Chapter 6: Research Planning and Hypothesis Development with AI
What You’ll Learn
How research planning connects to and builds on your literature review
Why AI works best as a collaborative thinking partner in the early planning phase
How to use AI conversation to sharpen a vague research interest into a testable question
How to stress-test your question for feasibility and novelty before committing
Where This Fits in the Research Process
If you think about the early stages of a research project as a problem framing phase, then literature review and research planning are two sides of the same coin. The previous chapter was about understanding the landscape: what has been done, what methods people use, where the gaps are. This chapter picks up where that leaves off. Now you have a sense of the terrain, and the question is what to do with it.
That transition is harder than it sounds. Most researchers can describe a general area they are curious about. Fewer have a research question that is specific enough to actually pursue — one that is testable, feasible, and meaningfully different from what already exists. The gap between those two states is what this chapter is about.
This phase is sometimes treated as if it happens naturally, like the question just arrives once you have read enough. In practice, it usually requires active work. You have to articulate half-formed ideas, push on your own assumptions, consider alternatives you might not have thought of, and gradually narrow from a broad interest to something you could actually write a methods section for. AI can be a genuine partner in that process.
AI as a Collaborative Thinking Partner
The most useful way to think about AI in the planning phase is as a collaborator you can think out loud with. That framing matters because it sets realistic expectations in both directions: AI is more than a search engine, but it is not a replacement for your domain expertise or your judgment about what is worth pursuing.
Here is one thing that consistently happens in these conversations. When you sit down and try to explain your research idea to an AI, you run into places where you cannot be precise. You thought you knew what you meant by “disease burden” or “patient adherence” or “information-seeking behavior,” and then you realize, mid-sentence, that you are actually using the term loosely and it could mean several different things. The act of articulating your idea to something that keeps asking “can you be more specific?” or “what would you need to measure for that?” surfaces vagueness you did not know was there. That is genuinely valuable, and it happens because you are being pushed to externalize thinking that usually stays implicit.
But that is only part of what happens. AI also brings something real to the conversation. It can suggest a framing you had not considered, connect your question to a methodology from a different field, or point out an assumption embedded in how you phrased the question. These are not just reflections of your own thinking back at you — they are contributions that you can accept, reject, or build from. Sometimes a suggestion completely misses the mark and tells you something useful about why your domain is different. Other times it is genuinely better than what you started with.
A useful analogy is working through a draft with a skilled editor. Sometimes the editor asks “what do you mean here?” and you realize you are not sure yet. Other times they suggest something that is simply sharper than what you had. Both things happen in the same conversation, and the polishing is collaborative. Your research question at the end of that process belongs to that back-and-forth, not to either party alone.
A Practical Workflow: From Vague Interest to Testable Question
The best way to understand this process is to see it in action. The following example walks through how a conversation with AI might develop, starting from a broad area of interest and ending somewhere more specific.
Suppose you have been working in chronic pain research and you are curious about how pain affects cognitive function in patients. That is a real interest, but it is not a research question yet. It is a topic.
A first prompt might look like this:
“I am interested in how chronic pain affects cognitive function. I work with clinical populations and have access to patient-reported outcomes and basic neuropsychological assessments. Can you help me think through what a research question in this space might look like?”
The AI might come back with several angles: the relationship between pain intensity and attention or working memory, the role of sleep disruption as a mediating variable, differences across pain conditions, or the effect of pain treatment on cognitive outcomes. It might also ask what you mean by cognitive function — are you thinking about specific domains like memory or attention, or overall cognitive load?
That last question is the kind of thing that moves you forward. You respond that you are most interested in attention, specifically in clinical settings where patients need to follow complex treatment instructions. Now the conversation has narrowed. You follow up:
“What would a feasible study design look like if I wanted to examine the effect of pain intensity on attention and treatment instruction retention in outpatient chronic pain patients?”
And so it continues. Each exchange makes the question more specific and more grounded in your actual situation: your data access, your population, your resources. The AI is not writing your research proposal. It is helping you think through something you are building.
A few practical notes on making this work well. Specific prompts produce better responses than vague ones. Telling the AI what you have access to — what kind of data, what measures, what patient population — helps it give you suggestions that are actually relevant. When it suggests something that does not fit your context, push back and explain why. That pushback often leads to a better alternative. And treat the output as a draft to react to, not a conclusion to accept.
Stress-Testing Your Question
Once you have a question that feels sharper, the next step is to pressure-test it before you invest heavily in design and data collection. This is another place where AI is useful, though the mode shifts. Instead of brainstorming and exploring, you are now asking the AI to play devil’s advocate.
Some useful prompts for this phase: “What are the main methodological concerns with this design?” pushes on whether your approach can actually answer the question you are asking. “What similar studies have been done, and how is this different?” helps you gauge novelty and make sure you have not missed something obvious in your literature review. “What are the most likely reasons this study might fail or produce uninterpretable results?” surfaces practical and conceptual risks early.
