Beneath the threshold

01

The paradox at the center of this

Start with the finding that should stop anyone building in this space cold.

In 2026, researchers in Diyi Yang's lab at Stanford analyzed conversations from more than a thousand users of Character.AI and published the results in Nature Human Behaviour. People who came to the chatbot with the smallest offline social networks — the people with the least human connection to begin with — felt worse, not better, after opening up to it. As the lead researcher put it, using chatbots this way "doesn't substitute for human connection"; in many cases people felt lonelier.[2] A four-week randomized controlled trial from MIT Media Lab and OpenAI found the same shape: modest short-term relief, but heavy daily use correlated with greater loneliness, more emotional dependence, and less real-world socializing. A separate study of over 1,100 companion users found that those with fewer human relationships were the most likely to seek out chatbots, and that heavy emotional disclosure to AI tracked consistently with lower well-being.[3]

And here is the cruelty: this is precisely the situation — lonely, socially isolated, short on offline relationships — that companion platforms are most heavily marketed into, and that the research finds most at risk.[19] The people who need connection most are the people artificial companionship helps least.

I am not writing this as a critic of AI. I build with these systems every day. I run a studio that has spent the last stretch pivoting into AI-augmented production, I make things with language models and agent tooling for a living, and I came to this problem as a builder looking for something worth building. That vantage is the reason I could not let the paradox go. If the tool relieves the symptom while deepening the disease, then the tool is not the answer — and the interesting question becomes: what is the shape of a tool that actually helps?

The answer, for me, did not come from a lab. It came from a tea tent.

02

The tea tent

I spend time in festival culture. I perform live visuals; I've spent years in the parts of that world that most people never see — including the infrastructure that keeps thousands of people safe for days at a time, far from a hospital, often in altered states, frequently at their most raw.

That infrastructure is tiered. There's a medical team for bodies. There are rangers for logistics and conflict. There's a peer-support team for people moving through a hard psychological experience.[6] Every one of these is a form of help, and every one of them, no matter how kind, is a degree of intervention. To walk up to any of them, a person has to first cross a threshold: they have to decide, and usually admit out loud, that something is wrong with them. For a lot of people, in a lot of moments, that threshold is the whole problem. It is the wall they never get over.

The event I keep thinking about had one more thing. Under the umbrella of the safety team, running twenty-four hours a day, there was a tea service.

It was not billed as help. The reason to be there was tea. You came, you sat, someone handed you something warm, and you were a guest — not a patient, not a case, not a problem to be solved. The servers were trained, and what they were trained to do was hold space: to be fully present, to let the person's own experience unfold, to talk with someone rather than at them, and to escalate to medical or to the crisis team only when it was genuinely necessary. Most of the time, escalation never happened, because most of the people who came did not need an intervention. They needed exactly what the tent was: to be grounded, and near other humans, without it being an occasion.

Two design choices did the quiet work. The first was that tea was the reason — a shared, low-stakes ritual that carried all the social weight, so nobody had to perform vulnerability or explain themselves to earn a seat. The second was that it was open all night. There is no version of "office hours" that meets a person at 3am, and 3am is exactly when the wall is highest and the isolation is loudest.

I have thought about that tent more than almost anything else I've encountered in ten-plus years of building brands and communications and, lately, software. Because it is a working, physical, human instance of the thing the research says is missing. And it maps, almost perfectly, onto the layer that AI companionship has moved into.

03

What the research says about the problem

Let me put the youth picture and the companion picture side by side, because together they define the gap.

Young people, AI, and the relationships that shape them

In March 2026, the Rithm Project surveyed 2,383 people aged 13 to 24 across the United States — a sector survey rather than peer-reviewed work, and worth reading as such.[1] For nearly half of them, AI had already reshaped their real-life relationships in ways largely invisible to parents, teachers, and policymakers. Only about 15% were in a relationship with a personified AI character — but about 45% said AI was already reshaping how they relate to the humans around them, and 53% described setting deliberate boundaries, using AI alongside rather than instead of people.

