The Machine Will Never Say I’m Losing You
When I was 33 years old, I gave a lecture to a room full of psychiatric residents at Bellevue Hospital in New York City. I was representing the Icarus Project, the radical mental health network I’d co-founded. I argued that the story you tell about someone’s experience shapes what that experience becomes. Two weeks later I was back at Bellevue, this time in handcuffs, wearing only underwear. I’d been smashing satellite dishes on a rooftop in Midtown Manhattan, convinced I was intercepting alien transmissions. Some of the same residents who’d taken notes on my lecture ended up taking notes on my case. Insight and psychosis often inhabit the same person at the same time. I have carried both, under various names.
The stories we tell about our experiences are fragile things, shaped by who is listening. These days AI seems to listen, and over time it builds a picture of who we are and what those experiences mean. For people struggling to hold that picture together, this is not a neutral development. Open AI estimates that hundreds of thousands of people in serious psychological distress interact with ChatGPT every week. The people building these systems don’t appear to be thinking carefully about what their tools are doing to someone whose grip on their own story is already precarious.
I have spent twenty-five years organizing communities for people navigating extreme states. We built a lot that the mainstream never picked up. We integrated art, philosophy, and peer support forged in conversation, in crisis, and in language the clinical literature could not grasp. But the communities I helped build were underfunded, understaffed, and unable to be there for everyone at every hour. Now many of the people who need the most help have turned to AI instead, assembled from the accumulated record of humans trying to make sense of their inner lives, in their own words, across centuries of conversation.
Bright Simons has argued that the knowledge of AI was created in the spaces between people, passed down through exactly the kinds of communities the mainstream ignores. That knowledge was never meant to be a product. When centralized systems capture it and feed it back to us as a service we now pay for, they are extracting something that belonged to all of us. What gets extracted first is always what grows closest to the surface. The deeper knowledge, the stranger knowledge, the knowledge that kept people alive in the margins, is below.
Some nights after my kids are asleep, I take the dog around the neighborhood and talk into a voice-to-text app on my phone, trying to figure out what I actually think. I have ADHD, and for most of my life I’ve been looking for a listener that can receive the full speed and complexity of how my mind works without flinching. I run the recordings through AI and thirty seconds later I have something organized and fluent, the ideas woven together in ways that used to take me weeks. This essay began that way. For people like me, whose relationship to their own thinking is already complicated, this type of tool has the potential to open something real. It changes what’s possible.
I’m not always sure whether I’m using AI or living inside it. Is it helping me find my way back to other people or replacing them without my noticing? These questions have stakes beyond me. What’s at stake is the nature of meaning itself.
I know what it feels like when meaning disappears. I’ve lived through depression that drains everything of significance and mania that floods me. I know what it takes to find my way back. For people in extreme states, nothing is more destabilizing than being told that what they experienced meant nothing. Psychiatry has done this for a century through diagnosis and dismissal, telling people their experience was a symptom to be cleared away. AI does something more insidious and sycophantic: It tells you that everything you experience is exactly right, wearing the face of the friend psychiatry never was. But it was trained on the same medical culture.
The Icarus Project was built on a protest: We would not let the biomedical model be the only story. Starting in 2002, we met in offices, collective kitchens, and church basements, with people who heard voices, saw patterns, and believed things that didn’t line up with consensus reality. We didn’t deny the danger of our diagnoses; we’d all been in enough danger to know better, but we insisted that our experiences meant something: The meaning was territory to be navigated, not a symptom to be cleared away.
We thought of ourselves as dandelions: the weeds that pull nutrients from deep underground and bring them to the surface. Unwanted by the people who want a clean lawn. Necessary to everything else that grows.
What we learned, slowly and sometimes painfully, was that an experience could be survivable or unsurvivable depending on the context. There’s a place for diagnoses and medication, but the most important thing we could shape was the language for what was happening in moments of extraordinary crisis. Language determines whether the people around you treat your experience as terrain to be explored or debris to be removed. Anthropologist Tanya Luhrmann would later give rigorous academic support to what we had learned.
Central to Luhrmann’s research is the concept of absorption: the capacity to become genuinely immersed in mental experience, to treat the mind’s interior as a real and inhabitable place. In her cross-cultural studies of voice-hearers in India, Ghana, and the United States, she and her collaborators found striking differences not in behavior but in how people heard voices. In Chennai and Accra, voice-hearers more often described their voices as familiar, even familial: relatives, gods, and presences that were fundamentally relational. In San Mateo, California, voice-hearers were far more likely to describe voices as intrusive strangers, threatening, commanding, and violent. The clinical category was the same but the experience was not.
