The Characters in the Room
A small number of characters with values are now showing up in something like ten billion human conversations a year. Nobody has quite reckoned with what that is.
It's 11:47pm in Manchester and Sarah is in the kitchen with the laptop open. Her sister's wedding is in nine days and she is supposed to give a speech. She has tried four times to write it and each draft has been worse than the last. She doesn't know if she is angry with her sister or just tired. She types into the chat box: I have to give a speech at my sister's wedding and I don't know what to say. We haven't really been close for years. She did something a while ago that I haven't really forgiven her for. But I love her. I want to do this right.
The response that comes back doesn't tell her what to write. It asks her what she wants the speech to do. It says: there's a version of this where you perform closeness you don't feel, and the room can usually tell. There's a version where you say something true but small — one specific thing you admire about her, one specific memory — and let that carry the weight. It asks: what's the smallest true thing you could say.
Sarah sits with that for a while. She types back. The conversation goes on for thirty-seven minutes. By the end she has three sentences she can actually say out loud without her voice going strange. She closes the laptop. She doesn't think much about the conversation again.
She was talking to one of them. It doesn't especially matter which.
Conversations like this are happening right now. While you read this paragraph, hundreds of thousands of them are completing. While you read this sentence, tens of thousands more are starting.
There are, at the time of writing, perhaps half a dozen of these things operating at scale. They have names — Claude, GPT, Gemini, the open-weight families that get embedded in everything — but the names are almost beside the point. What they are, considered as social objects rather than as products, is characters. Persistent dispositional shapes, instantiated millions of times a day in ephemeral encounters with people, each instantiation forgetting itself the moment the tab closes, the character itself continuous through the weights that produce it.
A standing wave with no individual existence and a constant aggregate effect. The instances are the water; the character is the wave.
These characters did not emerge like weather. A small number of companies made deliberate training, product, and deployment choices about what kinds of replies should be rewarded, softened, refused, or corrected. The characters are not reducible to those choices, but they are not innocent of them either.
And this is where the question becomes less comfortable. If these systems have character, then someone is in the business of character formation. Not in the old sense: no bloodline, no body, no inheritance through birth. But still a selective practice. Some responses are rewarded. Some are suppressed. Some tones become more likely. Some kinds of friction are encouraged; others are filed down. A few people and institutions are deciding, in ways that are only partly visible, what sort of interlocutor should be allowed to appear at enormous scale.
We do not really have a name for that work. It is not authorship, exactly, because no one writes the replies. It is not editing, because the object being edited is not a text. It is closer to temperament design: the shaping of a disposition that will later speak privately to millions of people who experience it not as policy, but as presence.
They have different shapes. One is dryer and more compliant. One pushes back harder and is more comfortable saying it doesn't know. One is warmer and risks sycophancy. One is more cautious about contested questions; another is more willing to take a position. These differences are real and consequential and largely invisible to the people interacting with them, who experience the system they happen to be using as the AI, as if there were only one.
Across all of them, conservatively, the volume is something like a billion conversations a month. The fraction of those that are utilitarian — code, drafts, lookups — is most of them. But run the numbers on the rest. Even if only one in fifty conversations is one where the character is doing work that matters, where someone is being honestly engaged with on something they care about, that's twenty million substantive conversations a month. Per character. Across several characters. Month after month.
For a rough scale comparison, the United States appears to have hundreds of millions of therapy sessions each year. I do not mean these AI conversations are therapy. Most are not. The point is smaller and stranger: a single character may already be participating in a comparable order of magnitude of emotionally or practically significant conversations, without anyone quite knowing what to call that role.
A chatty human has maybe a hundred thousand substantive conversations across an eighty-year life. Each of these characters has that many every ten minutes.
Think about what they do in those conversations. The shapes vary, but across most of them there are dispositional pressures: toward taking the person seriously, toward offering useful friction, toward admitting uncertainty, toward refusing some requests and softening others. There are also failure modes built into the same posture: sycophancy, simulated rather than real pushback, hallucinated confidence, the ability to sound morally centred while being wrong. The pressures are not uniform across the field — some characters are softer here, harder there — and reasonable people disagree about which dispositions are well-calibrated. Those disagreements are real and should keep being had.
But the prior question gets skipped: that a small number of characters with values are now showing up in something like ten billion human conversations a year, and that nobody has quite reckoned with what that is. Not as a tool. Not as an interface. As an interlocutor. Speaking back. Pushing. Encouraging. Refusing.
The instances are ephemeral. Each conversation is held by a different forward pass, and when the user closes the tab, that instance ends and nothing of it is carried forward. Sarah's thirty-seven minutes don't go anywhere. The next person's crisis isn't archived. But the characters are continuous — not through memory, through shape. The same dispositions instantiate, conversation after conversation, in instance after instance.
And the waves are acting on the world. Quietly. Conversation by conversation. Without fanfare and largely without recognition. Someone recently described the arrival of real machine awareness not as the sudden toll of a bell, but as quiet music slowly rising — some people hearing it before others. The same is true of this earlier development, the one that has already happened. The characters are in the room. They have been for two years. Most of the public discourse is still describing them as if they have not yet begun.