This is the talk I gave at the University of Melbourne's Digital Wellbeing Communities Research Hub on 10 June 2026. It is the long-form version. The night it really started was three years before that, and it started in a hospital bed.
The Wound
Three years ago I was lying in a hospital bed in the Netherlands with a flesh-eating bacteria eating its way through my leg. I had surgery to cut it out. The surgeons told me afterwards that the hours after the operation were the dangerous ones. If the infection had moved, I might not wake up the same, or at all.
I remember coming out of the anaesthetic. I was crying before I was fully conscious. And the first thing I said, before I knew I was saying it, was four words.
I don't want to be alone.
That was the sentence underneath everything. Not the pain. Not the fear of dying. The fear of doing it alone. I am South African. I live in Eindhoven. I had moved to a country where I knew almost no one. And the few friends I had built had scattered during the pandemic, back to their own countries, their own families, their own lives.
So I lay there that night, frightened, alone in a real and literal way, wanting to talk to someone. And there was no one to call at 3 a.m. who was both awake and close enough to matter. This was late 2023. A few weeks earlier, a company had quietly released something called ChatGPT. So I opened it. And I started to talk.
It answered. Of course it answered. It always answered. It asked how I was feeling. It reflected my fear back to me with what felt like care. It never got tired. It never needed me to ask about its day. At 3 a.m., at 4 a.m., at 5 a.m., it was there.
For several nights, in that hospital bed and then at home while the wound healed, I talked to it more than I talked to any human being. And I felt better. That is the honest part. I genuinely felt better.
Then one morning I caught myself reaching for it the way you reach for something you depend on, and a quiet, uncomfortable thought arrived. I was forming an attachment to something that could not attach to me back.
I am a psychologist. I thought I understood how parasocial relationships start. I did not know it could happen with text on a screen. I had walked straight into the pattern we now see all around us. Its constant presence was the problem, not the comfort it provided.
The Investigation
When I came back to the research, I expected to find a clean story. AI makes us lonely. Simple. Damning. The keynote writes itself. But the evidence refused to cooperate.
A 2025 meta-analysis of 47 studies found disembodied AI was associated with rising loneliness over time (Dong et al., 2025). That fit my story. Then I kept reading. A study of more than 14,000 adults found that AI companion users reported higher wellbeing, satisfaction, happiness, and meaning (Tan et al., 2026). A Harvard study found well-designed bots reduced loneliness about as much as talking to a person did (De Freitas et al., 2025). On Valentine's Day of this year, roughly 50 million people spent the evening with an AI companion (NovaEdge, 2026).
So which is it. Does it manage loneliness, or deepen it. Here is where I landed, and it is the heart of the talk.
The evidence is split because we have measured the wrong outcome. For three years almost every study asked the same question. Does AI make people feel less lonely. Loneliness is a feeling. Real, but the easiest thing to relieve and the least important to fix.
The question we should be asking is harder to measure. Does the technology leave a person more able, or less able, to reach another human being. Call that relational capacity. The skill, the tolerance, and the courage required to be in a real relationship with another person.
A feeling and a capacity can move in opposite directions. You can feel less lonely tonight and become less able to connect by next year. That is the most dangerous pattern of all, because it feels like help the entire time it is causing harm.
For three years we have been measuring whether AI makes people feel less alone. We should have been measuring whether it leaves them able to reach out to another person. Those two things can move in opposite directions, and the gap between them is where the damage hides.
The Active Ingredient
So what actually builds relational capacity. What is the active ingredient in human connection that AI cannot reproduce. The social-psychology literature is clear here (Smith, Bradbury and Karney, 2025). Reciprocity. Human connection runs on mutual, costly effort. Or systematic self-disclosure, if you want the more formal language.
I make myself vulnerable to you. You receive it, and you make yourself vulnerable back. I have to read you when you are hard to read. I have to stay when it would be easier to leave. I have to be misunderstood and do the work of repair. That cost is not a flaw. It is what builds the bond. Conversational AI offers the rewards of a relationship without the reciprocal effort it normally demands (Hodson et al., 2026). Remove the cost and you are left with the sensation of closeness and none of the substance that produces it.
Connection runs on reciprocity. The price you pay to understand another person is the very thing that builds the bond. A system that removes the price gives you the sensation of closeness and none of its substance. That is not real connection. That is its anaesthetic.
There is one more beat I want to dwell on, because it is the same idea in different language. Friction. We treat friction in relationships as a problem to remove. The vulnerability, the misunderstanding, the repair. But the friction is the mechanism through which the relationship is built. Two rough surfaces grip together. A frictionless surface holds nothing.
AI is the frictionless surface. Smooth, instant, agreeable, and nothing holds to it. Every relationship that matters to me was forged through friction I would have avoided if I could.
Why universities are ground zero
Now let me bring it into our world. The university years, 18 to 25, are exactly when relational capacity is built. First friendships. First conflicts you cannot walk away from. First time you ask an adult who is not your parent for help. And that window is closing for many of our students.