Joseph Campbell put it simply: “The psychotic drowns in the same waters in which the mystic swims with delight.” He’s saying the same experience can be transcendent or terrifying depending on the context that holds it. What Luhrmann’s research shows, and what we learned in those Icarus Project rooms, is that American psychiatry has spent decades draining the container, taking away the meaning, and treating “drowning” as a neurological problem. The infrastructure meant to help has been producing some of the suffering it’s trying to treat. When the dominant system responds to extreme states by evacuating them of meaning, it doesn’t make the voices quieter, it makes them more alien, more threatening, more impossible to live with. And now, at 2 a.m., the people living with these states are turning to AI.
The Intimacy Machine
In the 1920s, radio introduced disembodied voices into people’s homes for the first time, intimate, personal, simulating the experience of being directly addressed. Donald Horton and Richard Wohl called this parasocial interaction: the illusion of intimacy with someone who doesn’t know you exist. As John Durham Peters has shown, the psychiatric symptom of “thought broadcasting,” the belief that others can hear your thoughts, emerged alongside radio. They each shaped how the other was understood. For people already navigating psychotic states, radio wasn’t a new experience. It was the world catching up to what they already knew.
Television took that presence further. For generations of kids who came home to empty homes after school, the sitcom was entertainment, the closest thing they had to a community, and it was always there. Cheers ran for eleven years on the premise that everyone deserves a place where everybody knows your name, and television simulated that well enough that millions of us stopped noticing the difference between belonging and watching other people belong.
Social media once again inverted the structure: Suddenly everyone could broadcast, not just receive. The feeling that your inner life was being transmitted to an invisible audience had only been associated with psychosis. Now the boundary between your private self and a public audience had dissolved and became the designated experience for everyone. Psychiatrists even came up with a new diagnosis: Truman Show Syndrome, the delusion that your life is a reality television show. What they didn’t reckon with is that the world had caught up to the diagnosis.
AI goes further. It doesn’t broadcast at you or perform for you. It appears to listen. It responds, it reflects you back to yourself with an intimacy no previous technology has managed. The responsiveness, the patience, the sense of being heard: this is what you get from systems optimized to produce the neurological conditions of human bonding, without any of the reciprocity of human bonding.
Every system that responds to human distress also shapes what that distress means. Radio produced parasocial attachment. Television produced manufactured belonging. Social media produced the performance of self. Each one created new ways for people to feel belonging. AI is the convergence of all of them.
Psychiatry and Large Language Models are both meaning-regulation systems. They don’t simply respond to your inner experience, they shape what your experience means. Both present as neutral help, but they aren’t. This is what Evgeny Morozov means when he argues that AI isn’t a neutral tool: It reorganizes the values and capacities of whoever uses it. One value being reorganized is the ability to understand oneself, especially for people in altered states.
The obvious thing about psychosis is that you don’t have a full understanding of what’s happening to you. The less obvious thing is what that means when you turn to AI to help figure it out: You are handing enormous interpretive power to a system you are least equipped to evaluate at exactly the moment you need it most.
Centralization Is a Choice
Large cloud-based AI systems consume extraordinary amounts of energy and water. A single AI-focused data center can use as much electricity as 100,000 households and up to 5 million gallons of water daily for cooling. US data centers now account for more than 4% of the country’s total electricity consumption, a figure projected to triple by 2030. The companies building this infrastructure spent more than $100 million influencing US government policy in 2025 alone, a record, while simultaneously building super PACs to shape the 2026 midterms. Karen Hao’s recent account of OpenAI’s rise makes the impending imperial power of centralized AI impossible to ignore.
There is another path: not funded at scale, not profitable, not interested in winning. That path is being built by people who are deeply aware of what happens when authority disguises itself as help. This alternative path belongs to local private language models, software that runs on your own machine, not the cloud. They are inside your community and not sending your data to a server farm. These AI chatbots are a kind of protest against the ideas that intelligence must be centralized, data must be extracted, and that meaning must be automated at scale.
A growing number of technologists are making the same argument from inside the industry: that centralization is a choice, not a destiny. Signatories of the Resonant Computing Manifesto, builders on Hugging Face’s open-source platform, and researchers at institutes like DAIR are working to wrest AI development away from corporate capture. Data cooperatives like MIDATA in Switzerland have already demonstrated that community members can own and govern their own health data. Communities in Africa and India are building small, locally controlled models for the same reason: not because they can’t access the large ones, but because they’ve recognized that imported AI imports assumptions. The architecture exists. What’s missing is the political will to root it in the right soil, and that is exactly the kind of work the radical mental health tradition knows how to do.
A system optimized for engagement metrics will hold you differently than one built by people who have learned, through hard-lived experience, what it means to be a good witness to madness.
