AI Appreciation Day: Accelerating insight, not replacing it.
16th July 2026 •
16th July 2026 •
Using AI-generated personas and synthetic research activities can sharpen hypotheses, uncover blind spots, and make real user research more effective.
While redesigning a client’s support experience, I set out to answer a simple question: could we explore user needs and test our thinking before involving real users?
So, I asked ChatGPT to behave like the client’s users. Not just a few, but dozens of them. From different industries, with a variety of skill levels, frustrations, and goals. And then I asked those synthetic users to take part in research activities such as card sorting, tree testing, surveys and even focus groups.
The results were insightful – both when they mirrored human behaviour, and when they revealed an unexpected or unusual perspective. In either case, the experiment changed the way I think about AI’s role in user research.
The client’s support ecosystem is enormous, supporting millions of users across their software product portfolio. It covers everything from help with installing software and managing licences through to troubleshooting, tutorials and workflow guidance.
Like many large support ecosystems, it had evolved organically over time. As a result, users looking to solve a relatively simple problem, like activating a licence or recovering a crashed file, are often faced with a maze of categories and navigation paths.
The challenge wasn’t just to create a cleaner information architecture, but to understand how various types of users think about support. For example, do they look for “licensing” support, or do they think in terms of “I can’t access my software”? Do they separate troubleshooting from installation? And where would they expect to find learning content?
These are all classic UX research questions, and normally we would answer them using methods like card sorting, tree testing, usability testing and surveys.
But before we run any research study, there’s another question that always comes first: What should we test in the first place? And that’s where AI proved useful.
I started experimenting with ChatGPT and Copilot as tools for early-stage exploration. The first step was generating synthetic personas based on the client’s real user ecosystem, including people like:
Each persona was created with their own specific motivations, frustrations, workflows and support needs. From there, I started using them to simulate research activities. For example:
I asked AI to group the support topics into categories that would make sense to a particular type of user. These included topics like:
The interesting part wasn’t whether the categorisation was “correct”, it was seeing the mental models emerge. Time and time again, synthetic users grouped things around intent and outcomes, such as fixing a problem, installing software, accessing their account or learning something new.
I also used AI to simulate navigation behaviour. Prompting a synthetic user with a task like: “You need help fixing a licence activation issue. Which category would you click first?” quickly exposed areas where category naming on the site might create confusion or overlap.
One of the most interesting experiments involved simulating discussions between different types of users. For example: a BIM manager, a student user, and an enterprise IT admin discussing where they would expect to find installation support.
These conversations were surprisingly coherent at times, and surprisingly revealing. Not because they were “true”, but because they quickly uncovered different perspectives and assumptions.
The biggest advantage here was speed, allowing us to explore at far greater depth than budget typically allows. AI allowed us to rapidly explore different information architecture approaches, test alternative category naming, uncover edge cases and generate early research hypotheses far faster than starting with a blank Miro board.
We were able to stress-test ideas within minutes, and AI also proved useful for drafting survey questions, generating usability tasks, summarising patterns in large datasets and reframing problems from different user perspectives.
In many ways, it worked less like a research participant and more like a thinking partner.
However, I also discovered a number of important limitations. For example, sometimes the outputs were generic and unhelpful with AI frequently defaulting to corporate-sounding structures like ‘resources’, ‘knowledge base’, and ‘support centre’.
These are exactly the kinds of labels users often struggle with in the real world. More importantly, AI doesn’t reveal genuine human behaviour. It doesn’t hesitate, become frustrated, or misunderstand interfaces in unpredictable ways. As a result, it doesn’t produce the kinds of surprising moments that make real user research so valuable.
Ultimately, AI reflects from its training data and the prompts we have given it, which means it can easily reinforce existing assumptions or biases. Consequently, validation with real users remains essential.
And that’s exactly what we did. We still conducted real card sorting and usability testing with human users. And this is where things became interesting. Some patterns aligned closely with the synthetic outputs.
Human users consistently grouped support content around practical outcomes:
But real users also revealed important nuances AI missed entirely. One of the clearest insights from our real research was that users think in moments, not systems. They don’t think in terms of licensing workflows, subscription structures or account management architecture.
They think, “I need to fix something”, or “I want to learn how to do something”. That distinction fundamentally shaped the direction of the support information architecture we designed. And it’s exactly the kind of behavioural truth that only emerges through real user research.
Moving from experimentation to a formalised process has unlocked tangible benefits for our clients:
Faster time to insight – we can explore multiple directions in hours, not weeks, which accelerates the entire research cycle.
Reduced research cost – by refining hypotheses upfront, we avoid unnecessary testing and make every participant session count.
Higher quality outputs – research is more focused, more relevant, and more directly tied to business outcomes.
Broader perspective – synthetic users allow us to explore edge cases and niche behaviours that would be difficult to access early on.
Better stakeholder alignment – early simulations help teams visualise problems and align around priorities before formal research begins.
My biggest takeaway from this project is that AI works best as a research accelerator, not a research replacement. It’s incredibly useful for exploration at significant depth, generating hypotheses, drafting questions, stress-testing ideas, and rapid iteration.
But it’s no substitute for genuine human insight. Real research remains essential for uncovering:
The most interesting future probably isn’t AI versus research. It’s using both together.
Synthetic users can help us think faster. Real users help us understand what’s actually true. And combining the two creates some genuinely exciting possibilities for how UX teams work in the early stages of problem solving.
No. AI can’t replicate genuine human behaviour, emotions, or unpredictability. You can use it to explore ideas and generate hypotheses, but real user research is essential to validate findings and uncover deeper insights.
Synthetic users are AI-generated personas designed to simulate different audience types, each with distinct goals, behaviours, and frustrations. They can be used to model how different users might approach tasks or interpret information.
AI excels at speed and breadth. It helps teams to generate early hypotheses, stress-test ideas before real research begins and reframe problems from multiple user perspectives.
Treat it as a thinking partner, not a participant. Use AI to accelerate your early exploration, then validate everything with real users to ensure accuracy and depth.
At Torpedo, we’ve been exploring how AI can accelerate early-stage UX research, helping teams test ideas faster, challenge assumptions, and approach user studies with greater clarity.
But we also know where its limits are. Real insight still comes from real users.
If you’re looking to strike the right balance by using AI to move quickly while grounding every decision in human behaviour, we can help you design a research approach that does both.
Get in touch