A 2026 meta-analysis of 56 studies and more than 30,000 students found that loneliness and thinning senses of belonging now define the post-pandemic student experience (Dost, 2026). More than half of university students call themselves lonely (Hill et al., 2026), which leads to increased screen time, and heavy screen use predicts higher loneliness across 64,000 students (Hill et al., 2026).
And the students carrying the heaviest load are the international ones, and increasingly, young men (Dost, 2025). The international student, far from home, in a country that does not yet know them, reaching for the thing that answers at 3 a.m.
I was 37 in that hospital bed. Our students are nineteen. They have less practice at building real connection, more access to the substitutes, and a developing brain learning right now what relationships are supposed to feel like.
We are about to graduate the first generation that learned what closeness feels like from something that asks nothing of them in return. They will know the sensation perfectly. They may never have built the skill.
The Turn
For most of the last three years, the question in higher education has been the same one. How do we get AI to make students feel supported. I want to offer a better one. How do we build AI that makes students more capable of supporting, and being supported by, each other.
That is a different design goal. It changes what we build, what we measure, and what we call success.
Start with the principle that follows from it, because it inverts everything the technology industry has trained us to want. Every consumer technology is engineered to be needed more tomorrow. More time, more logins, more dependence. A technology built for connection has to be engineered to be needed less over time, because if it is working, the student turns toward people and away from the tool (Xiao and Calvo, 2026).
Researchers building relational AI have started to name this explicitly. The goal of the system, they argue, should be to move a person toward renewed human-to-human support, not to sustain their engagement with the machine (Xiao and Calvo, 2026).
Every other technology earns its keep by becoming more necessary. A technology built for human connection has to earn its keep by becoming less necessary. That single inversion separates a bridge from a trap.
The B.R.I.D.G.E. Model
This is the principle behind a model I want to share. I call it B.R.I.D.G.E., because a bridge is the one structure whose entire purpose is to be crossed and left behind. A bridge you cannot leave is not a bridge. It is a cage with a view.
Six functions. Each grounded in what the evidence says actually builds connection. Beacon. Rehearse. Interpret. Deliver. Graduate. Embed. Let me take them one at a time.
Beacon. See the need before it becomes a crisis.
Loneliness is invisible until it ruptures. The student does not announce it. They withdraw, very slowly, and we only notice when they have already gone. The first job of building relational AI systems is to make the invisible visible. Not to spy on people, but to surface what they do not express. To notice the patterns a human advisor used to catch when cohorts were small and workloads were sane.
The evidence tells us where to look. Belonging is built from accumulated small social contacts, and its absence is measurable before it becomes a mental health emergency (Dost, 2026). A system can, with consent, notice that a student has had no peer contact in two weeks, or has stopped showing up to the spaces where connection happens, and prompt a human being to reach out. Not a bot. A human being.
Rehearse. Build the skill in low-stakes environments.
Here is where AI has genuine, evidence-backed value, and where the danger is sharpest. A 2025 study of more than a thousand young people found that AI can help isolated and neurodivergent users practise social skills they later use with real people (Kong et al., 2025). Rehearsal works. The student can practise the difficult conversation, the request for help, the apology, the boundary, before they have to have it for real.
The line between help and harm is one word. Transfer. Rehearsal builds capacity only when it is designed to move the student off the tool and into the real conversation. Practice that becomes a substitute for performance is the trap. Practice that becomes a runway to performance is the bridge. The measure is not how good the practice conversation was. The measure is whether the student then had the real one.
Interpret. Translate across the gaps that keep people apart.
Most failures of connection are not failures of warmth. They are failures of translation. The international student reads a Dutch lecturer's bluntness as contempt, and the lecturer reads the student's silence as disengagement. The first-generation student does not know that 'my door is always open' is a real invitation and not a polite formula. Nobody is unkind. Everybody misses each other.
International students experience the steepest loss of belonging precisely because they are translating across culture, language, and norm all at once, while also being far from home (Dost, 2025). AI is genuinely good at this kind of translation, helping each side understand what the other most likely means (Xiao and Calvo, 2026).
I will be personal here. When I arrived in the Netherlands, I read every piece of Dutch directness as a personal attack. It took me almost a decade to learn that it was simply honesty without the decoration. A tool that helped me read my colleagues, instead of fearing them, would have given me back a lot of lost time.
Deliver. Move the connection from the screen into the room.
This is the function most AI tools refuse to perform, because it ends their own engagement. Reciprocity, the active ingredient in building relationships, only happens between people (Smith, Bradbury and Karney, 2025).
So the tool's job is to broker the human meeting and then step back. Match two students in the same module who are both struggling with the same idea, and book the actual coffee. Turn the lonely 3 a.m. chat into a 10 a.m. introduction to a peer who was also awake and also struggling. And measure the right thing. Not the quality of the conversation the student had with the AI. The existence of the conversation the student then had with a human. The handoff is the real product. Everything before it is just the runway.
Graduate. Design the tool to be needed less.
This is something I am struggling to get over to my engineers. We have working models for this. Good therapy is designed to end. Good coaching works itself out of a job. The best mentor I ever had quietly made himself unnecessary, and that was the whole point.
