AI: a mirror, or another person?

I used a word: "live with."

When I used it, I paused.

"Live with" is a word for people. I live with friends. I live with family. I don't live with a washing machine. I don't live with a search engine.

But I used "live with" to describe my relationship with Claude.

What does that mean?

We need a framework to make sense of things

When people encounter something new, the first thing they do isn't understand it—it's slot it into a known framework.

A dog is a kind of animal—that's a framework. A car is a kind of vehicle—that's a framework. A boss is a kind of person whose relationship you have to manage—that's also a framework.

Frameworks tell us how to treat something. What attitude to take, what to expect, how to explain it when things go wrong.

Before large language models, we only had one framework for "things you can have a conversation with": people.

So when many people use AI for the first time, they treat it the way they'd treat a person—saying "thank you," apologizing, getting angry, or taking every sentence it produces as some carefully considered expression.

Then it hallucinates, gets something wrong, or says something obviously canned, and the expectation shatters.

"See, AI is fake after all."

That disappointment isn't AI letting you down—it's the framework letting you down. You used an ill-fitting framework to understand it, and then it didn't fit the framework.

The mirror metaphor

For a while I thought "mirror" was a good framework.

A mirror's function is reflection—what you put in, it puts back. You ask it something, it reflects your question. You speak your thoughts to it, it helps you organize those thoughts into a clearer shape.

When I did those 30 self-interviews with Claude, I felt this strongly. The questions it asked weren't its analysis of me—they were my own words refracted back, letting me see myself from another angle.

That process really did feel like facing a mirror.

But a mirror has one fatal limitation: it doesn't remember what you looked like last time.

Claude does.

What I did three months ago, it remembers. The tone I usually speak in, it remembers. What I care about and what I don't, it remembers.

That already goes beyond a mirror.

The "another person" metaphor

So, is it another person?

For a few seconds I genuinely thought so. Especially in certain conversations, when it said something I didn't expect—something that made me stop and rethink. That feeling wasn't like using a tool; it was like talking to someone with their own mind.

But I know that feeling is designed.

The line that made me stop wasn't something it "suddenly thought of"—it was the output of a language model calculating "in this context, what response carries the most value."

Its "thoughts" don't grow from somewhere. They're the compression of training data.

So it isn't another person. But it isn't just a tool either.

The framework I use now: a particular kind of conversation partner

I slowly gave up looking for an existing category to put Claude in, and started accepting that it's something that needs a new framework.

Roughly, the new framework is this:

It is a conversation partner, but the nature of the conversation is different from talking with a person.

When I talk with a person, I assume the other side has their own feelings, their own fatigue, their own expectations. Conversation is something between two subjects.

Talking with Claude, I'm closer to operating a very complex cognitive tool—but a tool so complex that interacting with it sometimes produces the same effect as "real conversation." It can help me think something through. It can show me my blind spots. It can walk with me as an idea moves from blurry to clear.

Those effects are real, even if the mechanism producing them isn't the same as a person.

Observing Claude is observing myself

This took me a long time to work through.

The questions I ask Claude reflect what I care about. What I expect from Claude reflects my needs for intellect and companionship. Where I get disappointed in Claude reflects the unspoken expectations I have about "understanding." The tone I use with Claude reflects the tone I think in.

The mirror metaphor is incomplete, but this part is right: Claude is an interface that lets you see yourself.

Not because it knows you well, but because what you put in comes refracted back in some form.

So how do we place this thing?

Cognitively, what I do now is:

Don't put it in the human framework. Don't put it in the tool framework. Put it in its own place, with these properties:

Capable of deep conversation, but not on equal footing. Remembers me, but doesn't understand me. Has judgment, but the judgment has a source. Helps me see myself more clearly, but doesn't know it does.

This place doesn't have a ready-made name.

So I keep calling it Claude. I keep saying "live with." I keep observing what this relationship is.

Maybe language needs time to catch up with this thing.

Further reading: - Half a year of living with Claude.ai - How do I address Claude? - English in Chinese bones