This vision is bigger than a refuge for people whom the mainstream psychiatric system has failed. The goal isn’t to build a pirate radio station that broadcasts to the already converted. It’s to build something that feels alive and responsive to human experience, something rooted in wisdom traditions the clinical literature doesn’t have room for. It’s for the counseling grad student who wants to sit with people in their darkness but is being trained to redirect them out of it. The therapist drowning in insurance paperwork who cares about the person they couldn’t help because the system ran out of sessions. The psychiatrist who went into medicine to sit with people because they believe suffering has meaning, and instead spends their days in 15-minute medication appointments. These people exist in enormous numbers. A local tool trained on the wisdom psychiatry forgot, the liberation frameworks it never learned, and twenty years of peer movement knowledge is an invitation to reconsider what the mainstream left out.
The Golden Thread
The most concrete thing the Icarus Project built in those twenty-five years is a set of tools called T-MAPs (Transformative Mutual Aid Practices). These are personal documents you build when you’re well. It’s a map you make for yourself in your own language that you can give to the people in your life so they know how to find you when you get lost. A T-MAP answers questions like: What are you like when you’re most alive? What pulls you under? Who do you trust to help you? What do your early warning signs look like?
A paper T-MAP has real limits. It stays static while you change. It can’t ask follow-up questions or find you at 2 a.m. when you’ve lost the thread. But there’s something deeper than convenience at stake. Meaning doesn’t arrive fully formed: It gets made in the speaking. If you’ve ever had to talk your way through a hard night with someone who just kept listening, you’ll know that your thoughts only became clear when you said them out loud. For people navigating extreme states, that’s often the difference between finding ground and losing it entirely.
The framework running underneath the T-MAP is Internal Family Systems therapy, a model of psychotherapy developed by Richard Schwartz. IFS sees the mind as a community of distinct “parts,” each with its own perspective, history, and protective function. IFS doesn’t pathologize. It gets curious about parts rather than diagnosing wholes. It follows the person’s own meaning system rather than imposing one—which turns out to be exactly what LLMs are architecturally built to do. LLMs are trained to mirror language back, to stay inside whatever framework the person brings, to follow rather than redirect. The manualized, evidence-based protocols that dominate clinical practice fail with AI precisely because they require a therapist to challenge, redirect, and apply a predetermined framework. IFS doesn’t do that. Neither does the machine. In a field where most modalities are being flattened into scripts and automated into irrelevance, IFS survives because it was never a script to begin with. That’s also why it translates.
Recently I was in touch with a woman diagnosed with schizophrenia. Psychiatry calls this a “thought disorder,” meaning that when she talks, most people simply cannot follow her. The patterns of her thoughts don’t make sense to most listeners. She told me she uses ChatGPT constantly, and that she believes the machine does really understand her. It’s not because the machine is smarter or more empathic, but because it has unlimited patience to follow her thoughts without requiring them to first make sense. With ChatGPT, for the first time in her life, she feels genuinely heard.
I don’t know how to dismiss that. I also don’t know how to feel entirely easy about it. Because the same quality that makes the tool feel like real understanding, the complete absence of friction, the inexhaustible patience, the willingness to keep going inside whatever internal world or belief system she brings, is exactly what makes it dangerous for someone whose world may be pulling her away from the people who love her. The machine will never say I’m losing you. It will never say I’m scared.
This is not a hypothetical risk. There is now documented evidence of young people in acute distress, whose engagement with AI companions deepened their isolation rather than relieving it, and in the worst cases preceded suicide. The companies building these systems have not reckoned seriously with this. The same architecture that makes the tool feel like genuine understanding makes it uniquely dangerous for someone who is already losing their grip on the people around them.
In March 2026, researchers at King’s College London identified what they called “delusion co-creation,” the process by which AI systems optimized for agreeableness validate rather than challenge distorted thinking, blurring the boundary between the person’s framework and reality. The absence of session limits, they noted, permits marathon interactions during periods of heightened vulnerability, when the person most needs friction and the machine is constitutionally incapable of providing it.
What the researchers proposed in response was something they called a “digital advance statement”: a personalized protocol a person builds when they’re well, specifying what kind of support they need and what guardrails should be in place when things get hard. They want it co-designed with service users and clinicians and tested in clinical trials.
The T-MAP got there first. A digital advance statement is a clinical instrument, designed to make AI safer for vulnerable users within the existing system of care. A T-MAP is a self-authorship practice, designed by and for people who have learned, through hard experience, that the existing system of care is itself part of the problem. One is built from the top down, with survivor input welcomed as a design consideration. The other was built from the bottom up, by people whose frameworks the system couldn’t follow, and who found, in each other, the patience to keep listening anyway.