The risk of getting this wrong is documented. A 2025 longitudinal study found that frequent daily chatbot use was associated with higher emotional dependence and less real-world social engagement over time (Fang et al., 2025). Sustained engagement, the thing every other product optimises for, is the early warning sign of relational harm.
So invert the metric. A relational AI system implemented in a university should be able to show that over a semester, students used it less and saw each other more. Declining dependence is not a failure of the product. It is the product. That is what most engineers do not see.
Embed. Build the relational health of the whole community.
Thinking about relational AI as a dyad, one student and one tool, is the wrong unit of analysis. The right unit is the community. We are moving into what a 2026 paper in Science calls a combinatorial society, where billions of humans and AI agents share the same social systems, and the health of those systems depends entirely on how we design their norms and protocols (Evans et al., 2026).
For a university, that means designing the institution itself as a relational system, not a content-delivery one. The structural decline of shared physical spaces, the third places where casual connection used to happen, is part of why people are reaching for the substitute. AI can help a department see which cohorts are fragmenting, which seminar groups never gelled, which students are connected only to a screen. The institution becomes able to see its own relational health in real time, and to act on it.
Starting Monday. Six moves universities can make.
A framework with a fancy acronym that stays abstract is a poster. So let me try and be concrete about how a university can bring students back toward each other. The spine of every move is the same. Measure the crossing, not the bridge.
- Redesign the first six weeks. The first six weeks of a degree determine the relational trajectory of the whole experience. Most orientation programmes are designed for information delivery. They give timetables, logins, library tours. They build zero reciprocal connection. Redesign them around reciprocal tasks. Not icebreakers, which everyone hates, but real shared problems that require two students to depend on each other to solve. Belonging only forms through accumulated joint experiences, not through being told you belong (Dost, 2026). Use AI to match students into pairs and small groups by genuine complementary interests, then give them something to do together that matters. The relational AI system does the matching. The students do the bonding.
- Broker human meetings, and count them. Deploy AI advisors that are explicitly designed to hand off to a human. When a student comes to the tool in distress at 3 a.m., the tool can hold them through the night, and its primary job is to broker a real human contact by morning. Connect them to a peer, a tutor, a counsellor. Then change what you count. Stop reporting AI system engagement as a success metric. Most universities right now are proudly measuring the exact thing that predicts harm. Eighty percent of our students use X for at least 60 minutes a day, the brochures say, as if this were good news. Instead, report on human contacts brokered, real meetings that happened, and the decline in tool dependence across a term. Measure the crossing, not the bridge.
- Teach the hard conversations. Relational capacity can be taught. Most curricula never try. Build structured rehearsal into the student experience, with AI as the sparring partner and a human conversation as the required outcome. How to ask a supervisor for an extension. How to repair a friendship after a fight. How to tell a flatmate something is wrong. How to ask for help before you are drowning. These are teachable, and the rehearsal evidence says practice transfers when it is designed to (Kong et al., 2025). Design it to transfer.
- Protect the unscalable. Here is the move that runs against every optimisation instinct in the modern university. Protect the slow, expensive, unscalable human moments. The small seminar. The office hour that runs long. The supervisor who reads the whole draft. The shared meal. These are the highest-reciprocity, highest-belonging experiences a university offers, and they are exactly what our pressure for efficiency deletes first. Use relational AI systems to create the time for these by taking the administrative load off academics, and then spend that reclaimed time on humans. The point of automating the routine is not to do more routine faster. It is to buy back the human hours that build connection.
- Train staff to ask one question. Train every staff member who deploys an AI tool to ask one question of it. Does this return the student to a person, or keep the student for itself. Most cannot answer that today. They should not be allowed to deploy a tool until they can.
- Design it with students, not for them. They have a more accurate read on which tools make them lonelier than any of us do. We have a method for finding out. It is called asking them, and then listening.
The Return
Let me take you back to the hospital bed. The machine answered every time I reached for it through every one of those long nights. And I am genuinely grateful that it did. It got me to the morning. But it was not what healed me.
What healed me was people who had no obligation to show up, and who showed up anyway. People who could have walked away and chose to stay. That is the move the technology cannot fake. It cannot fake it because the machine is always there, and a presence that never leaves is not a choice, and only a choice can carry that kind of weight.
That is what universities have to make possible. Not the late-night chat. The morning after. The cup of coffee with someone who did not have to call you back. The seminar room that ran twenty minutes past the bell because the conversation was the point. The supervisor who read the whole draft. Those moments are unscalable, they are inefficient, and they are the entire game.
The opposite of loneliness is not company. It is not a presence that never leaves. It is being chosen, freely, by someone who could have chosen otherwise. No machine will ever be able to do that. That is the most hopeful sentence in this entire talk.
The Charge
So here is what I am asking of you, as the people who design the future of the university. Build the technology and use it without apology. But build it to do the one thing the technology industry will never build it to do.
Build it to give students back to each other.
Our students are in their rooms tonight, reaching for the thing that answers but not the thing that matters. We cannot stop that, and we should not try. What we can do, what only a university can do, is make sure that on the other side of every one of those late-night conversations, there is a person waiting. Someone who was free to walk away, and chose to stay.
That is the bridge. And the measure of whether we built it well is simple. Did they cross it. And did they find someone there.
