Day 1- Coexisting with AI: A practical guide for researchers to navigate tools, ethics, and integration

— Hi there, I’m here to talk about co-existing with AI

— To understand something about users in the past, you had to ask user researchers about it

— But now people are starting to ask ChatGPT, and researchers have legitimate reason for feeling uncomfortable about it

• We are accustomed to being authorities in room, and if that authoritative voice can come from AI, there is sense of taking away from researchers value

— Whether you believe that AI can effectively do your tasks, others believe it can, and it will change your role and how you are perceived

• But we have a lot of control though in how our value gets defined going forward

 

— So how to be effective researcher in world of AI?

• What is our role going forward and why do we matter?

— What does it matter with AI?

• Have role to adopt AI safely and effectively, but that depends on decisions made today

— Easy to get focused on short-term impacts, and time on UX lifecycle and focus on speed vs. quality in social media chatter

• Here there is much debate with little progress

— But missing bigger picture, in that how we use AI impacts our research relationships, and how we value business impact

— Let’s talk with business impact, and why we are worth the cost of our paychecks

• UXR has been in rough spot for many years, and we ae looking to prove value for more security and careful as to how to do it

— Market has been flooded with AI research tools, and we are often focused on research output like hours saved or studies run

— Model of research as production, and raw number of insights produced

• A particular lens from the industrial era

— Competing on volume is losing game, and we can’t compete with speed or quantity of insights with AI competitors like Deep Research

• Tools are capable of doing work for experts
• We are participating in commoditizing ourselves

— Most researchers want to drive meaningful change, despite emphasis on output

• Can’t put off figuring this out any longer, and anything other than studies run, should act and trace business impact of influence
  • De-commoditizing work and tying it to concrete outcomes

— But we are now on hook for bad recommendations though, and measuring impact is harder than measuring output

• Only way to do work as business partners, rather than tool jockeys

— Second impact is adoption of AI tools consistent with research values and what it means for us

• We have spent time telling people there is a right way to do research and the importance of real human stories

— Generative AI will have stakeholders look at us to practice what we preach

• Access is not sufficient reason to use it, learning best practices
• Error is costly

— LLMs and AI research tools confronted with dilemma, and if we are haphazard with AI adoption, it looks like our hand-wringing was just gate-keeping, and people will not trust our results

— Sense that for users we are their advocate, and as user voice in organizations, and we need to show if we can do this effectively over AI

• AI is inaccurate and more biased towards marginalized groups

— What other issues are introduced into UXR, and people will watch to see how tools are used

— Early stages of adoption are deciding the identity of UXR going forward, and determining if people see values of us and our research

• If wrong, UXR will be diminished, but if right, determine how used by skills we have— our authority and authenticity

— Research is the foundation for product strategy, and pressing need for integrating AI into research

— Two skills stand out for researchers

• Optimizing across insight methods
• Producing verifiably human data

— First, let’s talk about information optimization

• LLMs are mixed bag with user insights, but can expound confident sounding nonsense
  • When training to become UXR, learn strengths and weaknesses of different methods, and being great at tradeoffs and method of situation at hand. One more method to optimize for
• LLM is too new for broadly expanded use in social science, and we need to figure out what the use cases are

— Why is this?

• Lack of trustworthy authorities [we are either learning from marketers or our own trial and error]
  • Learning everything about survey design from SurveyMonkey has problems
  • Trial and error has less bias, but need deliberate training to do it well
• Getting good interviews requires specific skills, and LLMs require their own set of skills to master as well

— Prompting is a specialized skillset like interviewing, and it looks easy, but wrong wording in prompt, and you can get hard to recognize errors

• And many places where the errors can come in

— There is a baseline error rate for outputs in all AI tools used, with companies like Elicit, asking us to assume 90% of info is accurate

— AI mistakes are different from human mistakes, and signals for error distribution are not how it’s distributed, and citations are suspect

• Example of cherry picking phrases that look similar to generated text, that might not be relevant

— Even harder for stakeholders to figure out where AI output is helping/hurting decisions

— We are valuable because we have this gift for tradeoffs, and figuring out which tools to use when

• We need good info on strengths and weaknesses of tools

— AI strengths and weaknesses currently came from engineers, but they gave good material for us to reference

— We will eventually have AI for user research with agreed upon standards and best practices

• For the moment, listen to the engineers and scientists involved with AI

— Second we can also collect and make sense of data from verifiable users

• 1/3 of content of sites like Medium and Quora is now AI generated
  • AI is now creating training data and causing problems with real-time remote interviews

— Examples of AI content include deepfake interviews for university admissions, and eventually only thing people can trust may be an in-person conversation, and people will want that

— AI is not good at making novel connections, or dealing with novel populations, and differentiated insights come from human-to-human conversations

• If there is a product where differentiation matters, research will matter

— Given importance of helping orgs navigate AI tradeoffs and providing human insights, so what does UXR of future look like

• Unique strategic value of reasons outlined and enable more way to get insights into product decisions

— Research is separate from decision making, but what if AI enables us to drive decisions more directly

— Idea of AI powered triple-threats and giving skills in product development we don’t specialize in, and using AI to plan, ship, and code their own designs

— Entrepreneurial approach that allows fully autonomous PMs, and AI to solve problems that keep you from maximizing impact as researcher

— Using AI to solve problems researchers encounter themselves

• We often hear how hard it is to anticipate trends
• AI designed for trend analysis and signal application, and feeding with corporate memos and new articles

— If we have lack of peers to assist with research work, we can use tools like Dovetail can figure out video highlights

• Critiquing discussion guides, and not replacing work of trained researchers but error checking and quality

— AI prototyping tools and wizard to create prototypes based on prompts and rapid iterative testing

• Convincing design and engineering for what to build

— Helping tools to transform insights into product strategy and wireframes

— Problem of tracing business impact

• Helping understand stakeholder motivations, and broader company systems
  • Automated AI agents can help remove friction in this progress, and more reliable

— Unquestionably researcher in role, but it will look different

— We can be AI powered triple threat researchers, and extend AI specialization researchers on when to use AI tools versus traditional user research

• But this only happens with AI getting right today

 

— Authority going forward and helping navigate going forward

• All this stuff starts with decisions we make right now

— Best guidance of AI that amplifies what we already are.

• If just impact, will help us understand and impact

— Can choose not to use AI, and understand when to use it

• Choices will determine value of UXR going forward and choosing path that will align with our value as researchers, as we are not afraid to do hard work to get decisions right

Questions

1. Have we confirmed AI LLMs are good at trend analysis and summarizing type of work?
   1. Key question right now, and contentious question in community. Academic research, is that it’s so-so and depends on model and prompt
   2. AI tool very good at making you average in your field based on it’s training data
   3. Generally find AI output won’t be good as really great researcher, but better than what non-researcher would have produced
      1. Need to know where is it okay to use AI, and where to have the extra excellence.
2. Can you dig in more on what we mean by ‘triple-threat’ research role?
   1. If AI is good at bringing up level of novices, helps us have opportunity to become generalists, and people can produce prototypes and simple code of their own
      1. Not at same level as skilled expert, but it adds all of new skills
      2. Idea of T-shaped employee and one area of excellence
3. If you are UXR curious about exploring triple-threat role? How to do it?
   1. Leverage resources in slide with three engineers who talk about capabilities of AI at deeper level, and grasp how tools work from functional level
      1. Simon Willison
      2. Andrej Karpathy
      3. Mystery AI Hype Theater 3000