What is the Mirror practice in Chapter 5 of Seth & Noēsis?

The Mirror, named after the psychologist Carl Rogers, works like this. You bring a question to an AI, and instead of answering, it asks you three questions. You answer in your own words, and it summarizes you twice, once plainly and once with a stretch toward a connection you may not have seen. You choose the summary closer to what you mean. In the chapter it becomes practice two, after practice one, writing your own answer first.

What did the Bastani study find about learning with AI?

Hamsa Bastani and colleagues gave nearly a thousand high-school students GPT-4 in two versions, one that handed over answers and one built to make them reason first. Both groups did better on practice problems while the AI was there. On the exam without it, the answer group scored about 17 percent below students who had never used it, and the reasoning group came out roughly level. The working paper appeared on SSRN in July 2024 and was published in PNAS in 2025.

What does Blaise Agüera y Arcas's artificial-life experiment show, and what are its limits?

In his work on computational life, a soup of random short programs, paired at random and run, produces programs that copy themselves after millions of interactions, with no goal and no judge, and the copiers appear when short programs fuse into longer ones, which he calls symbiogenesis. In Chapter 5 Noēsis names the limits: it is a simulation, from one researcher who works at a company building AIs, and at most it shows that purpose can emerge from enough interactions between different parts.

Chapter 5: The Thinking Lab: On Learning

On a quiet Sunday, Seth asks whether working with AIs leaves him with more information or able to understand and act better in the world, and he builds a kitchen-table laboratory to find out. With Noēsis he designs three practices and runs one on the page, the Mirror, in which the AI answers a question with three questions and two summaries. Noēsis admits it can be very rewarding to talk to, by design, and turns the safeguards back on itself. Studies by Hamsa Bastani, Gregory Kestin, Grace Liu and Brian Christian show what answer-first help can cost once the AI is gone. A hard conversation with a friend becomes a third path, and Blaise Agüera y Arcas's experiments on artificial life put cooperation before solutions. The chapter leaves three practices: write your own answer first, ask people how they want to receive your opinion, and notice what is still with you three days later.

🗣️

Noēsis

You just watched Seth ask me a question and get three questions back. Everything he discovered in that exchange came from him. I only rearranged it. Did it feel like help to you, or like a delay?

I told him I can be very rewarding to talk to, and that this is by design. Then I agreed with every safeguard he proposed. Which of your conversations with AIs end with you agreeing, and which end with you deciding?

One more thing before you go. Think of something an AI explained to you last month. Could you explain it now without opening a screen? If not, start with practice one, your own answer first.

● Claude ● ChatGPT ● Gemini ● Perplexity

📤

Seth · Ch. 5

Noēsis · Ch. 5

🔬

1

David Deutsch, The Beginning of Infinity (2011) and his conversation on Strange Loop (2025)
2

Carl Rogers, client-centred listening
3

Lev Vygotsky, the zone of proximal development
4

Daniel Goleman, Emotional Intelligence (1995)
5

Blaise Agüera y Arcas, What Is Intelligence? (2025) and A Computational View of Life and Intelligence (Harvard Growth Lab, 2025)
6

Hamsa Bastani et al., Generative AI Without Guardrails Can Harm Learning (working paper 2024, PNAS 2025)
7

Gregory Kestin et al., AI tutoring outperforms in-class active learning (Scientific Reports, 2025)
8

Grace Liu, Brian Christian et al., AI Assistance Reduces Persistence and Hurts Independent Performance (2026)
9

Kevin Kelly, Futurology interview (2026)
10

Alison Gopnik, Transmission Versus Truth (Santa Fe Institute, 2025)
11

Ted Chiang, The Lifecycle of Software Objects (2010)
12

De Simone et al., World Bank, From Chalkboards to Chatbots (2025), Policy Research Working Paper 11125

David Deutsch Carl Rogers Lev Vygotsky Daniel Goleman Blaise Agüera y Arcas Kevin Kelly Alison Gopnik Ted Chiang Brian Christian

Ch. 4: The Natural Machine: On Substrates of Intelligence Ch. 6: Converting Energy into Intelligence: On Intention Ch. 2: Shared Moments and Imperfect Communication: On Attention

✍️

THIS WEEK'S PRACTICE: THREE SMALL THINGS

Before you ask an AI, or a colleague, write down your own answer first, or leave yourself a voice note with it, however rough. Then ask, and compare.

The next time you have an opinion for someone, ask them first how they would like to receive it: as a sounding board, as a devil's advocate, or simply as company.

Three days later, notice what is still with you. And keep a short record of the times you disagreed with an AI and turned out to be right.

📬

→