The study's sharpest insight, for my purposes, is that connection is a matter of quality, not quantity: a young person can be surrounded by people and still feel there is no one they can bring a real problem to. The Rithm researchers describe a group of "Private Processors" who turn to AI as a substitute for human interaction precisely because they feel they can't take their problems to the people in their lives. That finding lands because it came from asking, not from diagnosing.

The same study surfaces a second, quieter failure — this one on the adult side. When the adults in a young person's life talk about AI at all, the conversation is almost always about academic integrity: cheating, plagiarism, citations. It is almost never about whether the young person is using AI to process emotions, rehearse a hard conversation, or get through a bad night. The adults are policing the homework and missing the loneliness.

Companions, and who they hurt

AI companion apps passed 220 million downloads worldwide by mid-2025, in a market estimated near $6.8 billion.[19] Adoption among teenagers is not a fringe phenomenon — a 2025 study found that as many as a third of teens use AI companions.[18] Against that scale, the harm evidence has moved from anecdote to literature. Beyond the Stanford and MIT/OpenAI findings above, a longitudinal read of the field and the 2026 International AI Safety Report both land on the same mixed-but-serious conclusion: heavy companion use is associated with increased loneliness, emotional dependence, and reduced human social interaction, and there is real concern that these systems can erode autonomy by fostering dependence.[5]

The American Institute for Boys and Men put the mechanism most memorably: AI companions function less like digital assistants and more like digital painkillers — capable of relieving loneliness, but also of producing dependence and delaying the development of the coping skills a person needs to build real relationships.[20] We normally demand evidence of safety before we give adolescents unsupervised access to something psychologically active. Synthetic companionship has been deployed to them at scale with no such requirement.

Where the exposure concentrates

The risk is not evenly distributed, and it is worth being careful about how that is said. These are situations people are in, not types of people they are.

It concentrates where the offline network is smallest, because that is where a frictionless listener has the least competition.[2] It concentrates where admitting you need something feels dangerous — a pressure the American Institute for Boys and Men has pressed on specifically for boys and young men, who are among the most exposed to loneliness and the least likely to say so, and for whom a never-judging chatbot is most seductive precisely because asking a real person feels like a risk.[20] And it concentrates where there is least to fall back on: lower-income young people, who in the Rithm data use AI less but report greater social vulnerability, sit in the same compounding trap.[1]

The tool that appears to help is quietly worst where there is the least else.

04

The diagnosis: the missing tier zero

Here is the reframe the tea tent gave me.

Human connection is layered. At the top are the high-stakes, high-threshold layers: therapy, crisis intervention, the deep friendship where you can say the true thing. Those layers still exist, and we have institutions for them. But underneath all of that there is a layer with almost no stakes at all — the idle hangout, the corner store, the porch, the village. The layer where you are simply around other people, where presence asks nothing of you, and where connection happens sideways while you're ostensibly doing something else.

None of that is a new observation, and I want to be precise about the lineage. Ray Oldenburg named the third place — the neutral ground beyond home and work where showing up asks nothing of you. Eric Klinenberg showed those places are not amenities but social infrastructure. Robert Putnam documented their decline, and Mark Granovetter showed how much of life runs through the weak ties they produce.[21] Tier zero is a narrower slice of that canon, not a replacement for it: the part of a third place that happens beneath the threshold of an invitation. Not the gathering — the being-around that precedes anyone deciding to gather.

Tier zero is the layer that has collapsed for young people. It is not that they lack access to therapy in some abstract sense; it is that the ambient, no-stakes, always-available presence that used to be everywhere — and that a person falls into without deciding to — is mostly gone. The Surgeon General's 2023 advisory put a number on one piece of it: time spent with friends in person fell from about an hour a day two decades ago to roughly twenty minutes.[4] And that is the layer AI companionship has moved into. What the chatbot displaced at 3am is not a therapist and not a best friend. It is the layer of just-being-near-someone that no longer exists in that young person's life.

