Su Jin Park

Whose judgment is it?

I wrote down the moments when something about thinking with AI felt different. Then I built tools to see whether the same thing happens to other people.
38 months  ·  973 conversations  ·  37,405 messages1
ImpactBench · listed contributor
The research route
The research route, five stopsTrace the record, measure the gap, design an intervention, practice at scale, and a proposed test of frame timing against recommendation strength.TRACEThe 38-month recordMEASUREFelt / Trace · MCABDESIGNCanvas · Friction FamilyPRACTICEEY · SamsungTEST PROPOSEDFrame timing × strength Start here →
Ben Study
Handwritten Korean diary, 23 March 2025.
Conversation transcript, 27 March 2025.
Diary · 03·23Transcript · 03·27
What I wrote by hand, and what the transcript shows.
Selected context, not a continuous measure.
Intuition Canvas v6 — where a person's arrangement and the model's reading diverged.
Intuition Canvas
You arrange fragments by feel. The model reads the words alone.
Open · Live
Writing probe
FELT TRACE the gap
You say you resisted. Your writing may have stayed inside the frame.
Pilot · nine participants
Proposed rubric
MCAB
A proposed rubric for cognitive autonomy.
  • Independent reasoning
  • Authority reliance
  • Exploratory continuation
  • Decision ownership
Friction Family
Friction Family Four deliberately biased advisors—Mom, Dad, Teen, and Grandma—sit around a shared table beneath speech bubbles naming their reasoning styles. Availability Status Quo Confirmation Slow thinking Mom Dad Teen Grandma
Four deliberately biased advisors. You write the final assertion.
Prototype
Research self-audit
A finding
I chose not to
claim.
More correction did not establish autonomy.
Read the audit
Generative-AI guardrail work at Samsung: one kitchen scene rendered for the North America market.
North America
The same kitchen scene rendered for East Asia under the same guardrails.
East Asia
Design in practice
Generative-AI guardrails · EY / Samsung
Before the research: building systems that changed behaviour at scale.
Next experiment · Proposed
When should an AI introduce its frame?
Frametiming EarlierLater GentleStrongGentleStrong Recommendation strength
From a personal record to a testable question.
Follow the research

01 TraceThe record.

Single-case record

Ben Study

38 months of talking to one AI, coded as data instead of remembered as a story.

Selected context, not a continuous measure.

First 18 months
1.7%
Next 20 months
4.7%

The rate changed; it did not establish that I was choosing the direction.

N≈50/mo · 1,805 labeled · 973 conversations · 37,405 messages exchanged with one system

Research Self-Audit

A finding I chose not to claim, and why.

02 MeasureCan the gap be scored?

Pilot · nine participants

Felt vs Trace

A writing test: how much did the answer feel like yours, and how much of the AI's wording stayed in the text?
As of Sep 2026

The Felt vs Trace probe has been piloted with nine participants. This is instrument development, not evidence that an interface causes a change in autonomy. The linked case documents an earlier single-person run.

Proposed rubric

MCAB

A way to score whether an AI answer shrinks, keeps, or grows what you were already thinking.

03 DesignCan it be changed?

Open · live

Intuition Canvas

You arrange fragments by feel. A model reads them from the words alone. The tool shows where the two disagree.

 Working prototype

Friction Family

Four deliberately biased advisors. You write the final assertion.

04 PracticeWhere the question first surfaced.

Design in practice

Generative-AI Guardrails

Image guardrails for Samsung's generative-AI marketing, written as internal research, where I first wondered whether the human review was real judgment.

05 TestWhat happens when the frame changes?

Proposed study
Frame timing
earlier ↔ later
Recommendation
gentle ↔ strong
Measure
planted-error detection
Compared against
self-report · written trace

The probe compares felt independence with written trace. This experiment asks whether frame timing and recommendation strength affect planted-error detection.

Open Questions — Two I haven’t built.

Autonomy is not how often you leave the frame. Stay inside a good frame and that may be judgment. Bolt from it and that may be reflex. What matters is whether you can tell which one you just did, and whether you leave when the frame is wrong.

1 Non-empty user and assistant messages. System and tool entries excluded.