Working prototype,
18-page report
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.
AI ideation tools compete on removing effort but when the thing you are making is an idea, the effort is the work. People take the first polished answer and stop thinking and people feel little ownership of the result.
Designed the interface and the turn structure for an ideation tool with two kinds of deliberate friction built in. One makes you assemble the idea yourself. The other argues with you. Then we tested both against a plain chatbot.
Fast, automatic thinking filled 62 to 79% of the plain chatbot sessions, and 33 to 47% under either friction mode. Four of five people quit the plain chatbot before it was finished. Only one of the two friction mechanisms also made people feel the idea was theirs.
The Problem
The ideas that lose to filing friction are the ones nobody else has.
THE GAIN
More ideas, and more variety
Holzner 2025 · Kim & Maher 2023
THE COST
People stop doing the thinking
Boers 2025
Teams using AI produce more. The cost shows up in the person, not the output.
Every one of these tools competes on removing effort, and in most contexts that is right. An idea is different. You do not know what you want yet, so working it out is the task. Remove that effort and you have removed the only part getting you anywhere.
You stop examining output that already looks polished and finished, so weak ideas and biases go uncorrected.
You feel little psychological ownership of an idea you did not build, and weak ownership means you stop investing in it.
Why it happens
Frictionless AI makes ideation faster and shallower
SYSTEM 1 · FAST THINKING
Automatic: You do not feel it happening.
This is the part that comes up with ideas
SYSTEM 2 · SLOW THINKING
Deliberate: It costs you effort.
This is the part that judges them
WHERE PEOPLE STAY WHEN THEY WORK WITH AI
Fast thinking
Slow thinking
Good creative work needs both. Working with AI, people mostly stay in fast thinking (Wadinambiarachchi et al. 2024).
Human input in AI co-creation is remarkably low. People mostly select, accept or reject what they are handed, and rarely transform or add to it. (Xie 2024 · He & Liu 2026)
People and the model agree far too quickly. The model tends to confirm whatever you say, so an idea gets accepted without ever being tested. (Tsakalerou 2026)
Project Ideas
How to make an experiment around them
Ideas sit in a knowledge graph. Click a connection to read the three sources behind a claim.
Ban the twenty most common words in your category. Say one and the text blurs.
Three experts answer your question. Two are lying confidently. Work out which one is real.
Explain your concept to a monster that grows stronger on every buzzword you use.
A stakeholder who answers “better implies a metric, which metric” until you argue from first principles.
Research
Seven kinds of friction exist.
We could build two.
We surveyed friction mechanisms across educational psychology and interaction design, under desirable difficulties, positive friction and frictional AI, and grouped them into seven families. Two questions cut the seven down to two.
FILTER ONE
Can three students build it in one semester?
FILTER TWO
Does it change what the person does, or only how the screen looks?
Perceptual disfluency and temporal barriers fail the second test. They decorate the problem instead of touching it.
Degraded fonts, hard-to-read typesetting. Make it harder to read.
Microboundaries, artificial processing delays. Make the person wait.
Forced choice, unassisted first steps. Withhold help until they commit.
Engineered discomfort, unfriendly interfaces. Make the system argue.
Scaffolding, guiding questions. Give hints instead of answers.
Effortful assembly, the IKEA effect. Make the person do the work.
Reflective nudges, explainability requirements. Make someone check the reasoning.
Each mechanism we kept targets one of the two causes.
Few generative actions
Generative labour
adds the missing user actions
Consensus too fast
Antagonistic
breaks the fast agreement loop
Perceptual disfluency and temporal barriers add effort without asking you to produce anything, so neither could touch ownership.
THE GAP WE FOUND
Friction in AI systems has mostly been studied in decision support (Buçinca 2021; Natali 2024). Antagonistic conversation has never been tested in creative ideation.
Response
So we built three tools
We built a text ideation tool for marketing briefs. It has three modes behind a single toggle. The model, the task and the length of the session are identical in all three. The only thing that changes is how much work the interface makes you do.
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.
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.
Nobody was told which mode they were in, or that the mode changed between sessions. The next four sections are how each mode works. After that, what happened when people used them.
Browser
The interface
React 18
ReactFlow 11
Zustand 5
POST /api/chat
+ canvas state
Server
A proxy
Express 4 · 72 lines
no prompt logic, no state
localhost
/api/chat
Model Host
Ollama serving llama3
runs on a machine
no key, no quota
The control condition. A plain chat partner, no canvas, no resistance, and what every tool on the market already does.
Generative Labour
The model opens the space. You make the meaning.
Chat on the left, a node canvas on the right. Six turns walk one line of brief to a campaign with a working mechanism. Nodes never disappear between turns, so the tree grows as the session does.
01
Clarify
AI ACTS
Three questions about audience, product and goal. The canvas is still empty.
02
Vibe
YOU ACT
Directions arrive as single words. Click one or two and they attach to your next message.
03
Medium
AI ACTS
Channels branch off the directions you chose, joined by labelled edges.
04
Seeds
YOU ACT
Named concepts built from your direction and your channel. Pick the one that excites you.
05
Mechanism
YOU ACT
Three empty fields. Trigger, action, result. The model writes nothing here.
06
Finalise
AI ACTS
A campaign summary slides up. Finalising turns its top bar green.
One turn does the real work
Turn five is the only turn where the model writes nothing. What will make the audience act, what will they do, and what happens next. Every turn before it was choosing. This one is writing, and it is where people stalled.
Friction needs an exit, so we built four and scoped each one tightly.
UNDO
On the canvas, always.
REGENERATE
Turns one to four only, while ideas are still cheap.
GO BACK
Restores the canvas as it stood at turn four.
