AI in Research
Russell uses AI for specific tasks such as coding open-ended responses, probing survey answers, searching existing datasets, and exploring early hypotheses. Experienced researchers still design the study, check the evidence, interpret the findings, and make the recommendation.
Research design
A large set of open-ended answers may need consistent coding. A brief survey response may need a follow-up question. Each is a specific task with source material and an output a researcher can check.
We assess where AI can help within the study, including the work needed to prepare, review, and correct its output. The plan should explain what AI will do, how the result will be evaluated, and which decisions still need evidence from participants.
Read about AI analysis and survey probingGenerated responses do not add participants to a study. Keep modeled material distinguishable from what people actually said or did.
The role of AI changes what the output means. Reviewing collected evidence, asking a participant a question, and generating a possible response require different checks.
What is in the responses or study data?
AI can help code open ends, locate passages, query tables, or draft a synthesis. Researchers return to the sources to verify the answer, check calculations, and examine material the summary may have missed.
What did the participant mean?
A survey can use AI to select a probe based on an answer and approved instructions. Researchers test for relevant, non-leading questions and review the original answer with the follow-up. The additional evidence comes from the participant.
Which ideas need testing with people?
AI personas and synthetic responses can help develop hypotheses using existing material. Researchers document the sources and assumptions. Plausible generated language remains exploratory; it is not another interview or a measured customer reaction.
How does an AI answer affect the experience?
Participants can try an AI tool as part of a study, then discuss the questions they asked, the answers received, and how they evaluated them. Researchers distinguish this prompted exercise from evidence of how often people use AI naturally.
Illustrative distinction: asking a participant to explain an answer collects more evidence from that person. Asking an AI persona the same question produces modeled material. The report should make that difference clear.
Agree the outputs around the task and the people who will review or use the findings. Each should identify its source material, the checks performed, and any unresolved questions.
Findings organized around the research question, with a clear route back to the evidence. The interpretation distinguishes participant accounts, analytical findings, and modeled material.
Illustrative question
Does the summary reflect the range of responses, including people whose experience differs?
Where coding is part of the assignment, a file connects assigned themes to the original responses or transcript passages. Source references support review and correction.
Illustrative question
When a response mentions more than one concern, does the coding preserve both?
An account of where AI entered the work, what it was asked to do, and how researchers evaluated the output. The level of detail should allow your team to assess the proposed use.
Illustrative question
What was checked directly, and which conclusions still depend on an untested assumption?
Useful candidates include coding open-ended responses, locating relevant passages in transcripts, querying an existing study, and preparing a draft synthesis for review. The task needs defined source material, clear instructions, and a way to check the result. We start with the research question rather than add AI to every stage. Study design, interpretation, and recommendations remain researcher responsibilities. A task that cannot be checked against suitable evidence may need a different approach, even if the generated answer sounds convincing.
Yes, when the source material is available in a usable form and its permitted uses support the proposed task. Russell's published guidance describes querying completed studies and searching a controlled library of research. We review what each source covers, its definitions, dates, and limits before combining conclusions. Returning to the original responses or tables remains essential. Asking a new question of old research does not create evidence the study never collected; gaps should remain visible rather than be filled with generated answers.
AI can select a follow-up based on a participant's answer and the research team's approved instructions. If someone says they like the product features, a probe might ask which features they mean. The aim is to clarify the respondent's own account. Researchers need to test the probe behavior, avoid leading or repetitive questions, and consider the added burden on participants. The original answer and follow-up should be reviewed together, with the use of adaptive probing documented in the study method.
Reliability depends on the task, source data, and evaluation, so a general claim of equal or better quality would be misleading. For coding and synthesis, checks can compare themes with original responses, inspect exceptions, and verify calculations or quotations against their sources. Researchers also need to check whether important differences between audiences have been lost. AI assistance does not correct a poorly recruited sample or an unclear question. The validation should address the use being proposed and the decisions the output will inform.
Russell treats simulated responses as exploratory material that can help develop hypotheses or identify questions for further research. An AI persona built from a segment may generate plausible language, but that response is not another interview with a member of the segment. Synthetic answers do not increase the number of people measured or establish population percentages. When a decision depends on actual customer reactions, those reactions need direct research. Modeled material should be labeled and kept distinguishable from participant responses throughout the analysis and report.
The proposed use needs a review of the specific tool, the data involved, and the permissions governing it before processing begins. That includes what information is necessary, who can access it, how long it is retained, and whether the provider may use inputs or outputs for training. Those terms should be confirmed for the actual service and configuration; they cannot be inferred from an AI label. If the arrangement does not meet the study's requirements, the task needs a different setup or method.
It may reduce time spent on a defined task, especially when there is a substantial volume of text to organize or search. The comparison also needs to account for preparation, instruction design, checking, and correction. For a small or particularly sensitive assignment, that work may outweigh the saving. We assess the proposed use within the study scope rather than promise a standard reduction in cost or turnaround. Faster processing is useful only if the resulting work is suitable for the decision.
Yes. The method and validation note identifies the role AI played, the source material used, and the checks researchers performed. Findings should remain traceable to the relevant responses, transcripts, or records, with modeled outputs and unresolved questions clearly identified. The recommendation also needs to distinguish what the evidence supports now from what requires further research. That gives your team a basis for reviewing the work and deciding how to use it, rather than relying on the fluency of an automated summary.
Tell us what data you have, what you want AI to help with, and how you plan to use the result.
Talk with a Researcher