Why AI Is Bad at Research (and how to make it actually useful)
Summary
LLMs are everywhere, but when it comes to real research, they often fall short. Generic LLMs weren’t built for continuous research workflows, and product researchers quickly see the problem: the outputs are generic, lack full context, and struggle to connect multiple data sources. Instead of surfacing meaningful insights, they can amplify noise. In this session, Daniel will break down why AI often fails research teams and what’s missing. He’ll show how to make AI actually useful for continuous product research. Accelerating analysis, connecting insights across sources, and keeping researchers at the center, equipped with a powerful tool rather than replaced by one.
Key Insights
-
•
AI in research struggles with large datasets, often averaging results and missing subtle but important signals.
-
•
Curating and filtering datasets by removing irrelevant data improves AI research output quality.
-
•
Scoping research into focused projects or topics helps AI deliver more precise responses.
-
•
Asking one question at a time significantly enhances the quality of AI-generated answers.
-
•
Providing detailed contextual information (personas, company background, product details) to AI boosts specificity and nuance in responses.
-
•
AI hallucinations and trust issues necessitate human-in-the-loop processes to verify output quality and citations.
-
•
Iterative refinement of AI outputs, similar to app development, is critical for achieving polished research results.
-
•
Spot checking AI-generated citations can be an effective and efficient way to validate research quality.
-
•
Context passed as embedded knowledge rather than repeated in prompts yields better AI results.
-
•
Using multiple specialized AI agents to critique each other’s outputs can mitigate bias and improve research accuracy.
Notable Quotes
"AI has this strange weakness that when working with a large dataset, they often miss crucial, subtle findings."
"The larger the dataset you work with, the more costly it is to run a single operation on AI models."
"Whenever possible, you should be breaking down your work into specific research projects or topics."
"When you ask a question, try to ask one at a time so the model doesn't get lost."
"Context is everything — providing AI with a folder of your company’s knowledge makes responses more detailed and useful."
"Research with AI requires as much iteration and verification as building an app or prototype."
"AI-generated research reports should always be tied to real feedback that you can verify behind every sentence."
"There's no way to deny it: every industry needs to adapt to AI, but nobody really knows how yet."
"Human in the loop means constantly interacting with AI, documenting your thoughts and assuring quality."
"Some engineers build a council of agents that debate and generate responses, which can help with bias and accuracy."
Or choose a question:
More Videos
"There are no mistakes in jazz, just missed opportunities."
Jim KalbachJazz Improvisation as a Model for Team Collaboration
November 6, 2017
"Inclusion, collaboration, and iteration are the three pillars of how we work with authors to make a book."
Louis RosenfeldCoffee with Lou: Should You Write a (UX) Book?
March 7, 2024
"If you’re sitting there thinking I need to build out this project and nobody else is saying the same thing, it means you have to do it, period."
Catt Small Micah Bennett Brian Carr Jessica HarlleeWhat's Next for ICs: Exploring Staff and Principal Designer Roles
February 22, 2024
"I think we as UX researchers are uniquely positioned to forge bonds with data teams because of our storytelling and cross-functional work."
Marieke McCloskeyUser Science: Product Analytics & User Research
March 11, 2021
"User recordings are your most valuable asset but have become riskier due to biometric privacy laws."
Llewyn Paine[Demo] Deploying AI doppelgangers to de-identify user research recordings
June 5, 2024
"Designers have a hard time interpreting econometrics-driven quantitative research without bridging approaches."
Joshua NobleCasual Inference
October 6, 2023
"We’re selectively skeptical—skeptical about some things but not others based on what we want to believe."
Sara LogelYour Colleagues are Your Users Too
March 29, 2023
"If something makes you feel unwelcome, the code of conduct explains how to engage with staff to resolve issues."
Bria Alexander Louis RosenfeldWelcome
January 8, 2024
"When you design for the edges, you make things better, more fluid, more customizable for everyone."
Sam ProulxOnline Shopping: Designing an Accessible Experience
June 7, 2023