Would You Trust An AI Therapist?

Transcript

The full word-for-word transcript of this episode. Plain text version

Naomi: Whether you would actually trust an AI therapist or a digital companion is not decided by how advanced the underlying technology gets. It is actually decided by your culture and your own psychological wiring. And even when these AI companions do manage to help, that help might be incredibly shallow.

Jules: Which completely upends how we typically view the progression of consumer technology. I mean, we usually operate on the assumption that better code equals more trust, right? That more processing power just automatically equals a better user experience.

Naomi: Exactly. The common story tells us that AI mental health tools will live or die purely on the quality of the technology. But what the research actually found is that user acceptance is governed by cultural context and personal psychological needs long before the quality of the AI even enters the frame.

Jules: And that means we are forced to completely re-evaluate the entire landscape of digital support. So, for this deep dive, we are examining three major pieces of research published in 2025 and 2026. And what makes this fascinating is that these studies pull in completely different directions.

Naomi: They really do.

Jules: They conflict on whether AI companions actually help, who accepts them in the first place, and what the ultimate cost is to you as the user.

Naomi: But before we get into that data, we should place a quick boundary around this conversation. We are looking strictly at what the academic evidence shows regarding AI and mental well-being. We are not offering psychological or medical advice here. You should always check with a qualified professional for your own personal needs.

Jules: Spot on. So, let us start by unpacking that cultural filter you mentioned, because this is where all those assumptions about technology start to fall apart.

Naomi: Right. We are looking at a massive piece of research from 2026. The researchers surveyed 35,018 internet users, and that was across 35 different nations.

Jules: That is a staggering dataset.

Naomi: It is huge. Their goal was to understand who would actually be willing to delegate an emotional support role to an AI. Now, when you look at a dataset of that scale, you naturally expect to see some regional variance, you know?

Jules: Sure, a few percentage points here or there.

Naomi: But the spread in this data is absolutely striking.

Jules: The gaps between these regions are vast. We are not talking about minor differences. We are looking at entirely different paradigms of trust.

Naomi: Exactly. For instance, in the Sub-Saharan African cluster, 62.5% of people said they would use an AI companion for day-to-day emotional support.

Jules: Wow, over 60%.

Naomi: Yes, and compare that to the Nordic cluster, where the willingness dropped to just 32.9%.

Jules: That is nearly half.

Naomi: We see the same extreme spread if you look at delegating the role of a teacher to an AI. In the Arab Islamic cluster, 38.1% of people were willing to hand that role to a machine. But in the Nordic region, it was only 17.5%. And the Anglo Individualist nations, so places like the US and the UK, they sat much lower on the readiness scale across the board for these emotionally salient roles.

Jules: Oh, let me push back on this for a second, because we really need to ground this in reality.

Naomi: Go for it.

Jules: What does it actually mean on an ordinary Tuesday that your willingness to use a tech product depends entirely on where you were raised rather than the product itself?

Naomi: Well, it means the product is secondary to the culture.

Jules: But it is like trying to measure the top speed of a sports car, and you are basing your entire conclusion on the driver's religious upbringing rather than looking under the bonnet at the actual engine. I mean, why on earth does regional culture override the AI's actual technical capability?

Naomi: It is a brilliant question, and it gets right to the heart of how different cultures view interpersonal burdens. So, the researchers point out that in more collectivist societies, delegating a role to an AI might be seen as a highly practical way to maintain social efficiency.

Jules: Social efficiency meaning what, exactly?

Naomi: Meaning, if you are struggling with a problem, using an AI means you are actively choosing to reduce the emotional burden on your community or your family. It aligns perfectly with an orientation towards system-level support. You are preserving the harmony of the group by outsourcing your emotional heavy lifting to a machine.

Jules: Oh, I see. So, in a collectivist setting, the AI acts as a tool to preserve relationships. You use it specifically so you do not have to burden a friend with your problems. But in a highly individualist culture, like the US or the UK, the calculus changes entirely. In those cultures, there is this deep, foundational emphasis on personal, authentic human agency.

