😒 I feel immensely overwhelmed, my life is currently a jumble of work, courses, and side projects. And the reality of what Nigeria, my country (*whispers* unfortunately) may become in the next four years scares me.
For the past few months, Large Language Models (LLMs) like ChatGPT and Notion AI and how they might upend the tech industry's working class have dominated most conversations. Designers are wondering if the models will replace them. I am probably one of the few UXers that is not worried about the effects these models will have on my career. But I worry about their ethics.
I delayed this issue's publication because I wasn't sure of what it would be. More than an hour after brainstorming, I asked:
Me: What’s trending in the design world right now?
Notion AI: In the UX world, one of the biggest trends right now is the use of artificial intelligence (AI)...
Last year, my clairvoyant self wrote about how language models like LaMDA could be used to analyse big qual data and predict the context of users to build human-centred products.
Seven months later, we have products that might do what I envisioned: ChatGPT and Notion AI. I used both bots, if you will, to synthesise qualitative data.
About the Data
In 2020, I conducted a study to understand how 18-35 year-olds journaled, their journaling motivations, how journaling made them feel and affected their daily lives. I interviewed five participants, manually dissected the data with Microsoft word, synthesised it with Figjam, and created a persona.
I fed ChatGPT and Notion AI a small chunk of anonymised data and asked:
Me: (1) First, I want you to group these quotes into themes that are clear and actionable. Include the number of times a quote appears for each theme, who said the quote, and a snippet of the quote relating to the theme. Give me the output in a table. (quotes inserted) (2) Then I'd like you to write a summary of each theme. Let me know if you can find any relation between the themes. (3) Finally, using the information you generate, I want you to craft a persona and based on the information, tell me what this persona would like to see in a journaling app.
I had three sessions with both bots. I started with ChatGPT: after failing, in the first session, to make it learn from my data by prompting it step-by-step, I had to give it all the prompts and quotes at once. The same was so for Notion AI.
Here is what I discovered.
Small dataset
I used about 20% of the data I initially collected because:
The data set was too large for the LLMs to process.
The models are not transparent about what they do with the data.
Similar findings, skewed interpretation
The results were unsurprisingly similar even though they worked with 20% of the dataset. After analysing and synthesising the data, I came up with these themes:
Journaling Thematic Map
Journaling Rate: Some people journal on a whim, some have a set time to journal, and some find it hard to journal even though they want to.
Reasons for Journaling: Journaling helps people remember important experiences and express their emotions.
Rewards of journaling: Journaling about personal and emotional things relieve people and gives them clarity.
Journaling Tools: People want a flexible journal that allows them to pen down personal entries, create schedules, and jot down quick notes and ideas.
Unprotected Journal: People need a safe and accessible way to protect their journals because it is difficult to protect journals from snoopers and prevent data loss.
Between ChatGPT and Notion AI, the results vary slightly.
ChatGPT’s responses:
Response 1
Nostalgia
Pen and Paper
Journal Content
Emotional effect of journaling
Response 2
Preference for Hard Copy Journaling
Preference for Pen and Paper Journaling
Journaling to Remember Peculiar Moments
Journaling to Track Personal Growth
Journaling for Emotional Well-being
Journaling for Future Planning
Journaling Twice Daily
Using Multiple Journaling Tools
Response 3
Prefer hardcopy diary over digital
Journaling to document specific or peculiar events
Journaling to cope with ups and downs
Use of digital tools for convenience and security
No preference for hardcopy or digital
Notion AI:
Response 1
Hard Copy Journals
Loss of Things
Peculiar Occurrences
Ups and Downs
Writing About the Future
Writing on Phone vs. Pen and Paper
Apps Used for Journaling
Response 2
Hardcopy Journaling
Insecurity of Digital Journals
Writing About Peculiar Occurrences
Writing About Ups and Downs
Writing Spontaneously
Writing as a Form of Relief and Hope
Writing About the Future
Writing as a Tool for Accountability
Response 3
History of journaling
Security of hard copy
Peculiar Occurrences
Ups and Downs
Future Planning
Accountability
ChatGPT established connections between each theme.
Me: Can you find any relationship between the themes?
