Research with Generative AI: From Brainstorming to the Next Question
Published:
Do not let an AI discussion end when the chat closes
Generative AI can quickly offer options and explanations when you are stuck. That can be useful, but it is easy to close the conversation without knowing what you learned, what you decided, or what to do next.
The value of brainstorming is not only receiving an answer. It is being able to put your thinking into words, check evidence, counterexamples, and blind spots, and move to the next investigation or conversation.
This article explains how to use discussions with generative AI as part of the small cycle of research. It focuses on practices that work across fields, not on my research topic or actual conversation logs. In this article, generative AI includes both chat-based tools and AI agents. I use the term AI agent only when discussing agent-specific work with files and tools.
What you will learn
- How to make the context needed for a discussion available to generative AI
- How to compare your own idea with an idea from generative AI
- How to turn a conversation into working notes, a provisional decision, and the next question
- How to connect generative AI discussions to conversations with other people and the next validation step
Make the needed context reachable before brainstorming
If you ask only, “What should I do?”, it is hard to tell how an option should be compared. Before a discussion, you do not need to write every item in a fixed template. What matters is making the context needed for the current question reachable to both you and generative AI.
At a minimum, state what you want to check and what you want the AI to do. Then make the relevant evidence, your current hypothesis or idea, and any constraints available as needed.
With an AI agent, you can specify the files that contain the facts it must read and the files that contain your current thinking. Links between records help both the agent and you follow the necessary background. You can also set rules for material that must not be read or sent outside the workspace.
With a chat-based tool, you can serve the same purpose by typing the relevant facts and ideas, attaching material, or pasting a short summary. You do not need to put everything into one message. Add information when it becomes relevant during the conversation.
Whichever tool you use, do not mix verified facts with hypotheses or ideas. An explanation or option from generative AI does not belong in your factual record until you have checked its source.
Ask generative AI to check, not only to confirm
If you explain your idea and ask only, “Is this a good option?”, generative AI can easily reinforce it. Ask it to do more specific work instead.
- Check for contradictions or jumps in your explanation.
- Give counterexamples or conditions where the idea would fail.
- Suggest other approaches or comparison conditions.
- Identify evidence that still needs checking.
- Separate verified facts from inferences and suggestions.
As explained in Part 8, Using Generative AI Without Rushing Decisions, a plausible option is not a reason to choose it. And as Part 2, Research Does Not Start with a Method explains, check whether an option satisfies the properties needed to answer the research question. Check its evidence in the original sources, and consider what each option gains and loses.
The same applies when an AI agent reads research material. You decide what it must read, what it must not read, and what must not be sent outside. The research foundation introduced in Part 7 makes it easier to trace the records and sources needed for that check.
Compare an AI-generated option with your own idea
Even when I already have an answer in mind, I sometimes ask generative AI for an option before showing it my own idea. If its answer points in the same direction, I can check whether I missed anything. If it differs, I have another option to compare.
I do not reject the AI-generated option simply because it differs from mine. I use the following sequence instead.
- Look for problems in the AI-generated option.
- If you find one, explain it and ask for a revised option.
- If you cannot find a clear problem and the revised option still differs from yours, share your own idea.
- Compare each option’s strengths, weaknesses, assumptions, and trade-offs.
- Check the basis for the comparison, then decide for yourself which option to adopt—or whether to choose a different one.
This does not let generative AI choose the answer. It also does not mean looking for flaws only to defend your own idea. Compare how well both options meet the properties needed to answer the question, and where each becomes weaker.
When asking for a comparison, do not ask only, “Which is better?” Share the question, the necessary properties, and the constraints. Ask the AI to separate what each option can satisfy, where it is weak, what evidence still needs checking, and what you would lose by not adopting it. If there is no clear choice, that itself becomes the next question to investigate.
Turn the conversation into four kinds of record
Saving a whole conversation is not enough when you need to find its useful parts later. After a discussion, separate its contents into four kinds of record.
| Record | What to keep | What to watch for |
|---|---|---|
| Verified facts | Information confirmed in original sources or experiment records, together with its source | Do not use an AI explanation alone as evidence |
| Working notes | Options compared, hypotheses, counterexamples, and reasons for rejecting an option | Do not mix them with verified facts |
| Provisional decision | The direction you will use for now, why, and what would make you revisit it | Make clear that it is not final |
| Next question | What to investigate, test, or ask next | State what you need to learn, not only a task name |
It can be helpful to ask generative AI to draft this classification. Give it the conversation and ask for candidates under the four headings. Do not accept the classification without checking it. Review what has been placed where, correct any mixture of facts and hypotheses, and explain the problem if you ask the AI to revise its draft.
With an AI agent, you can have it draft changes to record files and review the diff before accepting them. With a chat-based tool, make the same check before copying the classification into your own notes. The goal is not simply to classify information quickly. It is to keep a record that lets you trace evidence and decisions later.
Keeping rejected options and their reasons in working notes is especially useful. Months later, you can see why the option was not adopted, what changed in the discussion, and which evidence was still missing. Do not move an AI-generated option into a provisional decision until you have checked what supports it. If evidence is still missing, leave it as the next question instead.
Use AI discussions to prepare for conversations with people
Do not treat brainstorming with generative AI as the final step for an important decision. When you are very uncertain, tired enough that your view has narrowed, or considering a change in research direction, prioritize a conversation with an advisor, senior student, or peer.
The value of speaking with people is not simply that they will give a more correct answer than generative AI. Explaining your thinking requires you to state the question, assumptions, evidence, and what you do not yet know. That process can reveal what was still unclear to you.
People can also bring in a perspective that is less tied to your own idea or to the direction of an AI conversation. Generative AI can be asked to provide counterarguments and alternatives, but it can still agree too readily or stay within the assumptions it was given. A person may ask a question outside the current option: “Should you analyze this result more before moving to an application?” or “Is there prerequisite work you need to examine before starting this project?”
Do not accept a person’s view without checking it either. Compare it with the research question, verified facts, and necessary properties. But for high-impact choices such as changing the research direction, making a conclusion, or changing the scope, do not keep the decision inside a conversation with generative AI alone.
The AI conversation is still useful preparation. If you separate the options compared, verified facts, unresolved points, and questions you want to ask, you can align your assumptions with the other person more easily. After the conversation, record what the feedback addressed, how it changed your thinking, and what you will check next. Then begin the next small cycle.
Summary: Return the discussion to the next question
Generative AI brainstorming is not for receiving an answer and stopping there. Make the needed context reachable, ask for counterexamples, alternatives, and evidence, and compare an AI-generated option with your own idea. Check the trade-offs and decide for yourself.
Then turn the discussion into verified facts, working notes, a provisional decision, and the next question. This makes a generative AI conversation more than a temporary source of ideas: it becomes a record that moves research toward the next validation and conversation.
This series has covered research questions, literature reviews, experiments, records, and working with generative AI. When I gain further lessons that can be shared without revealing unpublished research, I will share them in a form that is useful in practice.