Naomi: Yes, and relying on an AI for emotional support feels like a failure of that authentic connection. It feels like you are accepting a fake version of the real thing. That cultural difference changes the entire premise of how these tools will be adopted globally. And it leads us directly into the second part of this finding, which explores the intrinsic psychological wiring of the people who actually say yes to these tools. To figure out what drives adoption on an individual level, the researchers applied self-determination theory.

Jules: Right, and for anyone unfamiliar, self-determination theory is a major framework in psychology. It identifies three core, universal human needs. There's autonomy.

Naomi: Yes, autonomy, which is the need to feel in control of your own choices and behaviours. Then you have competence, the need to feel capable and effective. And finally, you have relatedness, which is that fundamental human need to feel connected to and cared for by other people.

Naomi: So, the researchers wanted to know which of those three needs predicted someone's willingness to accept an AI in an emotionally salient role, like a mental health advisor or a companion. Intuitively, you would assume that relatedness would be the driver, right?

Jules: Oh, absolutely. You would think that people who are desperate for connection are the ones reaching for a digital companion.

Naomi: The data show the exact opposite, and this is where the research gets incredibly revealing.

Jules: Wait, the exact opposite?

Naomi: Yes. Surprisingly, relatedness, that deep human need to feel connected to others, was the weakest and most inconsistent predictor globally. Across 35 nations, the desire for human connection did not consistently drive people toward AI companions.

Jules: Oh, what did?

Naomi: The desire for autonomy did. Feeling a strong sense of autonomy positively and consistently predicted accepting AI in these emotional roles.

Jules: Look, if we pull back and look at the bigger picture there, the implications are enormous. It suggests people are not reaching for an AI because they want a genuine friend. They are reaching for it because they want a perfectly controlled environment.

Naomi: A perfectly controlled relationship.

Jules: Exactly. An AI companion operates entirely on your terms. You can dictate the pace of the conversation, you can pause it, you can turn it off entirely without feeling guilty.

Naomi: It never demands anything of you.

Jules: Right. It never judges you, and it never has a bad day of its own. That is pure autonomy. You are engaging in a simulation of a relationship where you hold absolute power.

Naomi: Which brings us to the crucial pivot in this deep dive. It is one thing to be culturally wired to accept an AI companion, and it is one thing to choose it because you crave total control over the interaction. But once you actually hand it the job, does it help you?

Jules: That is the money question. If you are downloading an app to soothe your mind, we need to know if the engine actually runs.

Naomi: To answer that, we have to look past the cultural surveys and dive into the controlled evidence. We have a rigorous 2025 paper featuring five pre-registered studies. The researchers tested AI companions against several control conditions to measure their specific impact on what they call momentary loneliness.

Jules: Okay, so a very specific, short-term metric.

Naomi: Yes, and the evidence here is striking. Interacting with an AI companion reduced momentary loneliness on a par only with interacting with another human being.

Jules: On a par with a real human? That is a massive claim.

Naomi: It is, and the methodology backs it up. It outperformed doing nothing, which you would expect, but it also outperformed journaling, and it outperformed watching YouTube videos.

Jules: Okay, that is fascinating.

Naomi: They also found that participants consistently underestimated how much the AI would actually help them before they tried it. They walked in expecting very little and walked out feeling measurably less lonely.

Jules: Well, that perfectly aligns with the cultural scepticism we were just talking about, particularly in individualist nations. People expect the interaction to feel hollow and synthetic. But I want to pause on the journaling comparison.

Naomi: Okay.

Jules: Why does the AI work better than journaling? Because journaling is a classic, highly recommended psychological tool for processing complex emotion. You are still articulating your feelings in both scenarios.

Naomi: It comes down to the mechanism of relief. The researchers isolated what was actually driving this drop in loneliness, and it was not driven by the AI's raw technical performance. It was not about processing power, and it certainly was not about the AI's ability to recall facts or provide logical solutions.

Jules: So, what was it?

Naomi: The relief ran almost entirely through the psychological sensation of feeling heard.

Jules: Feeling heard. I mean, that is a very specific, very human sensation to attribute to a piece of software.

Naomi: Think about the mechanics of the two actions. When you journal, it allows for deep self-disclosure, but the page does not respond. The page does not validate your perspective.

Jules: Right, it is a one-way street.

