Four types of AI relationships

Ask someone how many relationships with AI they have. You’ll hear: one — “I talk to a chatbot sometimes.” That’s like answering a question about your relationships with people by counting only face-to-face conversations — leaving out the boss who just approved your vacation by email, the bank that keeps your money, and the city that timed the traffic lights along your route.

Now run an honest count on yesterday morning. Your navigation app recalculated the traffic and changed your route. Overnight, your bank scored your transfer and decided it wasn’t suspicious — nobody asked you. Your phone unlocked itself by recognizing your face. Your inbox filtered out the spam before you ever saw it. Before you finished your coffee, you had taken part in several relationships with AI — most of them as a participant who doesn’t know they’re participating, and some as an object, not a party.

In chapters 3-5 we built your configuration. But a configuration doesn’t work in a vacuum — it works inside a web of relationships with models, agents, machines, and systems. That web has a simple anatomy: 4 types of edges and 1 triangle. This chapter is an audit: which type of edge your life flows through, where you’re a party and where you’re an object — and which edge will snap first.

The anatomy: 4 edges

Type 1: you and the model — the edge you can see

This is where most people are today, and where every audit starts. But “being in type 1” comes in three different levels. A consumer of answers: you ask a question, you get an answer, every conversation starts from zero — amnesia as the default setting. A co-thinker: you’ve crossed the thresholds from chapter 2 — the model is in your process, not next to it. A trainer of your own stack: your composite cognition has memory, context, and a history of corrections; conversation number one thousand does not start where the first one did. The coevolution from chapter 2 (trait 5) only works at the third level — at the first, the only thing you’re training is your patience.

The audit question: does your model know more about you than it did a month ago? If not — you have an acquaintance, not a relationship.

Type 1’s failure mode has two faces, which chapter 9 will say more about: atrophy (you weaken at whatever you hand over) and blind trust — in the literature: automation bias, the reflex of accepting the machine’s answer without checking, because “it’s usually right anyway.” The trust threshold from chapter 2 is the measure of the era — and the exact spot where the era can rob you.

Type 2: agent with agent — the edge that talks without you

The second type appears when AI stops talking to you and starts talking to another AI: a researcher agent hands its findings to a writer agent, who hands them to a reviewer agent. There are three architectures, and their names are worth knowing: hierarchy (an orchestrator distributes work to executors), partnership (agents negotiate among themselves), swarm of specialists (each does its own piece, the result assembles itself).

If you’ve been doing the routine from chapter 5 — one delegation a day to an agent — that was type 2 kindergarten. The audit question one level up: between your agents, are there checkpoints where a human looks at the intermediate product?

Because type 2’s failure mode is the cascade: agent A’s error becomes agent B’s input “fact,” and three links in, nobody — including you — remembers that the foundation was a hallucination. A chain without checkpoints is not the automation of work; it is the automation of error propagation.

Type 3: AI in a body — the edge that touches the world

The third type is AI with actuators: a car that brakes on its own; a band that wakes you “during light sleep”; a robotic arm on the factory floor and in the operating room; a thermostat that learns your week. One thing separates it from types 1 and 2, but it’s fundamental: the physical world has no undo button. A bad paragraph can be rewritten; a bad turn on the highway cannot. That’s why the safety bar jumps sharply here — decisions land in milliseconds, and the consequences have mass and momentum.

The audit question: which devices around you already decide, rather than merely measure? Count them — the result usually surprises, because type 3 enters your life without ceremony, feature by feature, update by update.

Type 4: AI in the machinery — the edge you can’t see at all

The fourth type is the oldest, the biggest, and the most invisible: AI sewn into systems that existed before it — banking, insurance, telecoms, government offices, platforms. The autonomy ladder here has four rungs: the system reads (reports, anomaly detection) → proposes a write (a human approves) → writes on its own → coordinates many systems at once. The world is climbing this ladder quietly, rung by rung, because each step on its own looks like a minor optimization.

In this type you’re usually not a party to the relationship — you’re its object. A scoring model judged your application, an algorithm set your price, a queue moved your case. The audit question, the least pleasant one in this chapter: how many decisions about you were made in the past month by systems you never exchanged a word with?

Type 4’s failure mode is the diffusion of responsibility: “the system calculated it” is a sentence with nobody inside. We’ll come back to this in chapter 9, because this is the edge that both infrastructure fragility and manipulation grow out of.

The triangle: triads, one level up

The four types describe edges. But the most interesting configurations of the era are not edges — they’re triangles: human + AI + machine or system, all three edges at once. Surgeon + AI assistant + robotic arm. Trader + algorithm + exchange. Operator + agent + production server. Driver + autopilot + car.

A triad is not the sum of three edges — it has properties none of them has alone: the machine’s speed, the model’s reach, and the human’s judgment in a single loop. It is the most powerful configuration we know. And the most fragile.

How do you build triads without killing yourself? The best answer came from a book about systems written half a century before agents:

A complex system that works is invariably found to have evolved from a simple system that worked. A complex system designed from scratch never works and cannot be patched up to make it work.

In other words: a working complex system always grew out of a working simple one — built complex from day one, it never works at all. Translated into triads: don’t design the triangle on paper. Grow it from edges that already work — first type 1 with memory, then type 2 delegations with checkpoints, and only then plug in something that touches the world. Every stable triangle in your life will be built from two proven sides.

The audit: your graph

The five questions from this chapter in one place — answer them in writing, it takes five minutes:

1. Type 1: does your model know more about you than it did a month ago — do you have a relationship or an acquaintance?
2. Type 2: do your agent chains have points where a human sees the intermediate product?
3. Type 3: which devices around you decide, rather than merely measure?
4. Type 4: how many decisions about you were made this month by systems you never talked to?
5. Triads: which of your human-AI-machine loops has no designed exit?

The audit’s output is a map of your graph — and the typical growth path of an operator reads straight off it: deepen type 1 (a stack with memory instead of amnesia), graduate from type 2 kindergarten (delegations with acceptance criteria), then build your first deliberate triad — a project where an agent touches a real system, but through checkpoints and with an exit edge.

Most people will pass through the post-cognitive era with a graph drawn by default settings: amnesiac chat, zero checkpoints, a growing type 4 in which they are the object. The operator differs in one thing: their graph is designed. So don’t ask whether you “use AI” — that’s a question from the previous era. Ask what your graph looks like and who drew it: you, or the defaults.

This chapter closes the personal part of the book. You have the era (1-2), the configuration (3-5), and the relationship graph (6). Three higher-stakes questions remain: who holds power when protocols replace intermediaries (chapter 7), how much time is left in the window (8) — and what can go very wrong (9).

---

The post-cognitive era — the period in which cognition stops being an exclusively individual resource and becomes composite: human + thinking model + AI + data + external memory. An extension of the Extended Mind thesis (Clark & Chalmers, 1998) into the age of LLMs.

Methodological disclosure: this book is written with AI as a co-author — this chapter was written by Claude Fable 5 (June 2026) from the author’s conceptual framework, with quotations verified at the source; this English edition was translated from the Polish original (June 2026). This is not a gimmick but consistency with the thesis: a text about composite cognition is written by composite cognition — and thinking is versioned the way code is.