Human-AI Interaction
UT Austin
UX Research
AI ideation tools are built to feel effortless, and people stop thinking critically when they use them. We built an ideation tool with two deliberate kinds of friction, then ran a within-subjects study to test whether friction brings back deliberate thinking and a sense of ownership over ideas.
Overview
Ask a chatbot for campaign ideas and you get five polished ones before you have finished reading your own brief. It feels like progress, and for work you already know how to do, it is.
The ideas that lose to filing friction are the ones nobody else has.
When the output arrives finished there is nothing to push against. People take the first good answer, agree with the model within a turn or two, and feel little ownership of what they end up with.
Dual-process theory explains why. Slow, effortful System 2 thinking switches on when fast System 1 thinking hits a problem it cannot resolve. Fluent AI output never creates that problem, so the person never has to think.
My role
I designed and built the interface and every canvas interaction, set up the design system, ran the participant interviews and facilitated the user testing sessions. Ayesha Khan and Pranati Kompella led the prompt engineering and the experiment design. INF385T, UT Austin.
Research
Questions
1
How do generative labour and an antagonistic conversation style, used as friction mechanisms, affect System 1 and System 2 thinking during ideation?
Measured by: retrospective interviews coded against a dual-process codebook by two independent coders, with time per turn as a behavioural signal.
2
How do these two mechanisms affect psychological ownership of the ideas produced?
Measured by: 1–7 self-ratings of ownership and perceived AI contribution, with satisfaction and confidence as checks that friction did not just make people dislike the result.
Research
Deliberate friction has its own literature: desirable difficulties in learning science, and positive friction and frictional AI in interaction design. We pulled seven families of friction from it, then added five concepts of our own.
Two research questions
RQ1: Does generative labour or an antagonistic conversation style shift people from fast System 1 thinking to slow System 2 thinking? RQ2: Does it change how much people own the idea, and how much credit they give the AI?
We sorted every idea by two questions. Does it change what the person has to do, rather than how the screen looks? Can three students build and test it properly in one semester? Only two families passed both.
Illustrative reconstruction of the sort, rebuilt from our real selection criteria.
Five participants, aged 22 to 30, each used all three modes on a different marketing brief, drawn from fifteen matched briefs. A Latin square rotated the order. Every session was capped at six turns, screen and audio recorded, and timed turn by turn.
Whichever mode someone tries first gets their freshest attention, so the order rotated. Generative Labour has a fixed length, so every mode was capped at six turns, though people could stop earlier.
After each task I replayed the recording with the participant and asked what they were thinking at every turn. Two coders tagged each note as System 1, System 2 or cross-cutting, using a codebook built from dual-process theory, and agreed with a Cohen’s κ of 0.70 on the top-level codes.
We chose a replay interview over think-aloud on purpose. Talking through a task while doing it is itself slow, deliberate thinking, and would have pushed every mode toward System 2, hiding the very difference we wanted to measure.
People also rated ownership, AI contribution, satisfaction and confidence from 1 to 7, and completed the 52-item Metacognitive Awareness Inventory as a baseline, so a naturally reflective person would not read as a friction success.
Looking back for this case study, I re-sorted the study’s timing logs, ratings, interview notes, pilot comments and instructor feedback into six patterns. Every note on the map comes from the real data.
Key decisions
Every decision traded something away. These six shaped both the tool and the study, and each one came from something we saw or measured.
Instead of letting people connect anything to anything, Generative Labour walks six turns from abstract to concrete: clarify, vibe, medium, seeds, mechanism, finalise. The open canvas made the effort feel arbitrary. The cost is rigidity, since a confident user walks the same six steps as everyone else.
Instead of role-playing a marketer, then a customer, then a sceptic, the antagonist is one critic who pushes harder every turn. A rotating cast gives three shallow objections. One voice can follow a weak answer all the way down.
Instead of attacking from turn one, the critic waits until the person has picked an idea. The prompt swaps underneath them at turn three, so every challenge lands on something they already claimed. The cost is two friendly turns followed by a small ambush.