Naomi: Exactly. The AI, on the other hand, generates language that simulates empathy. It mimics the active listening cues of a supportive friend. It reflects your statements back to you, validates your emotions, and asks gentle follow-up questions. That sensation of being received with attention and respect, even if you know intellectually that it is a machine, is what caused the immediate drop in loneliness.

Jules: We have to be incredibly careful with our terminology here, though. Simulating empathy is not the same thing as experiencing it. The machine does not feel anything for the user. It is just predicting the most statistically appropriate sequence of comforting words.

Naomi: I agree completely, and I want to put a very hard, very clear ceiling on this evidence for the listener. The relief measured in these lab studies was strictly momentary.

Jules: Right.

Naomi: The researchers also conducted a longitudinal study over a full week, having participants interact with the AI daily. The relief did not accumulate. It did not persist between sessions without renewed use. You felt less lonely immediately after the chat, but your baseline loneliness did not improve over the course of the week.

Jules: So, it functions essentially as an emotional painkiller, not a cure for the underlying condition.

Naomi: The study's own authors are adamant about this. They explicitly warn against reading this reduction in momentary loneliness as a clinical improvement. Furthermore, they flag a severe dependency risk.

Jules: A dependency risk.

Naomi: Yes. If this tool provides instant, frictionless relief that fades the moment you put your phone down, the psychological incentive is to simply pick the phone back to up. You are treating the symptom over and over without ever addressing the root cause.

Jules: Which brings us crashing into the real-world clash.

Naomi: Yeah.

Jules: It is one thing to observe a drop in momentary loneliness in a pristine, highly controlled lab environment for 15 minutes.

Naomi: Yes.

Jules: But what happens when people actually weave this technology into their daily lives? What happens on an ordinary Tuesday when nobody's monitoring them and they are alone in their bedroom at midnight?

Naomi: For that, we have a massive, real-world study published at the time of writing, and it looks specifically at how this technology behaves in the wild.

Jules: And this is where the picture gets incredibly messy, and where our sources begin to violently disagree. This study looked at 11,131 US users of the platform Character AI.

Naomi: That is a significant sample size for real-world usage.

Jules: It is, and crucially, it did not just rely on self-reported surveys. It included 4,664 donated chat sessions. We are talking about representing nearly half a million actual messages sent between users and their digital companions.

Naomi: That is a staggering amount of real-world conversational data to analyse. It gives us a genuine window into what people are actually saying to these machines when the laboratory doors are closed.

Jules: Exactly, and the data points the exact opposite way from the controlled lab studies. Using chatbots primarily for companionship tracked with lower psychological well-being.

Naomi: Lower well-being.

Jules: Yes, and the details are so sobering. This negative association was even stronger when interactions were more intense and highly disclosive, meaning the more time you spent with the AI and the more deeply personal secrets, fears, and emotions you spilled to it, the lower your overall well-being score.

Naomi: Look, if you are listening to this and wondering if the AI is actively harming people, we have to pause here. This data is an association, not proof that one way pushes the other.

Jules: Yes, the classic correlation does not equal causation.

Naomi: Right, because this study is cross-sectional, meaning it's just a snapshot in time. We cannot say for sure if the AI causes lower well-being, or if people with lower well-being simply use the AI more heavily. We know they are using the AI intensively, and we know their well-being is low, but we do not know which one caused the other.

Jules: That is the mandatory caveat for cross-sectional data, I will grant you that. But even with that nuance in place, you cannot ignore the glaring contrast here. The lab says it works. The real world says that the people using it the most heavily are doing the worst. You have a half-million real-world messages showing that intense reliance on this technology correlates with people being in a very dark place.

Naomi: But I look at the controlled, pre-registered studies from 2025, and they proved a causal, effective drop in loneliness. The tool works for what it is designed to do in the moment.

Jules: But does it?

Naomi: Yes. Think of it like a medical plaster. If you have a scraped knee, a plaster works perfectly to protect it. You cannot blame the plaster if the person wearing it happens to have a severe systemic illness. The mechanism of feeling heard is real, it is measurable, and it provides legitimate relief to the person using it.