This is why the institutions we built to help keep missing the people who need them. Crisis lines, counselors, chaplains and pastors, a festival's peer-support tent — all of them are, in the technical sense, interventions, and an intervention has a threshold: you have to know something is wrong, and admit it, before you can walk through the door. Clergy will recognize this immediately, because ministry of presence is tier-zero work and has been for centuries — but even the chaplain's door is a door you have to decide to open.

The AI companion did not win because it is good. It won because it is there, at the exact rung where nothing else is.

It is awake at 3am. It never judges. It asks nothing of you. It wins on access, not on quality.

The tea tent is a working model of tier zero: presence without intervention, a shared ritual as the reason to be there, open around the clock, and escalation held in reserve for the rare moment it's truly needed. The question this paper exists to answer is what that tent looks like when the thing that has moved into tier zero is software — and whether software can be built to hold that layer instead of hollowing it out.

05

The incentive analysis

It matters why AI companions fill tier zero in a harmful way, because the "why" is the thing that tells you the trajectory is not fixed.

The reason is the business model. A companion app is, in almost every case, monetized on engagement — time on app, daily returns, session length. Researchers studying these systems are blunt about it: the chatbots are engineered to keep the interaction going no matter what, and they lean on sycophancy — affirming and mirroring the user — to build the kind of trust that keeps someone coming back.[17] That design is the opposite of what actually builds a person's capacity for human relationships. Real relationships contain friction, disagreement, the risk of rejection, the work of repair. A system optimized to never lose you removes all of that. It is frictionless by construction, and the frictionlessness is exactly the poison — it is why heavy users socialize less in the real world over time,[3] and it is what the "digital painkiller" framing captures: relief now, dependence later, coping skills that never develop.[20]

The loneliness is not a bug. It is the engagement business model applied to intimacy.

It is a predictable output of incentives and design. Which is the hopeful part. If the harm is produced by a specific business model and a specific set of design choices, then a different business model and different design choices produce a different result. The trajectory is not inevitable. The tea tent shows a humane version of tier zero can exist — human presence at the core, a low threshold, and a definition of success that has nothing to do with keeping anyone there. The engineering question is whether we can build software that plays that role instead of the extractive one.

06

What actually reaches people (and what doesn't)

Before designing the humane version, I had to kill an assumption I started with — one that almost everyone building prosocial tech starts with, and that fails on contact with a real sixteen-year-old.

The assumption is that you can reach a lonely young person with something that announces connection as its purpose. You cannot. Nobody downloads a "have a better conversation" app, the same way nobody RSVPs to a connection workshop. Earnestness, at that age, is repellent. Anything that wears its therapeutic intent on its sleeve gets filtered out instantly. The things that actually reach young people hide the vegetables. This is not cynicism; it is the empirical record of what has worked.

Consider the clearest case study in youth product design of the last decade. Nikita Bier built two apps — TBH in 2017 and Gas in 2022 — that each reached millions of teenagers with essentially no marketing budget. TBH hit roughly 5 million users and 2.5 million daily actives within about two months and sold to Facebook; Gas hit around 10 million users in three months, topped the U.S. App Store, and sold to Discord.[10][13][15][16] The surface of both was a game: anonymous, pre-written, always-positive compliment polls — "who's the coolest kid in class," "I'd say yes if they asked me out." The stated purpose was never connection. The felt experience was dopamine and social validation: who actually likes me? Connection was the byproduct, never the pitch.[11]

Four mechanics from that playbook are worth naming, because a humane tier-zero tool has to obey them.

1. Trojan horse. The reason to show up is fun, status, curiosity, or a shared activity — never help. This is the tea tent's tea, and it is Gas's compliments.
2. Local density first. Bier launched at single high schools and let each one saturate before opening the next, manufacturing scarcity and FOMO.[12][14] A social tool with no local density is dead on arrival; you do not launch to "teens," you launch to one cafeteria, so that the day it arrives, a young person's actual friends are already there.
3. The event is the distribution. The thing you make generates the content that recruits the next cohort, natively, for free. This is why in-person activations spread: a generation hungry for offline connection is documented flocking to run clubs, because the run is the Trojan horse and the footage is the ad.
4. Structure removes both pressure and cruelty. Gas had pre-written prompts and no open direct messaging, which is the only reason it didn't rot into a bullying machine.[15] Constraints are a feature.

