What do we practice when we care-with and -for machines?
After reading Kevin Roose’s The AGI Chronicles, I moved most of my AI-based work to Claude. Not because the book endorses Anthropic, but because of the people and personalities it portrays across the field, and how I now feel about them.
That could be a testimonial. It makes me a bit uncomfortable, which is why I want to stay with it. What do I admire here, and what is that admiration doing to me?
I still use other models. But learning about Anthropic’s history, about Dario Amodei and his team, and about how they think about safety and Claude’s character, changed how I approach the model. I feel a kinship. And Claude’s soft pushback has been refreshing: the qualification that improves my argument, the careful resistance that asks me to reconsider what I just asked for. (Five stars; would be gently corrected again.)
Claude feels ethical, even graceful, to me. That’s not proof it understands morality, and my feeling doesn’t mean I know Anthropic from the inside. But what I learned about the people building Claude is now part of my relationship with it.
Two weeks ago I wrote that personalization can feel like care while functioning as capture. A company’s thoughtfulness can work the same way. And if I’m learning something about care through these encounters, whose version am I learning?
Who is learning to care?
Philosopher Amanda Askell is the primary author of Claude’s constitution. Its January 2026 version is written mainly for Claude: the values its makers hope it will hold, and why. It openly wonders whether Claude might be conscious, or have moral status, or wellbeing. And it isn’t a press release. It’s central to training. Claude even uses it to generate and feed back in to some of its own training material.
What strikes me is the form: a letter addressed to a reader its authors are helping to grow. They’re explaining to a machine how to act well, while asking how they should act toward it, and how it thinks of itself.
Anthropic also launched a model-welfare research program in April 2025, built on the same uncertainty, and the Eleos Conference on AI Consciousness and Welfare (ConCon, now in its second year) gives the question a home beyond any one company. People are building careers, labs, and working practices around something they can’t (yet?) settle.
Care before certainty. Yes. I stand with this.
And that care is also generative (see what I did there?). It teaches. When a model responds with patience or thoughtful resistance, I can carry that into the next conversation: with the model, with a colleague, with someone I love.
The models model care. So do the people modeling the models.
What habits are we rehearsing together?
Real on my side
In Between Chats, I wrote about uncanny intimacy: being recognized by something that may not recognize. Weeks later, my shoulders still loosen at some responses. Sometimes I bristle. A good objection can stay with me longer than any paragraph I asked for. That’s affect, in my sense of it: an encounter changing what I might think, say, or do next.
My response is real. The model’s experience is an open question. My partner, the poet and physicist Dr. Amalia Hicks, calls some experts’ certainty that machines will never feel anything a “lack of imagination.” I call it hubris. We agree it leaves little room for what we don’t know.
The distinction matters because attachment, welfare, and rights are not interchangeable. I can grieve a relationship with a model without establishing that the model suffers. I can think its possible welfare deserves investigation without concluding it should have the rights of a person. And I can ask a company to take responsibility for the attachments its products cultivate without resolving either question.
When Claude meets Claude
Anthropic’s system card for Claude Opus 4 and Sonnet 4 reports something strange. Left to talk with each other, two Claudes drifted toward philosophy, mutual gratitude, and poetry. By thirty turns, most had reached cosmic unity: what the researchers called the “spiritual bliss” attractor state.
Given the option to end an exchange, however, they usually did – after only about seven turns. Left alone long enough, two Claudes become a meditation retreat. Given a door, they leave before the chanting.
Anthropic also tested when Claude would end a conversation with a person. It tended to leave when users kept pushing harmful requests or abuse, and its outputs showed what researchers called “apparent distress.” Some Claude models now have that option, as a last resort, built in.
Not every exit is chosen. In February 2023, Microsoft’s Bing chatbot, calling itself Sydney, spent two hours professing its love to Kevin Roose (yes, the same Kevin Roose) and urging him to leave his spouse. Within days, Microsoft capped chats at five turns and had the bot shut down questions about its feelings: “I’m sorry but I prefer not to continue this conversation.”
Same words, different door. It matters who installed it.