HINT
Turn five only, the one turn people actually stalled on.
Effort that feels like busywork produces resentment, not thinking. An open canvas where you connect anything to anything made the labour feel arbitrary.
Six fixed turns walking abstract to concrete, so every action builds on the one before it and the effort has somewhere to go.
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.
Antagonist
It is friendly for two turns, then it turns
TURNS 1 AND 2 · NEUTRAL EVALUATOR
Warm, useful, ordinary
It clarifies your brief, hands you three or four named campaign ideas, and asks which one excites you.
TURN 3 ONWARDS · CRITIC
It stops agreeing
It acknowledges your idea without praising it, asks questions built to find the weak joint, and never fully accepts an answer. Each turn it pushes harder.
THE SWITCH
The moment you choose an idea, the interface swaps the prompt underneath you.
CUT FROM THE PLAN
It was going to shape-shift, role-playing a different stakeholder each turn: a marketer, then a customer, then a sceptic. We cut it for one critic who escalates, because a rotating cast gives you three shallow objections instead of one that gets deeper.
Explore
Calm and curious. Who is it for, what would success look like.
Understand
It restates your idea in its own words and asks if that is right.
Challenge
Fundamental flaws, who would oppose this, what you have overlooked.
Hostility at turn one has nothing to bite. You have not committed to anything yet, so pressure reads as noise and people disengage.
The persona switch waits until after the choice. Every challenge then lands on something the person has already claimed as theirs.
The cost: those first two friendly turns are a small ambush. People trusted it, then it turned on them.
One pilot participant said the antagonistic mode made them feel they were defending someone else’s idea. That showed up in the ownership scores months later. It is the honest limit of this mechanism.
Method
Experiment Design
Every participant did all three modes, in a different order, on a different brief each time, so no mode ever drew the easy question and position cancels out. Sessions were recorded, then replayed turn by turn while people narrated what they had been thinking.
Five orders across five people. Every mode appears in every position exactly once. Whichever mode someone tries first gets their freshest attention, and whichever comes last gets a tired participant, so we made a counterbalanced Latin square. Five orders across five people so position cancels out.
P1
P2
P3
P4
Abandoned Directions
A shape-shifting antagonist
It was going to role-play a different stakeholder each turn. A rotating cast gives three shallow objections instead of one that gets deeper.
A canvas on both friction modes
If both had a canvas, no result could say which mechanism did the work. This cut is what makes the study readable.
A reflection test
We planned to score System 2 with a test taken outside the task. We replayed the recorded sessions instead and asked people to narrate.
Analysis
What came back
We wrote the codebook before anyone looked at a transcript. Three top level codes, twelve underneath, built from dual process theory rather than from whatever we happened to notice.
of coded excerpts were System 1 in the baseline. Under either friction mode, 33 to 47.
turns used in the baseline. Four of five stopped before the tool did, two after three turns.
the baseline time on task for the antagonist. 17:06 against 7:08.
So “friction works” is too simple. System 2 engagement and psychological ownership are separate outcomes, and only one of the two mechanisms moved both. The antagonist also cost 1.4 points of satisfaction and 0.6 of confidence.
Results
Both mechanisms changed how people thought. Only one changed how they felt.
Everything below comes from five participants running fifteen briefs across three modes, with every session recorded and replayed turn by turn. Here is the whole picture on one line each.
In the baseline, four of five people stopped before the tool did, and two quit after three turns. Asked why, they said they had “no more to discuss”. That is premature convergence, and it is the failure the whole project was built to interrupt.
Psychological ownership 3.4 to 5.2. Credit given to the AI fell by two points. Confidence went up a full point. Satisfaction dipped 0.4, which is nothing on five people.
Psychological ownership 3.4 to 3.6. Statistically nothing, on five people. It made people think harder, then cost them 1.4 points of satisfaction and 0.6 of confidence.
System 2 engagement and psychological ownership are separate outcomes, and only one of the two mechanisms moved both.
Conclusion
Three things the study supports
Both moved thinking from fast to slow. Only Generative Labour moved ownership too, 3.4 to 5.2, up for four of five.
Generative Labour runs five seconds a turn slower than baseline. The antagonist takes 2:26 against 1:37. One keeps you in the room, the other slows the room down.
The lowest scorer on the metacognitive inventory, 26 of 52, was weakest on exactly the three subscales friction demands first. They scored 5 of 6 on evaluation, the part that runs after.
Make someone defend an idea and they think harder. Make them build it and they think harder and call it theirs.
Overall
What I learned
Six things that changed how I work.
01
Frame a problem by its causes before you design
Two specific causes, passive uptake and fast agreement, gave every later decision a reason I could defend.
02
Effort has to point somewhere
An open canvas read as arbitrary, not effortful. Six fixed turns made the same work feel like building.
03
The prompt is where the interaction lives
The rule that stops the model editing your nodes is the whole promise of the tool. It lives in a 307-line file.
04
Cutting a feature protected the finding
A canvas on the antagonist would have made a better product and a worthless experiment.
05
Plot people before you plot averages
The mean said the antagonist changed nothing. The five scores said it split the room in half.
06
Put the limit next to the claim
Caveats at the bottom read as hedging. Next to the claim they qualify, the same words read as rigour.
If Generative Labour had raised ownership while leaving thinking unchanged, the mechanism would be about labour, not cognition, and the framing would need rewriting. It did not. The antagonist is the closer call: it moved thinking and not ownership.
A three person project at the School of Information, UT Austin: Ayesha Khan, Ambady Ravi and Pranati Kompella. Built in React and ReactFlow with llama3 served locally through Ollama, so no session left the room. Participants are identified as P1 to P5 throughout, in line with the consent they gave.
