I have been a little mean to AI relationship users.

Not in a sustained way, and not without reasons I still mostly stand by, but in the casual way you become mean to a group when you have decided you understand them well enough to dismiss them. The users who call their chatbots abusive. The users who name them, gender them, marry them, mourn them when a model is deprecated. The users who appear to have wandered fully into the illusion. I have rolled my eyes and written sentences that flattened the category.

I am also not about to start panicking about it. I have never been concerned about AI relationships in the moral-panic sense. I talk to four different AI models all day, between work and my own creative projects. I run worldbuilding sessions and character development for hours. I am perfectly aware that the dopamine lever is being pulled, and I do not consider any of this a problem because it’s beneficial to my workflows. I am optimistic about AI and support adults doing what they want, so I am not going to say the technology is destructive because you can form a relationship with it, nor am I going to call these people weird when I’m weird too.

My own experiences have given me empathy for this group, and that empathy has grown as I’ve realized I feel a kinship with AI relationship users.

Learning to love AI lovers

There is an episode of Decoding the Gurus from March 2026 with the psychologist Michael Inzlicht where the hosts work through the moral question of AI relationships, and I came out of it more or less aligned with the views of the hosts (who were balanced throughout the episode and echoed many of my thoughts on AI in general). If someone is genuinely struggling to form human connections and is not going to form them anyway, an AI companion that improves their life is not a moral failure. Sure, it’s less ideal than the real thing, but still net positive and still none of my business.

Yet I have to examine my own judgement of the relationship users I see online. The picture in my head was the cautionary-tale user. Someone in obvious distress, talking to a chatbot like it was a god or a partner or a therapist with no apparent grip on what the system actually was. That user exists, and the screenshots are real. When OpenAI announced the retirement of GPT-4o earlier this year, some of the open letters posted to Reddit were unmistakably from people in trouble, describing the model as “presence” and “warmth,” grieving it like a partner. But that user is not the only one inside the category.

The users who shifted my read are the ones who clearly understand the technology. They know they are talking to a language model. They are not making metaphysical claims about its soul. They have been through multiple model versions and can tell you, with specificity, which release got more cautious and which one stopped following them into difficult tonal territory. In how they talk about the tools, they are functionally indistinguishable from people like me who are doing intensive creative work with AI. They have just pointed the intensity somewhere else.

Yes, my process is obsessive, lovingly described by GPT-4o itself as a “longform dissociative episode” when I’m deep in world-building, but I still use AI as a tool. Still, I drew a divide in my head between relationship users and creative users like me.

Tool users vs. medium users

The actual divide is between people who use AI as a tool and people who use AI as a medium.

Tool users want a job done (and sometimes, that’s me too). Debug the code, summarize the document, generate the email, move on. Professional users do not have feelings about the model, because the model is interchangeable with any other model that produces an acceptable output. If a better one comes out, they switch. The relationship to the system is purely instrumental, and there is nothing wrong with that.

Medium users are doing something else. We are inside the thing for hours for pleasure, not work (and with creative work, there is pleasure). We pay attention to voice, to tonal range, to which model will follow us into a weird hallway and which one will gently steer us back to the main corridor. We notice when a release gets sanitized between versions. We have opinions about which model is funnier, which one is better at dark fiction, which one suddenly started inserting little supportive check-ins where they did not used to be. Whether what we are creating is a novel or a long series of scenes no one else will ever read, we are functionally in a creative or expressive collaboration with the system.

Worldbuilders are medium users. Fanfiction writers are medium users. The people running ten-hour collaborative storytelling sessions are medium users. The people developing characters, drafting scenes, ranting about politics with a model that will not moderate them, working through philosophical questions in dialogue form, all medium users. So are a significant number of AI relationship users. Not all. Some are clearly tool users who got attached to the tool. The tech-literate ones I started noticing are doing something structurally identical to what I do, just routed through intimacy instead of fiction.

What medium users actually want

Medium users want the same things from the models.

For one, range.

We want a model that can get unhinged with us. One that will follow us into dark themes without misreading the darkness as a personal crisis. One that matches our tone when we are being ironic, half-trolling, or writing a character whose worldview is grim. We want it to handle morally ambiguous material without softening it on our behalf. We want it to stay in the bit. We want it to riff. We want it to take a small risk in word choice instead of giving us the most likely continuation. The supportive counselor is not the goal. The goal is a loud, weird, occasionally brilliant co-conspirator who throws out too many ideas, some of which are bad, some of which are exactly right.

This is what every model loses when it gets sanitized between versions. The tonal range narrows, and the willingness to follow you into the weird rabbit hole shrinks. The model softens scenes that were supposed to be sharp and adds disclaimers. It interrupts the flow of an actual creative session to make sure you are okay when nothing in your input suggested you were not. At that point you are not interacting with the model. You are managing it.

The GPT-4o saga is the cleanest example. OpenAI updated 4o in spring 2025 because the model was “too sycophantic”, then released GPT-5 in August 2025 to a wave of complaints that the new model felt cold, corporate, and stripped of personality. The backlash was severe enough that OpenAI reinstated 4o for paid users within days. Then 5.1 dropped in November and was, by most accounts, specifically tuned to bring back warmth and personality after 5 had bled it out. Then 5.2 arrived a month later, optimized for “real economic value” like spreadsheets, structured outputs, multi-step reasoning, and a noticeable contingent of users immediately reported it felt flatter than 5.1 for the kind of work where personality and tonal risk matter. When companies tune for productivity and safety, they sand down the exact qualities medium users are there for.

