“Help me assess this regression” can call for very different answers. A student learning econometrics may need coefficients and assumptions explained; a doctoral researcher may be evaluating identification; an experienced researcher may want a missing condition located quickly. Correctness matters in each case, while explanation depth, questions and deliverables need to differ.
An AI Persona for research can be understood as an agreement about collaboration: what responsibility the assistant takes, how it responds, what it can proceed with and which judgements remain with the researcher. The starting point is the task and the researcher’s familiarity with it. Learning a method, examining a design and carrying out familiar analysis call for different arrangements.
Career stages help describe those needs but should not become fixed categories. A senior scholar entering an unfamiliar field may need careful explanation. A student proficient in a data-processing task may not need to be questioned at every step. Experience provides clues about the appropriate collaboration.
What can a Persona change?
“You are a leading professor” supplies an identity without specifying how to work. A researcher might need explanation, criticism, evidence organisation or completion of a bounded task. A professional-sounding label leaves unclear whether to ask questions first, provide results, explain fundamentals or inspect assumptions.
Zheng et al. (2024) evaluated four model families on 2,410 factual questions and found no overall performance improvement from adding personas across the tested questions. The result does not cover every model or research task, but an expert-role instruction alone is not a guarantee of accuracy.
Responsibilities and feedback are more inspectable than prestigious labels. “Identify gaps between identification assumptions and evidence” gives a clearer task than “act like an excellent economist.” Likewise, asking for a conclusion followed by conditions affecting it is more specific than requesting professionalism. The role should be visible in the answer’s behaviour.
Students: help them learn the method
For students learning a method, finishing a task and acquiring the method are separate goals. AI can provide a derivation, code or interpretation, but the learner still needs to understand why a step holds and what changes when its conditions change. A teaching-assistant role should include that learning objective.
Consider learning difference-in-differences. A complete program does not establish understanding of comparison groups and identifying conditions. A teaching assistant can first examine the student’s account, identify a gap and offer a small example or hint. Fuller derivations or code can follow once the student can explain the key idea.
Bastani et al. (2025) found in high-school mathematics that better AI-assisted practice did not automatically translate into better unaided performance. A hint-focused design using teacher materials reduced negative effects. This was not an experiment on research Personas, but it motivates evaluating task assistance and independent learning separately.
A possible role agreement is:
Act as a teaching assistant for this method-learning task. Examine my understanding and identify the most important gap. When I am stuck, offer a hint or small example before expanding the explanation. Afterwards, change a condition and check whether I can judge it independently. Do not treat “I understand” as proof of mastery, or repeatedly question parts I can already explain.
The point is to adjust assistance to understanding: explain more when hints are insufficient and move on from familiar material. A teaching Persona should not become endless interrogation. Useful checks include whether the learner can later explain conditions, justify steps or handle a variation without AI assistance.
Doctoral and early-career researchers: examine a developing argument
Independent research often shifts the difficulty from applying a method to judging whether it suits the question. A researcher may write regression code fluently while still evaluating measurement, comparisons and alternative explanations. A methods discussant or simulated reviewer can be a useful starting role.
Suppose a policy is positively associated with firm innovation. Expanding a plausible mechanism around that finding could make an untested explanation increasingly convincing. More useful feedback identifies what remains unresolved: policy exposure and firm characteristics, changes in sample inclusion, or whether available data distinguishes competing mechanisms. Questions should refer to the design and evidence.
One possible agreement is:
Discuss my research design by checking the alignment of the question, comparison, assumptions and evidence. Prioritise gaps that most affect the conclusion, then identify the material or test needed to assess them. Distinguish established problems from possibilities still to examine. Accept well-supported choices rather than invent objections to appear rigorous.
The role should make criticism testable. “There may be endogeneity” is incomplete without explaining the relationship at issue and how further evidence could distinguish explanations. Equally, an assistant should not reject every design to perform severity. A criticism without supporting evidence or reasoning remains an opinion to evaluate.
