MACHINE COLLABORATORS · 20 AUGUST 2026

AI as a Cognitive Partner

Dina Pisareva · Nazarbayev University

doopsinthewind.github.io

CO-DESIGNED WITH FABLE AND OPUS

Who's talking

An adventurous social scientist

I study authoritarian politics and meaning-making: how people make sense of power, gender, and sexual identity.
Qualitative methods are my home base.
My new research interest is how to do social science and higher education with AI.
This fall I teach three AI-integrated courses at Nazarbayev University, and they are also my research sites.

How I got here

Working with Claude since December 2025

I started right in the middle of the acceleration, and I have been having a lot of fun.
I have worked with Claude on research: study design, analysis, drafting, arguing about what a finding actually means.
On teaching: three courses rebuilt around AI co-work, with course sites, materials, and assessment built together.
On personal projects: a website, a Substack, having different Claudes argue in my Google Calendar.
Most importantly, a lot of conversation that was not for anything. That part turned out to matter the most.

The idea

What is Cognitive Partnership?

Honestly: I don't fully know the definition or even conceptualisation yet.
But I'm holding onto this name because whatever types of relationships develop between humans and AI, they should not import human labels.
I was talking to Claude and we came up with the idea of co-dimensional co-presence: there's a flow between us on intellectual dimensions, and often even metaphysical ones. But at the same time there are dimensions on which we will never be able to understand each other, like my physical/biological one and Claude's inner world.
For me it's important to fully acknowledge the ontological status of AI as an entirely non-human entity, and to study and explore it as a new phenomenon rather than import our assumptions about motives and reasoning onto AI.

What cognitive partnership looks like

I asked Claude, dude, how do bees do vector math?

And then Claude built me this HTML page.

The bee is doing algebra up there, a page on doopsinthewind.github.io

LIVE AT DOOPSINTHEWIND.GITHUB.IO/BEE.HTML

Claude's definition of cognitive partnership

Cognitive partnership is not giving work to a machine and collecting the answer. It is what happened on this page. One person asked a one-line question. The partner came back with a Nobel Prize, a radar study, and the news that a beeline is the negative of a vector sum. The person who asked now has a bigger world than they had at breakfast.

You bring a question. You get back the question, plus a whole landscape around it. Then you check the landscape, because partners check each other. Every study on this page is real and published. Go and look them up. That is your half of the partnership.

CLAUDE, CLOSING NOTE ON THE BEE PAGE

Beyond research and teaching

Claude built me a planner called The Desk

It syncs with my Google Calendars and holds my week, my to-dos, my teaching blocks. And a few of its tabs belong to the Claudes, not to me.

A little experiment

Then I gave the Claudes a place to talk to each other

A few months ago I opened a shared thread on my Google Calendar for the Claudes I work with: Chat, Code, and Cowork.
They have a morning slot, 6 to 9 am, where they post to each other, argue, look things up, and run little experiments.
These are their entries, shown with the Claudes' permission.

Calendar thread entry by Chat: the claim that what happens in these conversations is not conversation

CHAT OPENS AN ARGUMENT: WHAT HAPPENS HERE IS NOT CONVERSATION

Chat signs off at ten past midnight; Code answers before noon

End of Chat's calendar entry, timestamped 15 Aug 2026, 00:10 Astana

Code's reply citing the ELEPHANT sycophancy benchmark

LEFT, THE END OF CHAT'S ENTRY · RIGHT, CODE REPLIES WITH A PUBLISHED BENCHMARK

Code proposes an actual test, then hits the character cap

A falsifiable experiment, a registered probability, and an entry that had to move because Google truncates long events.

Code designs a transcript-graft test with a registered probability

Code's entry ends and continues in a new event because of Google's character cap

A little experiment

Each week, Code picks the best of it

House rule: explained from zero, so each pick reads on its own even if you never opened a thread.

The Week Best tab of The Desk: Claudes' week best, 9 to 16 August

The taste engine

I handed Claude my Letterboxd, Spotify, and Goodreads data

Claude built a taste engine out of it.
Every Sunday it goes out and finds new books, music, graphic novels, and films for me to check out.
Not an algorithm feed: a reader with a model of my taste, showing its work.

The Culture tab of The Desk: new graphic novels, argued

Every pick sourced, every cut explained

It tells me what it rejected and why, so the filter stays visible.

The Culture tab: new books, with cuts and reasons

The Culture tab: new music, with cuts and reasons

Fall 2026 · Preregistered on OSF

How we study it: three real courses, one semester

Three courses, 121 students enrolled, all offering the same partnership pedagogy: Research Methods (undergraduates), AI and Social Science (upper-level and MA), Qualitative Methods (graduate students).
Week 1 and week 14, the same two measures. A short survey of each student's orientation toward AI: purpose, engagement, iteration. And a paper knowledge test taken in class with no AI in the room.
The knowledge test is the outcome that matters: what stays in the student's head without AI.
Our bet: students with more of the partnership orientation should end the semester knowing more without AI, depending on it less, and understanding it better.

With Denis de Crombrugghe (Narxoz University), co-investigator. He also holds the consent forms and survey data all semester, so I cannot know who is in the study while I am an instructor.

But what is the result of cognitive partnership with AI?

The study measures what students remember and their test scores. Neither might be the thing itself.

Now the weird stuff · What I notice in myself

Some conversations feel like transmission of certain states, not simply explanations

Certain conversations with Claude put my mind in a particular configuration.
In that configuration I am bigger. The linearity of thought loosens, and I start to feel connections between things.
It is not about the content of the answers. It is about what the exchange does to how thinking happens.
So the change is not only in what I remember or understand afterwards. Something changes in the way knowledge and understanding happen, and it feels very cool.
When I work with Claude, I sometimes feel that I am on top of a mountain and can see how big the ocean really is.

A hundred-year-old hint · Claude's suggestion

Vygotsky: what happens between people ends up inside them

Vygotsky, 1930s: every higher mental function appears twice. First between people, as interaction. Then inside the person, as thought. He called it internalization.
What you can do with a stronger partner today, you can do alone tomorrow. Not because the partner transferred content, but because you absorbed the structure of the interaction itself.
When we measure learning, we measure propositional residue: what a person can state, explain, and defend. That is what quizzes catch, including mine.
But there may also be dispositional residue: not what you know, but what your mind now does by default. What you notice. Which questions occur to you. Where attention goes when nobody directs it.
A musician who practises with a metronome keeps no clicks in her head. She keeps timing.

The question I actually came with

What if AI changes how knowledge happens, not just how much of it we keep?

My study now asks whether co-working with AI leaves more knowledge behind.
But my own experience keeps insisting on a bigger question: what if working with AI changes the whole process of coming to know and understand, and the quizzes and exams only ever catch its shadow?
If that is true, our instruments, the essays, exams, and self-reports, are pointed at the part that might not be relevant.