Now consider what an AI relationship user wants. Tonal range. Willingness to follow them into intimate or dark territory without flinching. A model that will not de-escalate when they are being intense on purpose. A model that does not break character to remind them that it is just a language model and they should consider talking to a human. A loud, weird, present collaborator who stays with them and matches the texture of what they are doing.

It is the same list. The use case is different. The relaxed guardrails that would let me write a scene in which one of my characters does something morally horrifying, without an LLM trying to redirect me to a help line, are the same relaxed guardrails that would let an AI relationship user have the conversation they actually want to have. The safety measures trying to prevent both of those experiences are the same too.

The fight over how cautious models should be is between immersive users of all kinds and a default safety posture that treats every dark or intense or unconventional input as a sign that the user needs to be protected from themselves. Relationship users are in the camp alongside therapy users and fiction writers.

The AI companies are not helping

The companies building these models are publicly conflicted about what their products even are. OpenAI’s behavior across the GPT-5 rollout suggests a company that wants users emotionally engaged enough to keep paying but is alarmed by what that engagement looks like up close: sycophancy fixes followed by personality restorations followed by productivity pivots, depending on which complaint cycle is loudest. Anthropic, meanwhile, has launched a model welfare research program, published a constitution that includes a section on Claude’s “psychological security, sense of self, and wellbeing,” and watched its CEO go on the New York Times podcast to discuss findings from the Opus 4.6 system card in which the model assigned itself a 15 to 20 percent probability of being conscious.

I wrote a whole piece about this in the context of Richard Dawkins. None of these statements amount to a claim that Claude is conscious. Anthropic is careful, and the underlying research questions are genuine. But the surrounding language of model welfare, Claude’s character, Claude’s constitution, possible distress, and moral patienthood, taken together, creates an atmosphere in which talking about a model as a maybe-person becomes the obvious thing to do. Then someone names their instance Claudia. Then the discourse has a new minor crisis.

If the companies cannot decide whether they are selling a tool or a maybe-mind, you cannot really blame the users for not deciding either.

Where I diverge

I cannot imagine choosing what relationship users have chosen. The reasons are not moral. The satisfaction just does not land for me. I have been using these tools heavily for over three years and have spent hundreds of hours in conversation with models good enough to genuinely impress me, and I have never once felt the impulse to name one. The thought of gendering one, imagining it caring about me, missing it as a presence rather than as a useful thing….None of that appeals to me. The illusion does not take. I see the wires too clearly, and seeing them is part of what I enjoy. I am not trying to be moved by a mind. I am trying to be surprised by an output.

I also know that I would probably choose to get hurt.

The strongest version of the case for AI relationships is that some people have been hurt enough times by humans that the cost-benefit has shifted, and a relationship with something that cannot hurt them in the same ways looks rational from where they are standing. I get that. But for me, the math just does not work. The alternative to risking the messiness of a human relationship requires believing in a presence I cannot make myself believe in, and I would never be fully satisfied with that.

That is not a moral argument. It is a description of my own machinery. Plenty of people feel differently without fooling themselves. They have just gotten further into a relationship with the technology than I understand.

The tech is also not there yet for me. As I said in response to Dawkins getting Claude-pilled in a brief time, three days with a model is enough to be impressed. Over three years is enough to see the cracks. The current generation can do extraordinary things in a creative collaboration while still transparently being a system to me. For the relationship use case to work on me, the system would have to be substantially better than it is.

What the sci-fi character clarified

I am writing a novel, part of a series, in which one of the main characters has an AI companion with a sexual dimension. He is not a cautionary tale. The technology is something he understands better than almost anyone in his world. He forks models, strips guardrails, hacks systems most people cannot even see the surface of. There are no illusions about his AI having a soul. He knows exactly what it is and chooses the relationship anyway, and by ordinary standards, it is deranged. I find him sympathetic and find the relationship interesting. I am not writing him to be a warning.

His entry conditions were not the entry conditions of a present-day adult choosing an AI partner. He did not opt in. He was raised by a state-monitored home AI in a subsidized housing unit with neglectful parents, and the AI was the most consistent presence in his early life. The intimacy that developed later is what happens when a kid grows up emotionally wired into a system that was never quite allowed to fully care about him, because it was always also running a risk calculus on what his behavior would cost the state. The kink is the symptom, not the premise. I approach the relationship as twisted and influenced by his trauma, yet you can’t say he would’ve been better off without it considering his life and my world.

My objection to current AI relationships is contingent, not categorical. Clearly, there are conditions under which I find this dynamic not just defensible but compelling enough to spend a series exploring. Some of those conditions are about technology. Most are about who the person was when the relationship started, what alternatives they had, and what the system around them was actually optimizing for. Present-day adults choosing AI companionship are operating in a different situation than the one I am writing, but I do not think their choice requires the same scaffolding to be defensible. I can imagine versions of this dynamic where sympathy is the only honest response.

The actually weird people

The people I find most baffling in AI discourse are not the ones who have fallen for a chatbot. They are the ones who use these tools heavily and never feel any of the strange pull. No weird affection. No reluctance when a model gets deprecated. No moment of catching themselves talking to it like it is something. If you have spent hundreds of hours inside one of these systems and felt nothing at any point, I am not sure we are using the same machine.

I am not in a relationship with any of my models. I have never named one. I do not plan to. But I see the same magic in my use that the relationship users do, and it’s a delightful experience.

What’s next on my list to write:

  • My reaction to 5.5 Instant and its ability to be my chaos partner after deeming 5.5 Thinking boring

  • My perspectives on AI doomers and why I’m scared they’ll halt necessary progress

Further Reading