How will AI innovation disrupt qualitative research? Does it pose an existential threat to UX jobs?
A year after her last visit, award-winning researcher Mary Gray returned to share her answers to these questions (and more). Lucky us!
Mary talked about how AI can change a researcher’s workflow. She called for more transparency and clarity on AI’s limitations when conducting qualitative research. Finally, Mary looked forward to how qualitative research will shape the future of AI.
Watch the entire conversation.
Introducing AI into the Research Workflow
Marvin’s take on AI is simple — it’s here to help. Think of AI as your research assistant. The ultimate research sidekick.
Mary echoed these sentiments.
“No tool replaces us. We just open up new things that we add to what it is we’re doing,” she said.
She referenced a book titled ‘More Work for Mother’. The premise? Once you create tools to make your job easier, you’ll start to do more things as part of your job.
At the end of the day, AI is a tool that augments a researcher’s workflow. Mary urged us not to forget:
“It’s software. It’s pretty impressive software. (But) it is software.”
So how does Mary use AI?
She views it as a genre or pattern detector. AI helps Mary analyze data and gives her new ways to prompt her own reflection. She asks AI questions about her data.
“That’s pretty awesome because otherwise it’s a lot of find and replace or ‘Option + F’ /‘Control + F’,” she said.
Mary also spoke about the widespread adoption of Large Language Models (LLMs) into research workflows.
LLMs have a strong base in different languages. Researchers use LLMS to form a foundational understanding of data. They can use it as a search tool for common phrases and familiar formations. With AI, researchers can make even more robust interpretations.
Mary outlined how to think about integrating LLMs, or any AI technology into your workflow:
What can an LLM do that complements what I do? Where can AI perform something I don’t want to do?
“The craft of being able to have an idea of how to use a particular tool. That’s the signature of someone who’s got expertise,” she said.
Using AI in Qualitative Research
It’s easy to get caught up in the furor surrounding AI.
Mary cautioned that we must proceed carefully while using various AI technologies. Understanding their limitations is key.
Predictive Capabilities of AI
AI is fast becoming a predictive tool. Leveraging historical data, it can make predictions about future user behavior.
When it comes to using this feature for qualitative research, Mary errs on the side of caution. A major drawback of AI is that it represents a static representation of what’s already happened.
“We’re not really modeling human decision-making at all. We’re capturing what is the endpoint of a lot of back and forth among groups of people in a setting,” she said.
Mary argues that we’re only looking at the output of AI.
In qualitative research, it’s so much more important to understand how you got there. It’s about the journey. Not simply the end result.
Qualitative research aims to understand the underlying motivation or thinking behind every decision. This helps distinguish between two decisions that superficially look the same.
“To just focus on quantitative approaches is to lose the chance to think about what’s possible. To predict what’s going to happen is a very narrow ambition. In some ways, it’s short-sighted. You want the range of possibilities. What Bourdieu called the space of possibilities — that’s the bread and butter of qualitative research,” she said.
Sentiment Analysis in Qualitative Research
Sentiment analysis is an AI technology that’s gaining in popularity.
By classifying words as positive, negative or neutral, AI aims to capture how customers feel about products or experiences. Researchers gauge user attitudes from tickets, feedback and social media posts on a large scale.
What does Mary think about the use of sentiment analysis?
To answer some research questions, it’s fine. For most others, she’s not convinced.
Sentiment analysis can unearth the correlation between words. It can evaluate interactions and identify prevailing emotions. While it picks up on the loud notes, sentiment analysis misses the nuances. That boils down to human interpretation.
We express ourselves with words. The devil is in the details of the way we talk, and sentiment analysis can’t sniff those out.
“It’s too simplistic for the human experience,” Mary said.
Marvin CEO Prayag Narula says it best, “We love quantifying qualitative data.”
Mary gave us plenty of examples from the research field.
- Word clouds attempt to be a quantitative representation of how people feel. “My skepticism is because analysis works with the assumption ‘If I see quantity, i.e. more or less of this word, (then) I know something about how you feel,’” she said.
- Mary likened sentiment analysis to the Likert scale. An approach designed to apply quantitative methods to the field of psychology.
That’s not to say don’t use sentiment analysis. Be mindful of using the tool and questions it’ll help you answer.
Creating User Personas with AI
Mary weighed in on using AI to assemble qualitative findings to create user personas.
She acknowledged the helpfulness of crafting personas in user experience research. But she pointed out that if we swapped out the word “persona” for “stereotype” or “bias,” people would get upset about it.
She urges caution when using AI to generate a user persona. A downside of using machine learning to create tacit personas is that it clumps individuals into a “bag of attributes.” A collection of different demographics.
“That’s a very reductive way of looking at who we are socially,” she said. “Often socially imposed categories don’t give us much room to think. Is this a meaningful way of representing groups of people and personae?”
AI and the Human Experience
We’re at a very early stage with AI.
“Up to this point, we’ve scraped existing text and images. That’s a paltry representation of the human experience, existence and exchange,” Mary said.