You should also test feasibility more directly. What sample size would you need, and can you realistically recruit it? What measures or instruments would you need, and do you have access to them? What does the analysis pipeline look like at a high level, and is that within your team’s capacity?
None of these conversations replace expert peer feedback, a statistician’s review, or a frank conversation with your mentor. They are preparation for those conversations. Going into a meeting with a more thoroughly examined question saves everyone time and usually leads to sharper feedback.
It is also worth knowing when to stop. AI conversations in this phase can become circular, especially if you are genuinely uncertain about your direction. If you find yourself going in circles, that is usually a signal to step back and talk to a person — a colleague, an advisor, someone with domain expertise who can cut through the ambiguity in ways an AI cannot.
Tools for Research Planning
The conversational AI tools that appear throughout this handbook, such as ChatGPT, Claude, Gemini, and UM-GPT, are useful starting points for the brainstorming and question-development work described above. But a few more specialized tools are worth knowing about, because they are designed specifically around research workflows and can add something that a general chatbot cannot.
Elicit (https://elicit.com) is built around research questions rather than general conversation. You give it a rough version of your question and it surfaces relevant papers, extracts key variables from those papers, and helps you see what has already been tested in the space. This makes it especially useful during the stress-testing phase, when you want to check whether your planned question has solid empirical precedent or is genuinely novel.
ResearchRabbit (https://www.researchrabbit.ai) and Connected Papers (https://www.connectedpapers.com) let you visualize clusters of related work as a graph, starting from one or two seed papers. Rather than giving you a list of citations, they show you how papers relate to each other, which research groups are working in the same space, and where your planned question sits in the landscape. These are particularly helpful when you suspect your question might have been asked before, just by different people working in a different field.
Consensus (https://consensus.app) synthesizes findings across a large body of literature and gives you a quick read on what the evidence says about a particular relationship or claim. It is not a replacement for a systematic literature review, but as a feasibility check it can quickly tell you whether an assumption underlying your research question has broad empirical support or is actively contested.
NotebookLM (https://notebooklm.google.com) works differently from all of these because it reasons specifically from documents you have uploaded rather than from the open web (Google LLC, 2024). If you have already assembled a reading pile of key papers, you can load them into NotebookLM and have a conversation about gaps, tensions, and open questions grounded specifically in that set of literature. This is a more controlled version of the thinking-out-loud workflow described earlier in this chapter, and it reduces the risk that the AI will draw on papers you have not actually read.
None of these tools replace your judgment or your domain expertise. Think of them as different lenses on the same landscape. Using more than one of them at different points in the planning process is often more useful than sticking with any single tool throughout.
A Note on Ownership and Judgment
One thing worth being explicit about: the intellectual ownership of a research question developed through this kind of process is yours. AI assisted in the articulation and refinement, but the question emerged from your knowledge of the field, your sense of what matters, and your judgment about what is feasible in your context. That is not a trivial contribution.
What AI cannot do is tell you whether a question is worth asking. That judgment depends on things it does not have access to: your read on the field’s priorities, your knowledge of what funding bodies are interested in, your sense of what your institution can support, and your own scientific values. Those remain yours throughout.
Connecting Forward
A well-developed research question makes everything that comes after significantly easier. When you sit down to write a grant proposal, a clear and defensible question is the foundation everything else builds on. When you design your study, the question dictates the appropriate methods. When you write your IRB application, a specific question makes the risk-benefit analysis much more tractable.
As your question takes shape, it is also worth thinking about feasibility from a compute perspective. If your plan involves training a large neural network or analyzing millions of records, the computational requirements should inform your timeline and resource planning. See Chapter 13 for guidance on choosing computing resources, and Chapter 31 if you need to understand GPU specs and compute terminology.
The planning phase often feels like it should be over faster than it is. Researchers sometimes rush through it because it does not feel like “real” work yet. But time spent getting the question right tends to pay back later in cleaner designs, more focused data collection, and cleaner papers. AI can help you spend that time more productively.
Try This
Pick a research area you are currently thinking about, something you are curious about but have not yet committed to a specific question. Then try the following sequence over a single working session.
Start by writing two or three sentences describing your interest, as if you were explaining it to a thoughtful colleague outside your field. Then take that description to an AI tool and ask it to help you develop several possible research questions from it. Do not filter yet — generate options first.
For each question it suggests, evaluate whether it is actually interesting to you, whether it is feasible given what you have access to, whether it is specific enough to study, and whether it builds on what you already know from the literature.
Pick the one or two that seem most promising and go deeper. Ask the AI to help you refine the wording, identify what you would need to measure, and surface any obvious methodological concerns. Then ask it what similar studies exist and what would make yours different.
By the end of the session you should have something considerably more specific than what you started with. The goal is not a finished research proposal — it is a question you could plausibly defend to a skeptical colleague. That is enough to move forward.