ChatGPT: Yes, there are some relationships between the themes that emerge from the quotes. First, there is a strong emphasis on the use of hard copy diaries or pen and paper for journaling. Several participants mention the emotional connection they have with physical journals, and how writing in them helps them to better recall and reflect on their experiences. However, digital journaling is also popular, especially for those who want to keep their writing private and secure. Some participants use both physical and digital journaling methods, depending on their needs and preferences. The quotes also suggest that many people journal as a way of processing their emotions and experiences, especially during difficult times. Participants describe how journaling helps them to feel relief, hope, or a sense of accomplishment. Finally, there is an element of intentionality to journaling. Many participants describe how they make a conscious effort to write about certain topics or experiences, whether to capture important memories, process difficult emotions, or plan for the future. Overall, the quotes suggest that journaling is a flexible and adaptive practice that can take many different forms, but that it often serves as a valuable tool for reflection, processing emotions, and personal growth.
For the most part, I am impressed. But I disagree with one interpretation here. When I conducted the study, I concluded that both digital and notebook journalers worried about the safety and retention of their journals. Even though one participant depended on digital notetaking apps for creating content, there was no affinity towards digital apps for privacy and security reasons.
Notion AI’s responses, in all three sessions, were generic:
Notion AI: The quotes discussed in the themes related to journaling and writing had a variety of topics, from the use of hard copy diaries and digital journals, to the security and potential loss of information associated with digital journals, to special and rare occurrences, writing about ups and downs, spontaneity of journaling, relief in expression and hope in anticipation, writing about future plans and goals, apps used for journaling and writing, and the ease of writing with pen and paper. While each of these topics have their own distinct elements, they all relate to the idea of journaling, and the different ways it can be done and experienced. This makes the persona of someone looking for a journaling app that offers a combination of features from both hard copy diaries and digital journals quite clear.
And as I pondered over its shallowness, I remembered this passage from Ian Bogot's article:
But you may find comfort in knowing that the bot’s output, while fluent and persuasive as text, is consistently uninteresting as prose. It’s formulaic in structure, style, and content. John Warner, the author of the book Why They Can’t Write, has been railing against the five-paragraph essay for years and wrote a Twitter thread about how ChatGPT reflects this rules-based, standardized form of writing: “Students were essentially trained to produce imitations of writing,” he tweeted. The AI can generate credible writing, but only because writing, and our expectations for it, has become so unaspiring.
I will not continue with the responses I got from Notion AI because they are generic, too superficial to process.
Turns out that LLMs also make assumptions
Oh wow! Who would have known…
Me: What insights can you derive from the data I shared with you?
ChatGPT: Based on the data you shared about your users' demographics, it appears that your target audience is primarily young adults aged 18-34, with a slightly higher percentage of female users. This information can be useful in guiding the design and marketing of your journaling app to appeal to this demographic. Additionally, the reasons for journaling cited by your users, such as stress relief and personal growth, can inform the features and benefits you emphasize in your app. For example, you may want to include prompts or exercises that are specifically geared towards stress relief or personal growth, as well as features that allow users to track their progress and set goals for themselves. It's also worth noting that many users in your survey expressed a desire for privacy and security when it comes to their journaling, which should be taken into consideration when designing the app's user interface and security features. Overall, by analyzing the data you provided, you can gain valuable insights into the needs and preferences of your target audience, which can help you design a more effective and appealing journaling app.
What survey? How did it know my sample was majorly female?
ChatGPT could not learn from previously generated responses which begs the question: was it only spawning answers from my data, or was it branching out of the scope I set for it?
Both.
ChatGPT automatically triangulates your data with data from external sources it cannot specifically cite. If you do not give it all the prompts — and your data — at once, it will depend on other sources. It took me multiple tries during the first session to discover this.
And this leads me to my next point.
It veers off course and dismisses context
We must not be quick to trust the answers it generates especially when we synthesise data because there is always a chance that it will digress and potentially feed you data that does not fit into the context of your users. "We're suckers for explanation," Adam Rogers says, even if the explanation comes from a chatbot that does not care to verify the information it spurts.
Explanations make us feel smart and in control. It bubbles us in certainty we might not have to make conclusions. If we use ChatGPT to synthesise data, we can get swayed by its answers and forget that it does not bother to verify its responses.
Here’s another instance:
Me: Give me a punchy summary of all the themes you found.
ChatGPT: Sure, here's a summary of the themes I found: Personal Development: The pursuit of self-improvement and personal growth through a variety of means, such as education, introspection, and goal-setting. Relationships: The complex web of connections we have with others, including romantic relationships, friendships, and family ties. Mental Health: The state of our emotional and psychological well-being, including issues such as anxiety, depression, and stress management. Physical Health: The state of our physical well-being, including fitness, nutrition, and healthcare. Creativity: The expression of imagination and originality through various art forms, including writing, music, and visual arts. Productivity: The ability to efficiently accomplish tasks and goals, often through the use of tools and strategies such as time management and organization…
Me (clearly anthropomorphizing now): No, I mean the journaling themes you found that relate to the quote [sic] I told you to analyse earlier.