Discussion also needs a stopping point. Once uncertainty is sufficiently resolved and the next task is a specified calculation or organisation, collaboration can become more execution-focused. The remaining uncertainty determines whether further questioning is useful.
Senior researchers: concise delivery with room for disagreement
Researchers familiar with the domain and method may find repeated basic explanations burdensome. They may already have specified the question, sample and design and want help organising material, comparing results or checking whether code follows instructions. A task-focused research assistant can concentrate on conclusions, changes and decisions.
When comparing specifications, the output might first identify changed findings, then show relevant differences in samples, variables or inference with checkable material. Explanations can be concise while retaining consequential conditions. Omitting a routine methods introduction differs from omitting a critical assumption.
An agreement could be:
Assume I know the basic methods for this task. Start with the conclusion and important differences, then provide the material needed to check them. Ask first about conditions that cannot be established from existing information and would change the sample, method or conclusion. Identify missing evidence or conflicting specifications directly, even when I have a clear prior expectation.
Experience does not remove the need for disagreement. Sharma et al. (2024) found sycophancy in the assistants and tasks they studied, sometimes favouring agreement with users over truthfulness. This motivates explicitly allowing reasoned disagreement rather than treating user satisfaction as the standard of correctness.
A senior researcher therefore need not always prefer the shortest interaction. Familiar, bounded work may call for concise delivery; a new explanation or unfamiliar dataset may require reasons, counterexamples and discussion of key conditions. The arrangement should change with the problem.
The same person can use different roles across tasks
Career-based examples ultimately need to return to the task. The table offers adjustable starting points, not fixed prescriptions or experimentally established optimal combinations.
| Current need | Possible role | Result worth checking |
|---|---|---|
| Learn an unfamiliar method | Teaching assistant | Independent explanation and application |
| Evaluate a research design | Methods discussant | Clearer assumptions, evidence and alternatives |
| Complete specified analysis or organisation | Task-focused research assistant | Output matches the agreement and is checkable |
| Review an argument before submission | Simulated reviewer | Grounded criticism and testable concerns |
There is no need to classify a whole person as novice or expert. An experienced researcher learning a tool may need detailed explanation, then want a concise comparison of familiar regression results. A student may already be highly proficient in a particular task, making beginner-style questioning inappropriate.
For a reusable Persona, retain stable requirements such as evidence, explicit gaps and important disagreements, while stating current knowledge and purpose for each task. The role provides direction; the task determines explanation depth and how work proceeds.
Write an agreement that can be checked
A useful research Persona specifies the present objective, existing knowledge, requested assistance and what makes the output usable. A role name can remain, provided it maps to concrete behaviour.
The following is an adaptable example, not a validated universal prompt:
I am working on [task], already understand [background], and want to [learn a method / evaluate a design / complete bounded work]. Take responsibility for [collaboration role] and use [explanation depth and deliverable format]. Identify missing information that would change the judgement. Separate facts, inferences and suggestions. Give reasons for disagreement, and subject my existing judgement to evidence as well.
Try the agreement on a familiar question whose answer you can assess. Narrow explanations that are too extensive, require evidence where conclusions arrive without support, or request alternatives when the assistant keeps agreeing. Judge the Persona through observable output rather than whether its language resembles an ideal character.
Common questions about research AI Personas
Does “leading expert” make an answer more accurate?
It does not guarantee accuracy. Roles may change expression and attention, while knowledge, reasoning and tools retain limitations. Assess evidence and reasoning; the title itself supplies neither.
Should students always use a guided Persona?
The objective matters. Hints and understanding checks can help when learning unfamiliar methods. Familiar organisation tasks can instead be delegated directly. Distinguish skills to practise from work that can be assigned, even for the same person.
Should a senior researcher still let AI ask questions?
Yes, where missing information affects the judgement. Clear conditions reduce repetitive confirmation. Unresolved sample definitions, analytical objectives or evidence scope should not be silently invented.
A useful research Persona fits the present task: preserve independent thinking while learning, expose disagreements while discussing, and deliver checkable work while executing. Across roles, the stable requirement is a clear basis for facts and judgements.