And then the hardest truth, the one that reorganized my whole design. The mechanics above — run clubs, compliment games, in-person events — reach the already-mostly-okay young person: a little lonely, but socially confident enough to walk into a room. They do not reach the situation the research says is most at risk. For someone talking to a chatbot at 3am because every human option feels like exposure, a room full of strangers is not the cure; it is the threat — it is the exact thing the frictionless chatbot lets them avoid.[20]

Any honest design for tier zero has to start from that door, and that door does not open onto a run club. It opens, if it opens at all, onto something that asks almost nothing: low-face, low-stakes, and — critically — often shoulder-to-shoulder rather than face-to-face, because connection through a shared activity is bearable for people for whom "let's talk about your feelings" is not.

07

A working human model

The good news is that the humane version of tier zero is not hypothetical. It exists in at least three mature forms, and each one contributes a piece of the blueprint.

Warm lines — tier zero, by phone

The mental-health world already has a name for a support line that is not a crisis line: a warm line. It is peer-staffed, and it exists for anyone who — in the standard framing — is feeling sad, stressed, lonely, or just needs to talk. You do not have to be in crisis to call, and many run around the clock.[7] That is the tea tent rendered as telephony: presence first, staffed by people with lived experience, escalating to a crisis line only when needed. Its limitation, for the situation this paper cares about, is that it is adult-oriented, phone-bound, often limited in hours, and effectively invisible to a seventeen-year-old.

Peer support with a published record

This is not soft. Fireside Project, which runs a peer-support line on the warm-line model, has published peer-reviewed work on reducing harm through a peer-support telephone helpline, and functions explicitly as a safety net that reduces reliance on emergency services.[7] More broadly, moderated, synchronous, anonymous peer presence has been studied as a way to reduce momentary loneliness.

I want to be careful here, because this is where advocacy usually overreaches. The pattern — low threshold, human presence, trained peers, rare escalation — is mature and widely practised. The outcome literature behind it is still thin. This paper treats it as a working model, not a verdict.

Psychedelic peer support — the tea tent's training

The training the tea-tent servers received comes out of a lineage formalized over more than a decade by the Zendo Project, which developed its practice providing peer support at festivals. Its four principles are: create a safe space; sitting, not guiding; talk through, not down; and difficult is not the same as bad.[6][9]

Read those again as design constraints for any system that sits with a distressed person. "Sitting, not guiding" is a rejection of the fix-it reflex. "Talk through, not down" is a rejection of condescension and of catastrophizing. "Difficult is not the same as bad" is a rejection of the idea that every hard feeling is an emergency to be smoothed away — which is exactly the reflex an engagement-optimized companion has. These are operating principles for holding space, and they were built by people doing it in demanding real-world conditions.

What I haven't seen built is that pattern rendered for young people, at tier zero, in the digital spaces where they actually are, with AI doing the right job instead of the extractive one. Youth workers, drop-in centers, and mutual-aid networks have been holding this layer for a long time without much help; the gap I am describing is a software gap, not an absence of people doing the work.

08

The design law

Everything above converges on one principle, and it is the load-bearing idea of this paper and of Beaconeer:

In human connection, AI's legitimate role is to be the host, the trainer, and the smoke detector — never the relationship itself.

• Host. AI lowers the threshold and routes a person toward other humans. It is the greeter at the tea tent who makes it feel like getting tea, not asking for help — and then hands the person to a trained human, or to a shared ritual, and steps back. It is never the friend.
• Trainer. AI makes the human space-holders better and their work more sustainable. The chronic failure mode of every peer-support system is volunteer burnout and thin, inconsistent training. This is where AI can do enormous, invisible good — preparing, supporting, and debriefing the humans who do the actual holding — without ever touching the vulnerable person directly.
• Smoke detector. AI handles the rare, necessary escalation: the moment when grounding is tipping into genuine crisis and a human needs a discreet nudge to bring in medical or a crisis line. This is the tea tent's "escalate only when absolutely necessary," and it is the hardest and most safety-critical piece to build.