Instead of letting the model tidy the canvas, anything a person makes gets an id the model is forbidden to touch. It is enforced twice, in the prompt and in the code that reads each reply. A study about ownership means nothing if the model can rewrite what the person made.
Instead of the original plan of fifteen people each trying one mode, five people tried all three. Each person becomes their own control, and a Latin square cancels out order effects. The cost is that every order runs exactly once, with no room for a failed session.
BEFORE · THE TESTED BUILD
Rounded, navy, and every node a coloured box with a title. To know a box was a channel you had to remember what purple meant.
AFTER · THE REBUILD
Every node now says what it is. Colour backs the label up instead of carrying it alone, and a neutral square shell leaves the canvas as the only colour on screen.
Solution
You describe a brief, it hands you finished campaign ideas, you pick one. No canvas, no resistance. This is the control, and it is what every tool on the market already does.
The model puts rough fragments on a canvas and stops. You cannot advance a turn until you have edited them, added detail and drawn the connections yourself.
It is helpful for two turns. Once you commit to an idea it turns critic, acknowledges without praise, and pushes on your weakest answer, harder every turn.
Same local model, same brief format, same turn cap. The only thing that changes between modes is how much work the interface asks of you. Here is the solution in four parts.
The model puts rough fragments on a canvas and stops. You cannot move to the next turn until you have selected, edited or connected them yourself. On turn five the model writes nothing: you write the trigger, action and result.
A real session in the running app. Drag nodes, rename what the model suggested, add your own constraint, draw a labelled link, then write the mechanism yourself. Model thinking time is sped up.
Two calm turns to explore and restate your idea. Then it turns critic: it acknowledges without praise, names the people who would object, and pushes on your weakest answer.
Watch the progress bar move from Understand to Challenge, and the tone change with it.
Anything you add, rename or connect on the canvas is read by the model on every turn, but never rewritten.
Too many escape routes and the friction stops working. Too few and people quit. Undo is always there. Regenerate works on turns one to four, while ideas are still cheap. Go back restores the canvas to turn four. Get hints exists on turn five only, the one turn where people stalled.
Prompt x Interface
Most of this tool’s behaviour lives in prompts, not pixels. Ayesha and Pranati wrote and tuned the prompts; I conceptualised the interface in Figma and built it with Claude Code, so every interaction had to survive contact with what the model would actually do.
Everything runs on a laptop through Ollama, so no participant data leaves the machine. Each turn has its own prompt, and every reply comes back as strict JSON so the canvas can place nodes without guessing.
The full canvas state goes to the model with every message. It returns only what it is adding and never re-lists the rest. That one rule is what keeps six turns of someone’s work from being quietly rewritten.
Results
In the baseline, 79% of interview notes were coded System 1, and four of five people stopped before the session ended because there was “no more to discuss”. In both friction modes the split evened out, and cross-cutting moments, where a fast thought gets caught and reconsidered, appeared only there.
Generative Labour kept people in the session at the baseline’s pace, 1:42 per turn against 1:37. The antagonist slowed every exchange to 2:26. One keeps you in the room. The other slows you down inside it.
Ownership: clear for one mode, mixed for the other
Generative Labour raised ownership for 4 of 5 people, from a mean of 3.4 to 5.2, and the credit people gave the AI fell from 5.8 to 3.8. The antagonist was mixed: two people rated ownership higher, one the same and two lower.
P4 scored 26 of 52 on metacognitive awareness, the lowest, with planning at 1 of 7. P4 also rated both friction modes lowest and spent the longest in them. Both modes ask for planning up front. It is one person, read after the fact, so I treat it as a question for the next study.
Five participants and no significance test. Read these as directions, not proof.
Reflection
Three principles I would carry into any AI product.
Effort has to feel like it is going somewhere.
An open canvas felt arbitrary. Six turns that build on each other felt like progress.
The prompt is part of the interface.
The rule that protects your nodes lives in a string. I now review prompts the way I review a spec.
Design the exit with the friction.
Resistance without relief just makes people leave.
Try it yourself. The tool is open source and runs on your laptop.
