By now, you’re no doubt familiar with ChatGPT, an AI chatbot developed by OpenAI that has been taking the internet by storm since its launch in the fall of 2022. The sophisticated large language model uses natural language processing and machine learning algorithms to draw from a massive amount of text to generate conversation-like responses in seconds. It took no time at all for every industry to start asking: “How can we use this to be more efficient?” And UX has been no exception.
In the past few weeks, I have seen many blog posts and Twitter threads on using ChatGPT to expedite UX research – from smart suggestions to what I would argue are unethical ones. As a UX Researcher with a background in human-computer interaction, I definitely have some thoughts on the matter. In this post, I will highlight some ways that ChatGPT can augment and even streamline the research process while also discussing how to use AI ethically when conducting UX research.
Identifying the competitive landscape
Without a doubt, ChatGPT can be an excellent resource for getting a quick high-level overview of a topic. This can be helpful if you’re in the discovery phase of a study or if you’re working on a product in an industry, you’re less familiar with. For example, suppose I’ve been asked to conduct user research to test a new feature for a food delivery app. One of the first things I would often do is observe what competitors are doing. Next, I might ask ChatGPT what the most popular food delivery apps in my target market are.
Not only did ChatGPT provide me with the names of the most popular apps, but it also reported the market share for each. However, when I cross-referenced this with other sources, DoorDash has actually made considerable strides in the industry and now accounts for over half of the total market. So what’s the deal? As ChatGPT’s developers state:
Ah. So, we need to be careful about even fairly straightforward queries. Knowing that, I can still abstract some helpful information from this, such as who the industry leaders are. From here, I might go on to do a competitive analysis, maybe conducting a SWOT analysis or running a usability test on a competitor app and comparing performance against my client’s app.
Familiarizing yourself with KPIs or metrics for an industry
Another good use of ChatGPT is learning what key performance indicators and metrics for success are essential for a type of product or service. Continuing with my food delivery app example, I queried KPIs and received a few results:
Since these are standard, ChatGPT is easily able to pull this list. (I’ll get more into how that works below). Knowing the KPIs (and their respective metrics) gives me a good sense of what my client’s priorities might be and could influence what kinds of questions I ask a user or which features in the app we test for usability.
Calling attention to accessibility concerns
One use for ChatGPT I have not seen mentioned is identifying potential accessibility issues related to a product or service. So, sticking with my food delivery app example, I tested it out:
While the five types of accessibility concerns aren’t unique to food delivery apps by any means, some of the points feel especially relevant, such as the heavy reliance on images of food. These lists can be helpful if accessibility concerns aren’t necessarily top-of-mind. It can be enough to make you say, “Oh, right. I hadn’t thought about possibly needing an audio cue for visually impaired users!” But how should it work and in what contexts is it most valuable are questions to ask real users with these needs.
Where It Gets Sketchy
Experienced researchers know that user testing - from recruitment to running and analyzing sessions - can be time-consuming. So it’s only natural to want to streamline the process as much as possible. But there are only so many corners we can cut and truly call what we’re doing “user research.” For instance, I have seen people suggest using ChatGPT to generate use cases, likes/dislikes, personas, and even simulate user sessions using AI participants. And this is where I get squeamish as a researcher.
To understand my reasoning, it’s helpful to know how ChatGPT works. ChatGPT generates its responses by pulling existing text from all over the web, from books to blogs and everything in between. ChatGPT, and all chatbots based on large language models, do not need to understand your query in a literal sense. Rather, it strings together sentences word by word, based on the probability of that word appearing in that sentence and in that context. Have you ever known someone who seems to say what they think people want to hear, even if it’s not necessarily true? That’s ChatGPT.
But let’s see what happens when I ask ChatGPT to tell me what users like about UberEats and then ask what users like about Grubhub. Here’s what I got for each query:
As you can see, the results are very similar overall: a wide selection of restaurants, a user-friendly interface, reliable delivery, flexible payment options, and deals and promotions. It makes sense that there would be a lot of similarities, given these are both food delivery apps. ChatGPT detects Grubhub and UberEats are food delivery services and associates many words and phrases with that context.
That’s also why a lot of this information is overly vague. Both apps are reportedly “user-friendly,” but ChatGPT doesn’t tell us what specifically makes each app “intuitive.” I would want to ask users what specific features are helpful or not, or even test users’ ability to use each app. But if I delivered a report with findings this vague to a client, I sincerely doubt they would be pleased. There is no new knowledge coming from this information, and nothing actionable for designers to take away from it.
I also question where this information is coming from. Is it coming from online reviews of these apps? The app developer’s own website or press releases boasting about their features? Tweets written by bots? With ChatGPT, it could be any and all of the above. Of course, it’s possible to get some data based on actual user sessions if that data has been published online somewhere (which is complicated with NDAs). Even then, every study is unique, and I would not assume findings from one study carried over to another. Nevertheless, this list can help give us ideas for concepts that need to be validated through user testing.
On the note of sources… it was recently discovered that ChatGPT fabricates academic sources. One user on the data science subReddit explained that they used ChatGPT to get a list of references to start with for a literature review, but when they tried to look up the recommended articles, none could be found. Associate Professor of Economics at the University of Queensland, David Smerdon, provides a great breakdown of how ChatGPT incorrectly identified a made-up paper as “the most cited economics paper of all time.”
So, should I not use ChatGPT?
In its current state, ChatGPT does an okay-ish job of harvesting certain factual data that can easily be found online. (Although I’d argue we could find answers more quickly and reliably through a Google search). That said, it can be helpful for some basic secondary research and to give us a starting point to think through study design.
But I could never, in good faith, advise using ChatGPT - or any AI - for primary research. Part of why we conduct UX research is to learn what we don’t know about users’ attitudes, expectations, and behaviors. Our job is to uncover the unexpected, to find the “ah-ha” moments. But how will we ever make new discoveries if we rely on existing data? Moreover, in a world where most things are designed with able-bodied users as the assumed user, recycling existing data runs the risk of overshadowing the actual lived experiences of disabled users that we would otherwise get from user sessions.
I don’t know about you, but I’m all for keeping the “user” in user research.
At Key Lime Interactive, we pride ourselves on our human-first approach while always staying on the cutting edge of technology. If you share our mindset and want to get insights from a human, contact us to see how we can help with your next UX/CX project.