A topic gives you a place to start, not a question.
Try searching for “trust in AI.” You will find papers, surveys, product claims, and plenty of opinions. After reading them, you may know more about the topic and still not know what your own study should answer. Search finds material related to your words. It cannot decide what you need to learn.
Suppose a team asks you to study trust in a service that makes automated decisions. Do they need to know whether people understand a refusal, believe it is fair, or feel able to challenge a mistake? These are different questions. Choosing one tells you whose experience to examine and what evidence to look for.
Spend time on that choice before committing to a method. In his account of research practice, Jason Wei argues that choosing a problem can limit or expand a project’s contribution, however well the experiments are run.[1] A carefully executed study is more useful when it answers something people actually need to know.
Before searching for more material, decide what you need it to tell you.
Ask who needs the answer and what they could do with it.
A request such as “understand the market” or “explore user attitudes” rarely names the choice behind it. Ask the person requesting the work: what could you do differently after reading the findings? They might redesign a notice, delay a release, investigate a problem, or keep the current service. If the answer cannot affect any available option, clarify the purpose before collecting more information.
Not every study needs an immediate business decision. Research can change an explanation, test a theory, or reveal a question that a field has overlooked. The useful habit is to say what you hope to learn and why it matters. Published research agendas, such as the Anthropic Institute’s focus areas, make these priorities visible so others can examine them.[2]
Try completing this sentence: “This study will help [person or team] choose between [options] by finding out [unknown].” For the service team, it might mean deciding whether to rewrite the refusal notice or change the appeal process. If either option has already been ruled out, the study needs a different purpose or a more honest question.
Explain what you mean by words like “trust.”
You cannot measure “trust” simply by putting the word in a research question. You might ask people whether they trust a service, observe whether they use it, or check whether they understand its limitations. Each approach reveals something different. Someone may keep using a service because they have no alternative, not because they trust it.
A measurable sign used in place of a broader idea is often called a proxy. Choosing one affects what the study can find. Passi and Barocas’s field research shows how data-science teams shape their targets around available data, organizational priorities, and negotiation.[3] Questions about fairness can enter at this early stage, before anyone builds a model.
For each important term, write down what you mean, what you will observe, and what that observation cannot tell you. Also ask who could be disadvantaged if this measure becomes the official definition. In the service example, counting submitted appeals may miss people who wanted to appeal but could not find the instructions.
Using a service and trusting it are not necessarily the same thing.
Try a few different questions about the same problem.
Before polishing the first question, consider alternatives. How many people abandon an appeal? At which step do they stop? Does changing the notice help them finish? How do they describe the experience? Do delays happen after the appeal reaches the service? These questions concern the same process, but each could lead to a different explanation and a different kind of study.
AI can help you explore these alternatives by changing the people, setting, comparison, or possible explanation in a draft question. Treat its suggestions as candidates. Co-Scientist, for example, separates generating hypotheses from reviewing, comparing, and improving them.[4] Producing more possibilities is only the beginning of choosing a useful one.
As an exercise, ask for twelve candidate questions, then group them by what you would learn from the answers. The number is just a prompt to explore beyond your first idea. Remove versions that merely swap words. If two questions require different evidence, keep them separate rather than forcing both into one study.
Choose a question worth answering with the resources you have.
Compare the candidates using three questions. Would the answer matter? Can you obtain the evidence responsibly with the time, access, and methods available? Would different results help you choose between explanations or options? A question can be important but not yet feasible. An easy question can still tell you very little.
These are reasons to discuss, not a formula that produces an objective winner. An important question may justify a small preliminary study to see whether the required data can be collected. A modest question may be worth answering because a larger project depends on it. Write down the reason for the choice, not just a score.
For each candidate, note what it would help someone decide, which evidence would be hardest to obtain, what result would surprise you, and what you could investigate next. The team can then compare the actual studies it might run, rather than argue over which sentence sounds more impressive.
An illustrative question brief for an appeal process. Agree on these choices before selecting methods.
Write down what would change your mind.
Suppose you expect clearer instructions to help people finish an appeal. What result would make you reconsider? If people understand the new notice but still cannot complete the process, the main problem may lie elsewhere. Writing down that possibility before the study makes it harder to explain away an inconvenient result later.
Preregistration takes this further: researchers record their questions and analysis plans before seeing the results. It helps readers distinguish a prediction tested by the study from an explanation developed after seeing the data.[5][6] Even when a public registration is not suitable, keeping a dated plan and a record of changes can make that distinction clearer.
You can still explore unexpected findings. Say that they were unexpected and explain how they led to a new question or analysis. Do not present a pattern you discovered as something you predicted all along. Registered Reports apply a related principle to publication: reviewers assess the question and proposed methods before the results are known.[7]
Say whose experience and which setting you will study.
A short question cannot contain every detail, so keep a brief alongside it. Specify who or what you will study, where and when, what you will compare, and how you will define the outcome. These choices keep the work manageable and help readers see where the answer applies.
“Does AI improve research?” is too broad on its own. You could instead compare how many sources a particular team finds with and without AI during a fixed search period. Check how many sources survive review, and how much extra checking each approach requires. Now you have a task, a comparison, and outcomes you can observe.
Be clear about the people and evidence you leave out. English-language public records will not tell you everything people know about an issue. Interviews with current users will not establish how common an experience is among non-users. Stating those limits helps you identify who or what a later study should include.
Decide when to revisit the question.
A question that made sense at the start can stop being useful. You might lose access to records, discover that your measure misses the real problem, or learn that someone else has answered it. The decision the study was meant to support might also change. Time already spent does not make continuing the best choice.
Choose a point to review the plan before the work becomes expensive. Reconsider it if you cannot obtain the needed evidence, distinguish the competing explanations, or learn enough to justify the remaining effort. A published agenda can also evolve: the Anthropic Institute describes its research priorities as work that will develop as AI and its effects change.[2] Your own brief should allow for changes supported by what you learn.
If you pause or stop, leave a short record: the original question, what changed, what you learned, and which materials are still useful. Say what would make the work worth resuming. The aim is to use the remaining time well, not to keep every initial question alive.