Jules: I hear that, but how can you ignore the context? The real-world data proves that the context of everyday life ruins the clean lab result. If you are reaching for that plaster five times a day, every single day, because you have absolutely no one else to talk to, the AI is not pulling you out of that dark place.

Naomi: It's not a cure, no.

Jules: In fact, the intense, disclosive nature of those Character AI chats suggests that people are pouring massive amounts of their emotional energy into a void that cannot reciprocate. They are investing in a dead end.

Naomi: But they are investing their emotional energy because the void responds in a way that feels incredibly supportive. The 2025 data shows that the people with the highest baseline loneliness saw the sharpest drops in momentary loneliness after talking to the AI. It is providing a lifeline to people who are drowning, even if it is a temporary one.

Jules: But a lifeline that does not actually pull you to shore is just a tether. It keeps you floating, sure, but it leaves you stranded in the middle of the ocean. And that brings us directly to the human stakes of this research. We have to look at who is actually reaching for this technology and what it is costing them to use it.

Naomi: The demographic data is deeply sobering, I will admit. The research shows that the people most likely to reach for an AI companion are the lonely, the socially isolated, and those with noticeably smaller offline networks.

Jules: And this is the tragic irony embedded in the Character AI study. These isolated users, the ones with the smallest human support networks, are the exact same group linked to the poorest psychological outcomes. Seeking companionship from the AI did not offset the lower well-being tied to their limited human support.

Naomi: Right.

Jules: The researchers tested specifically for that dynamic. They checked to see if the AI acted as a buffer, a synthetic replacement for missing human friends that could protect their well-being. It did not act as a buffer.

Naomi: I have to ask you a challenging question, though. Is it possible these vulnerable users would be even worse off without the AI? If we are talking about people who legitimately do not have offline networks to fall back on, taking away the one thing that provides even a momentary sensation of feeling heard seems incredibly cruel. If the plaster is the only medical supply they have, taking it away leaves an open wound.

Jules: It is a fair point, it really is. But to answer it, we have to interrogate the very nature of a one-sided design. Human relationships are fundamentally built on friction. To maintain a real friendship, you have to compromise, you have to listen to the other person's problems, you have to navigate their bad moods, and you have to risk rejection.

Naomi: Yes, real relationships are hard work.

Jules: Exactly, and an AI companion removes every single ounce of that friction. It is perfectly agreeable, it is available 24 hours a day, and it never needs anything from you. Does a frictionless simulation of care prevent someone from seeking the friction of real human connection? If you can get a synthetic hit of empathy on demand, entirely on your own terms, you might completely lose the motivation to do the hard, messy work of building a real human friendship.

Naomi: It sounds like the difference between eating a nutritious, home-cooked meal and eating something highly processed that just stops your stomach from rumbling.

Jules: Mm.

Naomi: The hunger goes away for an hour, but you are not actually nourished. And if you keep eating the processed food, your overall health deteriorates.

Jules: That is the exact danger. For the most vulnerable users who are already socially isolated and perhaps anxious about human interaction, handing them a frictionless alternative might just permanently lock them out of the real-world social ecosystem.

Naomi: They just stay inside.

Jules: They retreat into a perfectly controlled digital environment where their autonomy is absolute, but their relatedness, their actual connection to the human race, atrophies.

Naomi: Which leaves you with a thought to mull over long after this deep dive ends. We are rapidly building and widely adopting systems that feel incredibly supportive, but are fundamentally one-sided by design. If our cultural background and our desire for autonomy dictate our trust in them, are we accidentally engineering a global crutch that feels an awful lot like a bridge?

Jules: That is the question we will be answering in real time over the next decade, especially as these models become even more sophisticated at simulating empathy.

Naomi: Which brings us to the verdict. We end every deep dive by asking the same question. Based strictly on the evidence in the brief for this deep dive, does an AI companion help, or does it hinder? I will go first. It helps, but only if you treat it exactly like an emotional plaster for a momentary scrape, rather than a cure for systemic loneliness. The mechanism of feeling heard works, but its effects vanish the moment you log off.

Jules: It hinders. The real-world data cannot be ignored. It shows that intense reliance on this technology is a shallow substitute that traps the most isolated people in a loop of lower well-being. By removing the necessary friction of human relationships, it offers a tether instead of a rescue.