Accessible only to conference ticket holders.
Log in Create account Buy conference recordings
For 90 days after a conference, only paid ticket holders can watch conference videos. After that, all Gold members have access.
If you have purchased recording access and cannot see the video, please contact support.
Human vs. machine: Testing AI’s ability to synthesize and analyze research
Summary
Nielsen Norman Group (NNG) has conducted and continues to conduct extensive research testing various large language model (LLM) tools designed for research synthesis and analysis. Our goal was to determine whether these AI-powered tools could meaningfully accelerate the work of experienced UX researchers. Through rigorous testing across multiple models and specialized research tools, we’ve found that while a few tools provide modest speed improvements for experienced researchers, none come close to replacing human expertise in research synthesis and analysis. The core problem is that these tools consistently exhibit critical flaws: they hallucinate findings, fail to identify meaningful patterns in qualitative data, cannot adequately consider nuanced research questions, and produce only superficial, high-level summaries of participant behavior. What makes this particularly dangerous is that these AI-generated outputs often have the veneer of legitimate research results—they look professional and sound plausible. However, closer inspection reveals significant gaps, inaccuracies, and missed insights that would mislead stakeholders and result in poor design decisions. The appearance of competence masks fundamental limitations that make these tools unreliable for serious research work. While we’ve found several places in the research process that can benefit from LLM usage, analysis and synthesis consistently falls short. In this talk, I can share the specific research we’re doing and explain what actually works.
Key Insights
-
•
AI tools frequently produce insight-shaped outputs but often lack the rigor and accuracy of trained human researchers.
-
•
AI moderators cannot currently assess user behavior beyond spoken words, missing key usability observations like failed or inefficient tasks.
-
•
Contextual elements such as environmental interruptions are critical in research but are invisible to AI tools.
-
•
Synthetic users generated by AI tend to produce overly positive, unrealistic feedback that can mislead product teams.
-
•
AI excels at finding semantic connections and grouping codes in large, already coded qualitative datasets quickly.
-
•
Meta-analysis of large repositories using AI can uncover recurring user themes, like change aversion, much faster than manual methods.
-
•
Integrating AI with organizational systems to pull in diverse data sources improves context but requires expert setup and is not yet simple.
-
•
AI’s context window limitations cause it to forget earlier input, affecting the accuracy of multi-turn interactions.
-
•
Even trained researchers must use AI outputs cautiously, vetting insights to maintain research quality.
-
•
Effective user research depends on human synthesis, collaboration, and contextual understanding, areas where AI currently fails.
Notable Quotes
"AI can generate insights, but it does not do them as well as a moderately trained human researcher."
"There is a world of difference between what a participant says and what they actually do, and AI misses that completely."
"AI tells you what you want to hear, which is dangerous if you’re making product decisions based on synthetic feedback."
"Our job as researchers is not making reports or interviewing users; it’s providing actionable, correct insights."
"AI tools are incentivized to produce final deliverables, but that’s an output, not the essence of research."
"AI is pretty good at finding semantic patterns among codes after human researchers have done the initial coding."
"Nobody is going to be satisfied by insight-shaped answers or high-level summaries masquerading as breakthroughs."
"AI cannot notice body language, tone, or environmental context during a research session."
"Using AI to scan large archives of research is a game changer for meta-analyses, even if it’s imperfect."
"Well-set-up AI systems pulling data from multiple company sources will have more context, but it’s still limited compared to human understanding."
Or choose a question:
More Videos
"If it’s not usable it’s not valuable."
Product and Design at Bloomberg: A 15-year Evolution
December 6, 2022
"Breaking up design research training into short exercises over time helps get people out of the building and engaged."
Louis Rosenfeld Jose Coronado Rachel Posman Guneet Singh Crystal YanThe Bigger Picture: A Panel Discussion
October 23, 2019
"I learned about gross pay equity between me and my white counterpart despite doing more work."
Lisanne NormanWhy I Left Research
March 27, 2023
"No one knows yet exactly what bringing truth to power means, and that’s the beauty of it."
Jemma AhmedTheme 2 Intro
March 26, 2024
"Design ops people are mutants coming at this profession from different directions."
Holly ColePanel Discussion: Growing People and Teams
November 8, 2018
"If you don’t feel connected and you’re standing alone, the challenge looks much bigger."
Marc Fonteijn Ru ButlerIncrease your confidence, influence, and impact (through a Professional Community)
December 3, 2024
"Future thinking helps us create future memories, essentially to hack our neurobiology and trick our brains into creating pathways to imagine possible futures."
Ilana LipsettAnticipating Risk, Regulating Tech: A Playbook for Ethical Technology Governance
December 10, 2021
"Humans tend to give too much authority to autonomous systems, which can lead to overtrust."
Helen ArmstrongAugment the Human. Interrogate the System.
June 7, 2023
"Burnout exists because we made rest a reward rather than a right."
Zariah CameronReDesigning Wellbeing for Equitable Care in the Workplace
September 23, 2024
Latest Books All books
Dig deeper with the Rosenbot
What strategies motivate cross-functional teams to input their research findings into a centralized repository?
What alternative decision-making structures support more equitable power sharing beyond majority voting?
Why is design education’s emphasis on critical making important, and how does it interrelate with critique and theory?