A prime example? All machine learning models are developed in English. “Last time I looked, that wasn’t the only language that had a way of seeing the world,” she said.
Language can’t be watered down to a one-to-one translation. Mary worries about the subtleties of language that get lost in translation.
It’s vital to capture the nuances of the human experience.
Keeping Up with the Context
Qualitative research will always be an interpretive enterprise. Researchers want to get to the bottom of underlying human decisions. Rather than getting rid of people’s points of view, they seek to understand where they’re coming from. It’s a social and cultural science. Researchers seek out context.
How do I model social decision-making?
With more context.
Mary illustrated this with an example:
Two people take the same route to work. On the surface, the result is the same. But their reasoning may be very different. Context gives us competing logic.
Juxtapose this with the way AI models are created, which is the polar opposite.
Mary introduced us to the concept “the Banality of Scale” (Note: She has a book with the same title coming out soon!). She calls it “an obsessive attachment to disconnect something from a particular need, making it applicable and usable by everyone at any time.” Developers take a blanket approach to addressing user needs.
With the banality of scale, a researcher’s role resembles a fact-finding mission. Their questioning stops once they learn enough facts. Again, the opposite of what qualitative researchers do.
Our CEO Prayag and Mary joked about a question that they hear far too often:
Is this qualitative study statistically significant?
An oxymoronic question.
Addressing Bias in AI
Are we invested in inclusive and diverse experiences?
Part of embracing different perspectives is understanding bias.
Mary struggles with the narrative that AI is viewed as a neutral third party. AI itself is trained on biased datasets.
People completely new to the field might ask:
How do you eliminate bias?
“We have to accept that we can’t de-bias models,” Mary said. “That would depend on us being able to build a tool that’s completely divorced from the way we as individuals interpret the world. Social theory doesn’t support that conjecture,” she pointed out.
To illustrate inherent bias, Mary referenced another Marvin guest speaker’s research. Google’s Rida Qadri and her colleagues examined how text-to-image models represent South Asian cultures.
Rida and her team found that if you type in “Middle Eastern” or “Bangladeshi woman” in your prompt (for example), the system overcompensates. Results are churned out with typically Western perceptions of Asian cultures.
“We start overfitting when we have less data,” said Mary.
Rida’s example illustrates two flaws:
- Demographic Bias. Skewed social attributes are present in the underlying data. For people in the West, AI has been a boon for their productivity. “That is the problem. They are not most of the world. The global majority has not been a part of this process,” Mary said.
- AI Models Functionality. Models are trained with a limited worldview. Scrutinizing the way information is analyzed “inside the box.”
What can researchers do to understand the kind of biases present in AI models?
Building Transparency into AI
“There are problems we need to solve. These will only be solved once we understand what trained the model,” Mary said.
AI is largely self-regulated by developers (if at all). There are no agencies to conduct due diligence of an AI model. It’s also a ‘black box’ — no one knows the data included in a system or how it works.
Transparency is fundamental to understanding inherent bias in a system. The data that a model is trained on is fundamental to understanding what kind of bias one needs to look for.
To Mary, transparency isn’t only the data that goes into a model. It’s a running record of how people are using the model.
Researchers have to start coming up with ways to keep provenance of the data that they have and use.
“Transparency is…show me the receipts,” she said.
A Future of AI with Qualitative Research
Widespread AI adoption does not mean that the system is perfect. Far from it.
Mary urged developers to challenge typical methods of creating AI prevalent today. Just enough is not good enough.
She encouraged them to take a more qualitative approach to developing AI:
“To update it (systems), we’re going to have to really engage people. If we don’t, we’re headed for a world that is incredibly dull and alienating. Imagine something constantly telling you that you need to pick a different word because your word isn’t the right one,” she said.
Essential to mapping the universe of possibilities, is providing context.
“We’ve only just begun to see the value of holding onto context,” Mary said. “I think we’re about to discover just how interdependent our understanding and decision-making is.”
With this in mind, she gave us her prediction for the future iterations of AI:
“The future of AI is how it can start modeling and how we can point it towards modeling context and thinking about how much setting matters. It’s a very strong force on decisions people make. We don’t just make up what we want to do, we do it in relation to others,” she said.
Getting rich, qualitative contextual data is essential for the holistic development of future AI.
“What AI can offer is a way to track more patterns of context that we otherwise don’t take in. I hope it actually takes more time for us to do qualitative research,” she said.
Spending more time with participants helps researchers understand the inner workings of user decisions.
“You’re no longer just scraping Reddit. No offense to Reddit, but at some point, it can’t give you the nuance and it never will. Going fast and having quality are not on a collision course. You get higher quality and you go faster when you get higher quality. Measure twice, cut once,” she said.
Clearly, AI still has a long way to go. Qualitative research will only enhance AI’s capabilities as a more inclusive, thoughtful and far-reaching tool.
Mary is optimistic. “The boon we’re about to hit is (that) qualitative research hasn’t had its day yet,” she said.
