Human–AI research · In development

What makes
progress possible?

What helps people make meaningful progress with AI, and what can one attempt teach the next?

We are developing ways to connect consequential decisions with the work and tests that follow them, and investigate what those experiences could contribute to later attempts.

Explore our first rehearsal

Three translucent doorway studies connected by a pink thread, tracing a change across successive designs.

01 · The questions

Follow the change
in the work and the person.

A finished result leaves questions open. What changed the direction? What did the person learn to judge? What would still be useful somewhere else?

What moved the work forward?

A suggestion may matter because someone tests it, revises it or knows when to seek help. We want to understand those connections through the decisions, work and evidence.

How did the person change?

Expectations, judgment and ways of working may change with AI. Better performance with assistance and an ability someone retains are different questions to examine.

What helps the next attempt?

The longer ambition is knowledge, methods and technology that improve later work. Their value depends on being useful in a meaningfully different situation, under that work’s own tests.

02 · Our first rehearsal

September 2026

Understanding the decisions
behind the work.

Documentary storytelling follows the people, relationships and choices that shape a meaningful attempt. Our research examines what that context contributes to understanding decisions in human–AI collaboration.

In our first research rehearsal, we traced 23 decision episodes from a real creative collaboration and developed a method for comparing AI reconstructions of the work. The analysis connected briefs, instructions, revisions and outcomes, while additional contextual accounts supplied details about contributions and personal criteria.

Historical checklist scores out of 40. Original record, 22 and 25. With synthetic descriptions, 25 and 22. With real contextual accounts, 24 and 27. The respective averages are 23.5, 23.5 and 25.5. — The six AI reconstructions were run using **OpenAI**'s GPT-6 Astra xhigh. Each dot is one AI reconstruction. The short mark shows the pair’s average. Scores count sufficiently complete, source-supported answers on a fixed 40-component historical checklist.

We’re building on this work to investigate how contextual, cultural recording can contribute to understanding research decisions.

Develop documentaries or research human–AI collaboration?Start a conversation

03 · Toward useful methods

Help a later person
choose a better next step.

A reliable way to understand an attempt could eventually help people evaluate AI assistance, recognise when expertise is needed or make a better-supported decision in unfamiliar work.

Two related doorway studies linked by a pink thread beside an unfinished paper model.

The research opportunity

We want to establish what can be learned beyond the interactions and files usually left behind. Useful comparisons and scrutiny from relevant experts are central to that work.

The longer venture

Repeatable methods and technology could serve people beyond an individual filmed journey. That requires evidence of usefulness in later work and a clear need from the people who would use it.

The public film and research evidence have distinct purposes and permissions. The person’s account, observable actions and relevant tests each contribute something different to understanding an attempt.

A question behind the work