ChatGPT: I apologize for the confusion. Based on the quote you provided earlier, "The only way to do great work is to love what you do", here are some potential journaling themes that may relate to this quote: Passion and purpose: Explore what you are truly passionate about and what gives your life a sense of purpose. Ask yourself what you would do if money was not an issue, and write about how you can align your work
I restarted the process.
Me: Let's try again. I want you to group these quotes into themes that are clear and actionable. Include the number of times a quote appears for each theme, who said the quote, and a snippet of the quote relating to the theme. Give me the output in a table. Here are the quotes: (inserts quotes)
ChatGPT: Sure, here's an attempt at grouping the quotes based on themes that are clear and actionable.
I hope this helps!
Me: Please give me a summary of each theme and tell me how they relate to other themes, if possible.
ChatGPT: I'm sorry, but I'm not sure which specific themes you're referring to. Could you please provide more information or context so that I can better understand your request and provide a more helpful response?
Me: All the themes that you listed out in the table above.
ChatGPT: Sure, here is a summary of each theme and how they relate to each other: Love and relationships: This theme explores the dynamics of love and relationships between individuals, including topics such as intimacy, trust, communication, and more. Love and relationships can intersect with many other themes, such as family, identity, and self-discovery. Family: This theme focuses on the relationships between family members, including the roles they play, the dynamics of family life, and the challenges that families face. Family can intersect with many other themes, such as love and relationships, identity, and personal growth. Identity: This theme explores the idea of personal identity and how it is shaped by various factors such as culture, society, family, and personal experiences. Identity can intersect with many other themes, such as self-discovery, relationships, and social issues. Self-discovery: This
Me: I mean the themes that you listed in the table above. The table with quotes and themes that pertain to journaling.
ChatGPT: My apologies for the confusion earlier. Here are summaries of the themes listed in the table above that pertain to journaling:
🤦🏽♀️🤦🏽♀️🤦🏽♀️
LLMs will hinder UX Researchers from creating the deep connections they need to understand their participants. Context is crucial to qualitative research. LLMs will disregard all forms of nuance. What will remain are bits of data rather than the rich information they need to inform design decisions.
Free service, but not transparent enough
Before I began, I scraped the data of all forms of Personal Identifiable Information (PII).
A few quotes from the data I collected and shared with ChatGPT and Notion AI
Unless you have a private LLM linked to your database, it isn't safe, at least for now, to use ChatGPT or Notion AI to analyse qualitative data. I am halfway through the Service Design course by the Interaction Design Foundation and I've discovered the chilling reality that free cloud-based services (GSuite, etc) are not free; we pay with our data and do not have full control of it. Regardless, these services make us feel in control to increase the value
we attribute to them.
Nothing is stopping the owners of the LLMs we have now from adopting this model.
Oh yes, bias
If in all that we do, humans and biases are inseparable, then AI models trained with human data are invariably biased. We can't pinpoint where the models retrieve their data, but whatever they collect will always be somewhat subjective. The creators and the models themselves are not transparent enough to inform us of the measures they have taken to reduce bias — I guess they are close to none. These bots, in all their disruptiveness and innovation, are — like humans, but less likely — unable to identify their biases. Garbage In, Garbage Out.
“Treat it like a toy, not a tool”
But I think the biggest issue is that we expect too much from these models. Because we anticipate too much from them, we expect truth and accuracy every time we use them. While they help to streamline search results, these models are anything but transparent.
The more I think of these language models, given my recent trial, the less I see how they would participate in in-depth qualitative research. They are good at summarising. But qualitative research needs more than summaries. I have used them to remind myself of some forgotten programming syntax, I have used them to brainstorm and gather inspiration. They are fun to play with and useful for small and mundane tasks.
But one aspect of qualitative research I have come to appreciate is that each participant is human. As you listen to them, you have to try as much as you can to understand their stories, needs and goals deeply. You are not just learning about the participant for your product's sake, you need to know who they are within the context of that product. And this function is as human as it gets. LLMs are good for brainstorming, but for them to truly participate in qualitative research, they need more context, and to retrieve and comprehend more context, they need to be as sentient as humans. We are yet to understand our sentience, what makes you think an LLM can get it right?