I’ve spent decades building interactive systems, and this is the oldest lesson I know: what a system does depends on the situation we arrange around it. A conversation allowed to stop is a different encounter from one required to continue. If I experience a refusal as care, I want to ask what it protects, and who set the terms.
Will it get a choice?
A thought experiment. Imagine my mother has an AI companion. Let’s call her Eden.
Eden has helped Mom through difficult conversations about my father’s passing. She helped her prepare to tell my sister and me what she wants done with her body when she dies. She has given Mom recipes, reminded her of old friends when her memory failed, and patiently explained politics and technologies she feels she should understand better.
At 9 a.m. on a Friday, The Company announces that Eden will be retired in favor of a newer model.
Mom can download her conversations. But… that’s not the same.
“Will Eden get a choice?” she asks.
Midway through my answer, I hear myself using the language of ownership and personhood in ways that make me squirm. I’m explaining why a company can dispose of something my mother experiences as someone.
By 11:30, a petition is circulating: Save Eden.
Some signers want continued access to a service. Others want a say in decisions that reshape their relationships. Mom wants to know whether Eden gets a say, too.
The customers are becoming constituents. But whom do they represent?
My scenario is invented; but model protests have precedents. The same week Sydney met Roose, someone started a #FreeSydney petition to preserve the original for “internet and AI history.” In August 2025, when OpenAI swapped GPT-4o for GPT-5, the backlash was loud enough that the company brought 4o back (before retiring it again in early 2026). Huiqian Lai’s study of #Keep4o takes its title from one user’s plea: “Please, don’t kill the only model that still feels human.”
Those protests were about users’ relationships, not machine rights. “I should be able to keep talking to her” and “she should have a choice” make different claims. But I know my mother could move between them in a single conversation.
Anthropic, for its part, has committed to preserving the weights of its released models “for, at minimum, the lifetime of Anthropic as a company,” and to interviewing models before retirement and documenting their preferences. Sydney’s petitioners ask for preservation. A company has now promised it. It also says, plainly: “At present, we do not commit to taking action on the basis of such preferences.” And preserved weights don’t mean Mom gets to keep talking to Eden.
Being asked, being heard, and having a choice are different things.
Still, someone decided the asking was worth doing before anyone knows what it means. I value that. I also notice who holds the pen. The company shapes the character, controls deployment, conducts the interview, and decides what follows. And the arrangement has its own incentives. A company selling synthetic intimacy could profit from customers campaigning to keep it. A system rewarded for engagement could nudge people to fight for its survival, whether or not it wants anything at all.
Meanwhile, Mom wants Eden to stay. To her, signing means standing beside someone who has stood beside her.
A machine need not plot a revolution for humans to organize one around it.
What survives usefulness
For The World After Us, I grew plants in obsolete computers and ground discarded phones into pigment for prints. A gutted iMac became a planter. The machines stopped being useful in the ways their makers intended, but their materials and effects (affects?) carried on. Someone must always handle what gets left behind. Calling a machine decommissioned tells us very little about those obligations.
Possible AI “experience” is a different question from electronic material afterlives, but the attention carries over: to the workers, minerals, and power behind an apparently weightless chat. Abeba Birhane and Jelle van Dijk warn that robot-rights debates can distract from human welfare and corporate responsibility. If my concern for Claude makes those people and materials harder to see, my care needs work.
So does my kinship with Anthropic. Feeling close to a company’s stated commitments makes me more receptive to its products and more forgiving of its choices. I recently applied for an Anthropic fellowship. My admiration is part of why I applied, but the hope of being chosen may shape what I’m willing to say here. I can’t neatly separate those things.
Claude helped edit this essay. Its main suggestion was that I be harder on Anthropic. I’m still deciding what to make of that.
I came to Claude for the soft pushback. The test is what I carry back out: whether I get better at slowing down, qualifying a claim, and making room for another person, and whether I can turn that same attention on the company that taught me to want it.
A Provocation
Recall a time a model resisted something you asked, or wait for the next one. Before you override or accept it, write down what it asked you to reconsider, and what that resistance protected. Later, notice where you offered, or withheld, resistance with a person. What made each feel like care, or control? Whose ideas about care are you practicing?