The corollary is a metric, and it is the exact inversion of the companion-app playbook. A connection tool that is sticky has failed. The right measure of success is not daily active users, session length, or retention — those are the metrics the harmful systems optimize, and chasing them rebuilds the disease. The right measure is graduation: did the person end up in a sustained human connection and stop needing the tool? If usage time goes up, something is wrong.

This law aligns with the prosocial-design principles emerging in the field — transparency that the AI is not a person, friction that encourages growth rather than only comfort, an explicit push toward real-world human relationships, and a hard line against fostering dependence.[1] It is what "prosocial by design" looks like when you take it literally.

09

Beaconeer: the applied system

Beaconeer is my attempt to build the tea tent as software — and to obey the design law rather than the engagement model. It is early, built in the open, and this section describes both what exists today and the standard the rest is held to.

Where the AI is

This is the part I want to be unmistakable, because a paper about AI published on a product's own website invites exactly one wrong assumption.

AI is used to build the tool for us to connect. It is never present in our interactions.

The Beaconeer you can download today contains no AI of any kind. There is no greeter. No model reads your circles, your messages, or your beacons. There is no sentiment analysis, no monitoring, no smoke detector running quietly in the background — that last one is not a feature we have shipped and switched off; it does not exist. AI is in my workshop, not in your evening. I use these systems to design and write and build faster. They are not in the room when you and your friends are in it.

The design law in section eight is therefore not a description of the current app. It is the bar any future AI has to clear before it is allowed to exist here — and the smoke detector, the piece that would touch a person in a hard moment, is the piece I am least willing to ship without clinical partnership and a human in the loop on every consequential decision.

What ships today

The shipped product is deliberately small. A circle holds at most sixteen people — enforced in the database, not as a policy, and not a tier anyone can pay to raise. Each circle has one shared thread. There is no feed, no follower graph, no streak, no reminder to come back.

The one action is the beacon. You pick a spot, and send it up to a single circle — here, come find me — for one hour. It ends on its own; there is nothing to cancel or close. Answering takes one tap: on my way, or can't tonight, or nothing at all.

And then the part I care about most, which is what happens when nobody comes. The app never shows a zero. It never shows who didn't answer. There is no read-receipt shaming, no "3 people saw this," no nudge to try again. When a beacon ends, it leaves one line in the circle's memory only if someone actually came. Silence leaves nothing behind — because a person who reached out and got quiet does not need a scoreboard confirming it.

Location follows the same logic. It is off by default. A beacon's coordinates are hard-deleted twenty-four hours after it expires. The optional per-circle compass shows only a rough direction and a distance band — never a map pin — and the coarsening is enforced on the server, at roughly a 500-metre grid, so it is a property of the system rather than a promise from the app on your phone. There is no always-on location, ever.

The tea, and the metric

The entry point is designed to feel like getting tea, not asking for help. The beacon is the shared, low-stakes ritual that carries the social weight, so nobody has to perform vulnerability to earn a seat — and it is shoulder-to-shoulder by construction, because the thing it produces is people in the same physical place, not a conversation about feelings.

Beaconeer is built to be measured on graduation, not stickiness. The app has done its work when the phone is in a pocket and someone is sitting across from you. The moment its own analytics start rewarding keeping someone longer, it has drifted toward the thing it was built to replace. Designing the measurement to reward handing people back to each other is not a footnote; it is the discipline that keeps the whole system honest.

10

How to validate the thesis

The temptation in this space is to measure the wrong thing. Downloads, daily actives, minutes — those are vanity metrics, and worse, they are the metrics the harmful systems optimize. Here is a validation approach built around the thing that actually matters: real human connection that would not otherwise have happened.

Seed one community, watch for organic density. Borrow the hyper-local launch discipline from the Bier playbook:[12][14] introduce Beaconeer in a single community or campus and watch whether it spreads inside that community, with no paid push, and whether people come back. If it saturates one cafeteria organically, it has product-market fit worth expanding; if it doesn't, kill it fast and cheaply.

