The questions about how to use AI in research aren't slowing down. If anything, they're getting sharper. Not "should we use it?" (most teams have moved past that). The harder questions are the ones that come after: How do you know when AI output is actually useful? How do you keep your research from going flat? When have you stopped doing the thinking yourself?

At IDEO, we don't have a fixed answer. The tools are changing fast, the landscape is shifting, and our perspective is evolving alongside it. What we do have is a point of view we keep coming back to: AI shouldn't do everything.

The goal isn't to hand your research process over to a model and collect the output. It's to figure out where AI genuinely expands what's possible and where human judgment, lived experience, and time in the field are still irreplaceable.

That's the lens Angela Kochoska brings to this conversation. As a Design Researcher and Data Scientist at IDEO and co-instructor of IDEO U's new course, Human-Centered Research with AI, Angela sits at an unusual intersection: she spent years building machine learning models for NASA and the European Space Agency before falling in love with design research. She knows how these models actually work. And she knows why the conversation your team has in a coffee shop after a field interview still produces insights no model can replicate.

After Hannah Rosenfeld's episode on IDEO’s framework for using AI in research, listener questions kept coming in, so we dedicated an entire episode to answering them. Below are highlights from the conversation. Listen to the full episode for more.

Listen on Apple, Spotify, YouTube, or wherever you get your podcasts.

Article Summary

• How do you validate AI-assisted synthesis and tell a real insight from an AI-generated pattern?

• How do you stop AI from producing themes that flatten nuance or overstate confidence in your research?

• Is it self-serving to say AI makes humans more important in research?

• How do you know when you've crossed into outsourcing your thinking to AI?

• What does good prompting actually look like in a research context?

• How do you move from AI-generated insights to meaningful human-centered decisions?

• What AI tools do researchers actually use?

• How do you protect the depth and specificity that makes your research valuable?

• My organization is skeptical of AI. How do I show its value for research?

How do you validate AI-assisted synthesis and tell a real insight from an AI-generated pattern?

The answer, Angela says, is already in the question. AI generates patterns. That's what it does. What we call an insight at IDEO is something different, and the distinction matters.

"By default, we don't consider whatever AI outputs an insight,” Angela clarified. “It's just a pattern that it found in the data. And then we use those patterns alongside our own thinking and work and experience of doing research in the field with people to figure out what is that meaty, juicy, interesting and actionable point that all of these patterns are pointing us to."

In practice, Angela takes all the notes and research data, runs a first round of AI-generated patterns, and then brings the team in to go through them critically, flagging what's genuinely interesting, what's too widely known to build on, and what's worth carrying forward into synthesis. The AI gets you to the starting line. The human judgment is what runs the race.

Insight: AI finds patterns. Humans make meaning. The two aren't interchangeable, and treating them as such is where research goes flat.

How do you stop AI from producing themes that flatten nuance or overstate confidence in your research?

You can't fully stop it, Angela says. Flattening toward the average is part of how these models are designed. But you can work with that tendency rather than against it.

"You can play around with different prompts to see if you get the kind of level of nuance that you want,” she said. “Because it is capable of nuance. You just need to handhold and guide AI to where you want it to go."

Context management matters too. Too little context and the model defaults to means. Too much and it starts losing the threads you actually care about. There's also one-shot or few-shot prompting—giving the model an example of the output format you want—though Angela notes that trades one kind of flattening for another. 

The two words she keeps coming back to: relational and emergent. Nothing about working with AI is fixed, and nothing happens in isolation from how you interact with it.

Insight: Nuance isn't something AI withholds. It's something you have to actively prompt for, with the right amount of context and the patience to iterate.

---

Ready to turn your research into clear direction and actionable insights? Learn how to bring AI into your research process as an exploration and analysis partner in our course Human-Centered Research with AI.

---

Is it self-serving to say AI makes humans more important in research?

Angela says that’s worth taking seriously rather than dismissing: "I wouldn't say that AI makes humans overall more important in the process, but it makes aspects of humans more important—as participants, as researchers, even as thought partners—because AI can't bring the lived experience of a human being."

The fear that AI will replace human researchers isn't irrational. We’re seeing it play out in other industries already. But the more useful frame, Angela argues, isn't replacement versus irreplaceability. It's difference. AI is a voice in the room with a specific perspective, and that's where it stops. It hasn't lived in Macedonia like she has, or grown up in India, or spent twenty years doing fieldwork in rural healthcare. That embodied knowledge isn't in the model. And in research, it's often exactly what you need.

Insight: The question isn't whether AI can do what humans do. It's what each one does differently and how to use both well.

How do you know when you've crossed into outsourcing your thinking to AI?

Angela has a few personal flags she watches for. The first is how much attention she's actually paying to the output—whether she's reading and internalizing it, or just skimming and asking for more.

"The ratio of AI text versus your own answer is something I always pay attention to,” she notes. “Once you get into that mode where AI is just vomiting at you and you're asking it to vomit more at you without really thinking about what you're getting, that's the sign."

A second signal: who's asking the questions and who's giving the long answers. If AI is driving the conversation and you're just saying yes, that's a flag. The clearest test of all, she says, is whether you could explain what you got to another person. If you stopped working with AI right now and had no idea what to do next, you've probably outsourced the thinking.