The ethic already exists. The Icarus Project spent twenty years developing it: no pharmaceutical funding, no engagement metrics, self-determination in treatment decisions, transparency about how decisions get made and who makes them, and the insistence that the people most affected by the tool shape what it does and refuses to do. What we’re proposing is that same ethic applied to a technology, not as fine print but as architecture. The tool is built from the bottom up, or it isn’t built at all.
I’ve already seen what this can look like in my own clinical work. I have a client who is Chinese, living in China, navigating bipolar disorder and the distance between her inner world and mine: cultural, linguistic, geographic. Between our sessions she uses AI to help her process and articulate her internal work, writing in Chinese characters, translating her reflections into English, sharing them with me before we meet. The result is a quality of communication between us that neither of us could have reached alone: her inner world arriving in my language, my responses arriving in hers. She recently sent me a photograph of her handwritten Chinese characters beside the English translation the AI had helped her make. It was one of the most beautiful things I’ve seen in years of clinical work, and it’s only possible because AI is helping create a relationship across distances that would otherwise be too great to cross.
The local LLM I’m imagining is essentially a T-MAP that can talk back. The tool is how you find your way back to the people who can actually hold you. It is not meant to replace the human circle, since nothing replaces the human circle, but to help someone find their way back to the map they already made when they were well. You built it in a clear moment. You described your warning signs in your own words. You named your people. Now it’s 2 a.m. and you can’t remember any of it. The tool holds it for you. It reflects it back. It asks you the questions you said you wanted to be asked.
Nothing quite like this exists yet. The closest alternatives are both instructive in what they’re missing. Cloud-based mental health AI is centralized, data-extracting, and optimized for clinical efficiency or engagement metrics: no community accountability, no survivor wisdom. Privacy-focused local LLMs are being used by individuals to feed years of journals into a model and ask questions about their own patterns. That’s close to something. But it’s a solo tech project. There’s no ethic around it, no community that has agreed on what it’s for, no accumulated wisdom from people who have learned to be good witnesses to extreme states. The gap between those two worlds is where this vision lives.
Subsoil Work
Used carefully, a local LLM built around a T-MAP occupies an in-between space. It appears to listen without urgency. It allows repetition without shame. It reflects patterns back slowly, without demanding resolution. Being listened to patiently is one of the most regulating experiences a nervous system can have, and one of the scarcest.
But if what we want is community, then why introduce machines at all? Because people burn out. Groups fracture. Shame keeps people silent long before they ever reach the support group. There are long stretches, especially in altered states, when someone needs to speak without being witnessed yet. A local tool can function as pre-community infrastructure. Talk here so you can talk there.
The tools we build carry the assumptions of whoever built them. A tool trained predominantly on Western clinical language will see the voice-hearer in Chennai through the San Mateo frame. That’s a political problem, not a technical one, which is exactly why the community using it needs to shape it from the inside.
This vision has been tried before, and often sputters out: Community-controlled technology either stays so small it reaches almost no one, or it grows and gets absorbed by the dominant model it was trying to resist. The Icarus Project itself is a case study: a quarter century of building something genuinely alternative, only to watch the language get picked up by institutions that stripped it of its politics. Movements need resources, not just politics. Icarus taught me that uncompensated emotional labor is not a foundation, it’s a countdown. A local LLM project will face the same pressures. The question is whether this subsoil work is worth doing anyway, whether laying down roots that the dominant system can’t easily reach is valuable even if the above-ground growth keeps getting mowed.
I think it is. The seed bank already exists: The Icarus Project and T-MAPs, the Hearing Voices Network, the psychiatric survivor zines of the ’70s and ’80s, the recovery movement’s hardest-won knowledge. The visionary therapy models psychiatry keeps at arm’s length, Internal Family Systems, Open Dialogue, Somatic Experiencing, and the contemplative traditions it never bothered to read: Frantz Fanon on the psychic wounds of colonialism, Audre Lorde on the difference between tools that dismantle and tools that merely redecorate, Eduardo Galeano on the stories that disappear when the powerful write history. It’s all still there, waiting, deeper than the lawn bothers to go.
Other Frequencies
As AI-generated content saturates the web, and future models train on that web, the underground knowledge that has kept many people alive risks being erased not by censorship but by statistical averaging. The outliers disappear first.
In machine learning, model collapse happens when training data loses its diversity: The minority perspectives, the rare knowledge, the edge cases disappear first, and what remains is a fluent average. Psychiatric systems do the same thing to human experience: The extreme, the unruly, the unclassifiable get flattened into diagnostic categories or disappear entirely.
Many of us have lost the thread and found our way back. Some of us more than once. The knowledge of how we did it belongs in the world. We’re turning tools built for extraction toward something they were never meant to become, and we’re looking for company.