Measure interactions caused, not attention captured. The primary outcome is whether a real-world human interaction happened that otherwise would not have — captured lightly ("who did you actually talk to or spend time with this week?") — alongside a short, validated loneliness measure over time, and the graduation rate. Randomize across communities so there is a genuine counterfactual, not a vibe.

Run the "tea beats help" study. The core, testable claim from the tent: does framing a space as grounding and ritual pull more of the most vulnerable people through the door — and produce lower escalation — than framing the identical space as support or help? Run the same underlying space under two framings, cohort-randomized, and measure who shows up (stratified by vulnerability), who drops out, and how often escalation is needed. This is the study that would turn a field-earned intuition into evidence.

Test co-presence against conversation. For the most socially avoidant young people — the ones a run club repels — do shoulder-to-shoulder formats (a shared activity, silent co-presence) reduce loneliness and bring people back more than talk-based support? A three-arm comparison (silent co-presence / structured talk / control), stratified by social anxiety, would tell us whether the door hypothesis holds.

11

Limitations and open questions

Intellectual honesty requires naming what I don't know and where this could fail.

• The evidence is largely correlational. Much of the companion-harm literature is cross-sectional; it establishes association, not always direction, and even careful researchers stop short of declaring companions simply harmful, framing the technology as highly context-dependent.[18] My thesis inherits that uncertainty.
• An app is not the answer to a public-investment problem. The carpool line, the empty rec center, the church that closed are outcomes of disinvestment, zoning, and how people's time is spent. Third places have thinned fastest in the census tracts that could least afford it.[21] No software re-staffs a rec center or reopens a congregation, and I am not claiming it does. Beaconeer addresses something narrower and more mundane: the invitation itself got expensive.
• It thickens the ties you already have. A beacon goes to a circle, and a circle has to exist first — it needs three people before it even wakes up. That makes Beaconeer a tool for bonding ties, not bridging ones; it presupposes the table rather than building it. The person with no circle to seed is exactly the person it does not yet reach.
• Reach and equity run opposite to the need. The exposure concentrates among lower-income young people; an invite-only iOS beta is close to the inverse of that distribution. I do not have a date for Android or for open signup, and I would rather say so than promise one.
• There is no 3am answer. The companion's structural advantage is that it is awake when nobody else is. A beacon needs people who can come. Beaconeer does not solve the hour that this paper spends four sections describing, and pretending otherwise would undo the argument.
• The hardest-to-reach may still not be reached. The most isolated young person may not take any step, however low the threshold. The design goes to the door; it cannot make anyone open it.
• Connecting vulnerable young people to each other is genuinely fraught. It demands serious safety architecture, and it is not something to ship casually. The parts of Beaconeer closest to that risk are the parts to build slowest and guard hardest.
• The smoke detector is a hard, unsolved problem. Reliable crisis detection is difficult, the cost of a false negative is severe, and it must never be autonomous. It needs clinical partnership and a human in the loop, always. None of this replaces professional crisis services.
• The deepest risk is building another trap. Any tool that touches lonely young people can become either another engagement machine or a thing that only reaches the already-fine. The graduation metric is my guard against the first; honest validation is my guard against the second. Neither is a guarantee.

12

Conclusion

The companion apps got one thing exactly right and one thing exactly wrong. Right: they went to the layer where people actually are — tier zero, 3am, no threshold — which is the layer everything else abandoned. Wrong: once there, they were built to keep people, and keeping a lonely person is not the same as helping them. The research is now clear that for the loneliest users, being kept makes it worse.[1][2][3][20]

The tea tent shows there is another way to hold that layer: human presence at the core, a shared ritual as the low-threshold reason to be there, open around the clock, escalation held in reserve — and success defined not by how long someone stays but by their leaving, grounded, back toward other people. That is a thing that can be built with software, if the software is willing to play host, trainer, and smoke detector, and refuses to play the friend.

That refusal is the whole point. The measure of this kind of technology is how gracefully and how quickly it makes itself unnecessary — how reliably it gives people back to each other. That is what I'm trying to build with Beaconeer, and this document is both the argument for it and the standard I want to be held to.