Insight: The work isn't in the prompting. It's in what you do with what comes back. If you can't explain it, you haven't internalized it.

What does good prompting actually look like in a research context?

There's no single right answer, Angela says. It depends on the data, the context, and the output you're after. But there's one mindset that applies across all of it.

"The whole point of being a researcher is asking good questions, asking the right questions, asking interesting questions,” she said. “So if you can use AI more in that capacity, you are becoming a better researcher through using AI and not losing research skills by letting AI do the thinking for you."

Good prompting in research isn't about getting better answers faster. It's about using AI to help you ask better questions. That shift keeps human judgment at the center of the process rather than on the sidelines.

Insight: The researcher's job is to ask good questions. AI is most useful when it helps you do that better, not when it answers them for you.

How do you move from AI-generated insights to meaningful human-centered decisions?

Angela points to a concept her colleague Will Notini developed called team intelligence, and it reframes what research output is actually for.

"The human-centeredness in innovation comes from the team intelligence that is developed when you're immersed in the field, in the problem space you're working on. It's a different thing when you're just looking at a screen and reading copious amounts of text about what is going on in the world versus actually being in the world and internalizing the learnings while you're in the process."

So much of the best synthesis at IDEO, she says, happens in a coffee shop after an interview, not in a report. The conversation in the room after fieldwork, where everyone's still buzzing from what they heard, is where the real insights form. 

Insight: Human-centered decisions happen in the room with the team, not in the output that gets referenced in a meeting.

What AI tools do researchers actually use?

Angela is upfront that this answer has a short shelf life. The field moves fast enough that what's true today may not be true in six months. That said, here's her current toolkit:

Claude is her go-to for general-purpose work. "I feel like I trust it, which is kind of a weird thing to say, but if I had to choose between all these available models, I feel like I trust Claude the most." She also uses Claude Code for prototyping, giving it access to her GitHub and VS Code environment.

Perplexity has become a regular tool for research, particularly because it selects the best smaller model for a given task rather than locking you into one approach. She's seen strong results from it for market research specifically.

NotebookLM is her knowledge management tool of choice. She loads sources, pulls papers, and uses it to get oriented in a new topic area. Her favorite feature is the mind map. It clusters content into sub-themes she can probe rather than flattening everything into a podcast or summary slide deck.

SciSpace rounds out the list for academic and scientific research, with a table-based output that shows sources, relevance, and citation status.

Insight: The tools matter less than the mindset. Pick what fits the task and expect it to change.

How do you protect the depth and specificity that makes your research valuable?

This is the question every field facing AI disruption is grappling with, Angela says. Her answer starts with an uncomfortable truth: AI will get you to average. Faster than you could get there yourself, but average is its goal.

"If you just stop there, you stop at average and you don't go beyond that,” she admits. “The depth and specificity come when we intentionally go through that friction of learning things ourselves."

She tells a story from a recent project where she spent two days trying to get Claude to solve a participant-grouping problem. Through many iterations, it just wasn’t getting it right. When she stepped back and thought through the logic herself, she solved it in one prompt. The friction of doing the thinking herself was what made the solution possible.

Insight: AI can get you to average quickly. Getting past average requires the work only you can do.

My organization is skeptical of AI. How do I show its value for research to a team that isn't on board?

Start by asking what you're actually trying to use it for, Angela says. If the answer is to outsource existing work that people already do well, the skepticism is probably justified. That's not a good use case, and it's the one that generates the most fear.

"Where I would think people get on board is when you can show how AI can expand and diversify what you can do in a research capacity,” she said.

The shift she's seen work at IDEO is when researchers start using AI to do things they couldn't do before, like vibe coding functional prototypes to bring into the field, rather than using it to do existing work faster. 

Insight: Skepticism fades when AI expands what's possible, not when it replaces what already works.

---

Explore More

Take the course 

Angela teaches Human-Centered Research with AI alongside Hannah Rosenfeld at IDEO U. It's a five-week cohort course where you'll learn how to bring AI into your research process intentionally, not just to go faster, but to do things you couldn't do before.

Listen to Hannah's episode 

If you haven't heard Hannah Rosenfeld's episode yet, it's a great place to start, and a strong companion to this one. 

Read Angela's paper 

Angela and two IDEO colleagues wrote a research paper called  "Encounters of Human and Artificial Intelligence in Innovation,” which goes deeper on the concept of team intelligence discussed in this episode.

Listen to more episodes 

Subscribe to the Creative Confidence Podcast to hear conversations with today's most thoughtful creative leaders and explore past episodes at ideou.com/podcast.

---

About the Speaker

Angela Kochoska 

Design Researcher and Data Scientist, IDEO 

LinkedIn

Angela Kochoska is a Design Researcher and Data Scientist at IDEO whose work sits at the intersection of rigorous analytical methods and human-centered design research. Before joining IDEO, she completed a PhD in astrophysics and built machine learning algorithms for sky surveys used by NASA and the European Space Agency. At IDEO, she brings that technical foundation to design research practice, helping teams figure out where AI meaningfully contributes to research work, and where human judgment is irreplaceable. Angela co-teaches Human-Centered Research with AI at IDEO U alongside Hannah Rosenfeld.

---