For each: scope (population + intervention + outcome), research question, why it's defensible at [undergrad / masters / PhD] level. Reject any that still need further narrowing.3. Pressure-test feasibility
My thesis topic: [topic]. My timeline: [months]. My resources: [data access, funding, advisor support, lab equipment]. Pressure-test on three axes: feasibility, scope, novelty. Output the single most likely failure mode and one mitigation.4. Find adjacent gaps
Pull these abstracts from Semantic Scholar or Elicit first — do not let the model invent them.
I want to work in [field]. Below are 10 recent paper abstracts. Identify 3 underexplored gaps adjacent to these papers — not the same, not a trivial extension. For each gap: why no one has done it, what new method or dataset would unlock it.
[paste real abstracts]5. Defend novelty
My thesis: [topic + central claim]. Articulate the novelty in three buckets: genuinely new, refinement of existing work, overlap with prior work. Then write 3 reviewer concerns I'd face and a 2-line response to each.6. Topic from a personal problem
I noticed this problem in my own life or work: [observation]. Could this be a thesis topic at [level]? If yes, give 2 candidate research angles, each with a concrete method. If no, name the academic gap or methodological reason it fails.7. Reverse from advisor’s interests
My advisor's recent papers (real abstracts pasted below). Suggest 4 candidate thesis topics that align with their interests but extend in new directions. For each: which paper it builds from, what's the new direction, why my advisor would say yes.
[paste real abstracts]8. Topic-to-proposal scaffolding
Run this in NotebookLM with your reading loaded so the prior-work section cites real sources, not invented ones.
My topic: [topic]. Scaffold a 2-page proposal with sections: problem statement (1 paragraph), prior work (3 key references), my contribution (3 bullets), methodology (concrete steps), timeline (by month), expected outcomes (artifact + finding). Mark anything I haven't decided yet.9. Methodology fit check
My research question: [question]. Candidate methodologies: [qualitative interview, quantitative survey, computational modeling, archival analysis, etc.]. For each: what it would let me claim, what it would NOT let me claim, time and resource cost. Recommend one and explain why.10. Topic cross-disciplinary stress test
My topic: [topic] sits between [field A] and [field B]. From [field A]'s perspective: is this a real contribution or a recycled idea? From [field B]'s: same question. Which department's committee would defend this most aggressively?11. Pre-mortem the thesis
Imagine it's 18 months from now and my thesis on [topic] failed. Write 5 plausible failure stories: data didn't exist, advisor moved on, scope collapsed, method didn't work, finding was null. For each: what early signal I'd ignore, what I'd do at month 3 to prevent it.12. Elevator pitch to a non-specialist
Pitch my thesis on [topic] in 60 seconds to a smart non-specialist. 3 sentences: (1) the problem in their language, (2) what I'm doing that's new, (3) what changes if I'm right. No jargon. If a sentence needs a definition, rewrite it.Common mistakes
- Too-broad topics (“AI in healthcare”) with no population, intervention, or outcome.
- Methodology that doesn’t fit the question (qualitative claim, quantitative method).
- No conversation with advisor before committing — and no plan B if they say no.
- Confusing topic novelty with a literature gap. No one studied X often means X isn’t worth studying; verify the gap with Semantic Scholar’s citation graph before claiming it.
- Letting the chatbot generate references for prompts 4, 7, and 8. Models still fabricate plausible-looking citations; ground every reference in a real paper you pulled yourself.
- Optimizing for impressive-sounding instead of finishable.
FAQ
Which model should I use to pick a thesis topic? Brainstorm wide with a fast model (GPT-5.5 Instant or Gemini 3.1 Pro) to get options on the table, then switch to a reasoning model — GPT-5.5 Thinking or Claude Opus 4.7 with extended thinking — for the narrowing, novelty, and feasibility prompts. Those calls are judgment-heavy and benefit from the model spending more time.
Can AI write my whole proposal? Use it to scaffold structure and stress-test logic, not to invent content. The methodology and prior-work sections must reflect real papers and your actual constraints. Prompt 8 is a skeleton you fill, not a finished draft, and anything AI-written needs your committee’s standards applied on top.
How do I stop the AI from making up citations? Don’t ask a general chatbot for references. Pull real abstracts from a grounded tool — Semantic Scholar (free), Elicit, or Consensus — and paste them in for prompts 4 and 7. For your own PDFs, use NotebookLM, which cites the exact source line and refuses to answer outside your loaded documents.
Is using AI for topic brainstorming considered academic misconduct? Brainstorming and stress-testing ideas is generally fine, but policies vary by institution, and disclosure rules differ. Check your program’s AI-use guidelines, and treat AI as a sparring partner for your own thinking, not a ghostwriter. The defensibility of the topic still has to be yours.
What if every AI-suggested topic feels generic? That usually means your prompt was generic. Feed it your actual constraints — specific datasets you can access, your advisor’s real papers (prompt 7), a problem you personally observed (prompt 6). The more specific the input, the less the output reads like a list of buzzwords.