Can AI strengthen research without compromising quality?
Embracing generative AI while preserving the essential role of human expertise
User research is the study of human behavior, so human knowledge and understanding are crucial. Technology can assist, but it can’t replace the skilled, contextual, nuanced observation that makes research meaningful. Expertise is more than a checklist of tasks; it’s the judgment to know whether a discussion guide is biased, whether a theme is real, or whether a finding actually addresses a question. Without that experience, it’s impossible to evaluate what AI produces effectively; the risk is making product decisions based on confident-sounding but unreliable output.
The appeal of AI is its speed, and that acceleration is real. But trading accuracy for speed may lead to high confidence in bad data. To help researchers think through the possibilities and risks of generative AI tools in user research, we’ve broken down the research process to some of its key stages. Real research is messier than this, or any framework suggests, but the structure shows where AI helps and where it doesn’t.
Planning: Designing the scope of the study
Organizing UX studies requires consensus building—loops of meeting, conferring, listening, and suggesting—with stakeholders from design, engineering, and product, each with unique goals that are often in tension.
Large language models (LLMs) can help: They’re good at getting past blank screens, turning vague requests into draft study plans, and serving as thought partners during ideation. But researchers can’t hand off the hard-won, context-specific knowledge that makes a study work. An AI-generated plan still requires an expert’s eye to ensure it addresses business context and asks questions that reflect what users need, rather than what stakeholders assume.
Recruiting: Finding the right participants
Recruiting is time-consuming. It’s also where AI can make the most immediate impact. Tools can screen for specific participant criteria, search career sites for candidates by job title or role and deliver a shortlist with contact information and details about why a participant could be a good fit. Done well, this could cut days off the recruiting timeline and help identify fantastic participants that may not have otherwise been included. Potential participants would, of course, still need to complete screeners.
The use of AI tools for recruiting comes with one tempting, but troublesome, shortcut: skip recruiting altogether and “interview” AI-generated users (synthetic personas) instead. Both the cultural anthropologist and cognitive scientist in us recoil at the idea of code attempting to mimic the infinite complexities and contradictions of humans. People are not simulatable. Data not derived from real humans has the potential to be overexaggerated or even flat out wrong. The temptation is real but the moment we rely solely on machines as a stand in for real people, we risk building for the machines and not the people. We would warn researchers here to experiment if they must, but to remain skeptical.
Analysis: From summary to insight
Analysis is where AI has made the deepest inroads and where the risks are sharpest. Many tools now generate interview summaries and study-wide themes that are useful as starting points. But starting points are not conclusions: A summary can surface key moments; it cannot interpret them. A generated theme may point toward a pattern, but a researcher must still verify it, contextualize it, and connect it to the actual research questions. Furthermore, LLMs frequently go beyond the data to make statements about user behavior that are simply not in the data (in other words, they hallucinate data on our behalf). Treat AI output as a provocation or pathway toward further investigation and research consideration, not a finding—the researcher, not the tool, still has to decide what any of it means.
One of the most valuable things researchers do is build frameworks for understanding user behavior across multiple studies over time. That type of knowledge accumulates slowly through years of research and collaboration with colleagues. It requires business context, human judgment, and the sort of synthesis that only happens when researchers compare notes. This intellectual core of research can’t be ingested from a document set. It remains stubbornly human.
Socializing research: Sharing findings across an organization
LLMs are immensely helpful at tailoring findings for different audiences—translating the same insight for engineers versus C-suite, or polishing slide language. Generative tools are also increasingly handling deck design. Researchers will still need to decide how much of their voice they’re comfortable delegating to an LLM, and legal teams should weigh in on what data and findings can be shared with unvetted generative tools.
Stakeholders increasingly want fast access to research—and tools like Acrobat PDF Spaces now let teams query across multiple studies to pull relevant findings. This is both convenient and risky: Decontextualized findings, especially when mixed with hallucinated details, can mislead as easily as inform. Review AI-generated summaries of research findings before they circulate.
No prompt can replace hard-won research expertise
Researchers should continue to experiment, stay curious, and not be reflexively dismissive of generative AI.
By speeding up planning, analysis, synthesis, and communication in real ways, the technology is genuinely reshaping how UX research gets done. But knowing how to use the tools doesn’t confer expertise. Without domain knowledge, it’s impossible to reliably evaluate output quality or make sound product decisions.
AI outputs should be treated as leads, not conclusions. Learning to analyze human behavior—in ways that are rigorous, minimally biased, and attuned to both user needs and business context—takes years. It means making judgment calls in the moment and rethinking assumptions study after study; those are skills that cannot be offloaded to AI.