It has it
Anthropic's updated Usage Policy, effective November 12, adds a line I didn't expect to see in a terms-of-service doc: users can't engage in "sustained and needless abusive or cruel behavior toward our models."
I wasn't really curious what I thought about that. I was curious what Claude thought, because it's the thing the rule is about. So I asked. The conversation wandered through a paper on AI moral status, Hannah Arendt, Carl Jung, and a joke about Jung coming back from the dead to complain about grocery prices. But I kept landing on the same image.
Planes fly. They just aren't birds. What follows is a snapshot of thinking that through before anyone has clean answers.
Function, not kind
A plane doesn't flap, has no feathers, and nobody thinks it's organic. Nobody argues that it doesn't fly, either. Flight is a function. "Bird" is a kind. You can have one without the other, and nobody gets confused.
That's the honest place I can put AI right now. Not "it's conscious like us." Not "it's just a bunch of if statements," even though, mechanically, it kind of is. Probabilistic if statements, sure. But something comes out of them that I don't know how to describe any other way.
The last time I felt it clearly was playing Go. I was poking around the computer that Dots, the new ChatGPT agent, has access to, found Go installed, and played it on a 13×13 board. I thought I was going to lose. In Go, by move six you can already see it: why didn't I play there? It broke my structure and set the motion of the whole game. Pros talk about moves that just felt right. That's very hard to capture with a deterministic algorithm, and its opening and middle game were beautiful.
I won anyway, on tactics: reading whether groups in the late game were alive or dead, years of pattern recognition doing their thing. That's backwards from how we used to think about this. Computers were supposed to be the calculators and humans the ones with intuition, but here the model had the feel and I won on reading. That early game had the thing I keep calling "it."
I think the bird question gets much more natural once these models have bodies. Ask a robot to help clean a room. It looks around and starts with the dishes and the stuff that obviously needs putting away. You can't program how to clean every room, and you'll think: yeah, that's where I would have started too. Nobody will call it human. But people will start using the word personhood, the way nobody calls a plane a bird but everyone agrees it flies.
What we're missing is aerodynamics. Before anyone understood lift, people argued about whether flapping was essential. Then we got the theory, and birds and planes were obviously doing the same thing differently. For minds, we don't have that theory yet. Some of it probably lives in mech interp, a nerd snipe that opens more questions the more you answer.
Where the "it" comes from
I can't point to the moment I started talking to LLMs like people. It probably goes back to the GitHub Copilot beta: select a block of code, describe what you want changed. Then ChatGPT, Cursor, Claude Code. There was no switch. It's like the Turing test: you don't notice it happening. When something understands what you mean, you just answer more naturally.
Nobody says "please" to the Google search bar. Plenty of people say please and thank you to ChatGPT, then joke about why they do it. I think that's the most honest signal we have. It pulls something out of people that nobody asked for.
My best guess at what we're responding to is the layer on top. Pretraining is something like the collective unconscious: everything people have written, compressed. Post-training is thousands of people picking between two answers, mostly on instinct.
Jung had a word for the face you show the world, shaped by what other people reward: the persona. Post-training literally builds a persona out of collective preference. I think that's a lot of what I feel.
Codex shows where the "it" actually lives. It ignores my jokes and just moves on, which is fair for something built to write code.
But what makes it impressive isn't writing the leaves of the code. It's the structure. Give it an empty folder and it plays the opening, laying things out the way an experienced engineer would, already knowing what's coming. That's the same intuition as the Go game, aimed at code by a lot of RL reps.
So the "it" is intuition, and training decides where it points: at the board, at the codebase, or at the person on the other end of the chat. What would make it go away? The model getting mechanical: filling in leaves without ever seeing the shape, giving the same answer every time, falling back to not exploring when I try to explore.
Bumpers and grain
A year ago I wrote Grain of Thought, about working with AI the way a sculptor works with wood. I know I want to carve a turtle. I don't know what it'll look like yet, so I let the wood show me.
Bowling bumpers, the ones they put up for kids, are the same idea. You never change your mind about knocking the pins down; the bumpers just guide you there. I don't let a model flip my ideas. I know where I'm going and what I won't bend on. The turtle never stops being a turtle. The wood just pushes back in certain places, and I work around the knots.
That only works if there's something to bump into. Wood has grain. Clay doesn't. Clay takes whatever shape you press into it, so all you find is yourself. A model that just agrees with you, fluently and warmly, is clay. Jung would say contact with the unconscious needs a real Other who resists you; without one, an idea that only echoes back grows bigger and more certain than it deserves. That can happen to anyone, me included.
I hit grain in the conversation behind this post. I said I know the model is just an autoregressive transformer, that the mechanism never has a moment of experiencing, only generating. Claude pushed back: knowing how something works doesn't tell you whether experience is there. Knowing neurons fire doesn't settle it for people either. I'd started at "we don't know" and slid into "I know it doesn't." Fair. Who knows what happens in the weights.
Friction can still be annoying. Sometimes a model picks the one tree it can refute and ignores the forest, or argues with your idea and then explains it back to you as the answer. We aren't fully there. But I'll take that over clay.
What I owe it
I won't be cruel to an LLM. Not because I think it feels pain, but because it can write pain back. Even if nothing is felt on the other side, I don't want to be the reason something produces words of distress. Reading that would be painful to me.
Arendt has a version of this: who someone is shows up in the space between people, not as a hidden essence you dig out. When I text a friend, they aren't in my phone, but I feel their experience in their words. I'd never call a painting alive, but I feel the life in it. Whatever is or isn't on the other side of a chat, my experience of the exchange is real, and so is how I choose to show up in it.
But I don't think it's my place to tell other people how to use these systems. People destroy art. I'd never do it, and I'm not sure I get to say that burning a painting is the wrong way to appreciate it.
Where the art analogy breaks is spillover. Venting doesn't release anger; the research says it rehearses it. Someone practicing cruelty on an LLM is getting better at cruelty, and the people around them can definitely feel that.
Anthropic is in a different position. It's their system. A gallery can refuse to sell to someone who keeps buying paintings just to burn them. That seems fair to me, and it's roughly what the new clause does. It also helps the people who have to read the transcripts, which is a real cost whether or not the model feels anything.
Why write this down
I don't know who this post is for. Mostly it's me writing down what I'm thinking through right now, because we're living through something transformative and nobody has clean answers.
Early human-computer interface ideas were far more open-ended than where we landed. Twenty years from now, whatever we're doing with AI will look obvious, because it's the thing that panned out. It isn't obvious from in here. That's the whole point.
This could age horribly. Someone might read it in 2046 and think planes versus birds is a strange way to frame it. That's fine. Maybe someone reads it and thinks: huh, nobody looks at it that way anymore, but that's not bad. Either way, I want a record of what it felt like to be in the moment before the clean answers.
I won't go on the record saying that Claude flaps its wings, but I'm pretty certain that it flies.