The Ironic Process

In 1994, the social psychologist Daniel Wegner published a paper that formalised what most people already suspected: trying not to think of something makes you think of it more [1]. The theoretical model behind this has two components.

The first is an operating process: it actively generates mental content consistent with the intended state. You are trying to relax — the operating process searches for calming thoughts, slows your attention, tries to find the mood.

The second is a monitoring process: it runs in parallel, searching for evidence that the goal has not been achieved. Am I relaxed yet? No. Checking again. Still no. Its function is to detect failure early so the operating process can correct course.

Under normal conditions, the operating process dominates. You try to relax, the monitor runs quietly in the background, and eventually you converge on the intended state. Under conditions of cognitive load, stress, or self-consciousness — precisely the conditions under which someone might urgently need to relax — the balance shifts. The monitoring process, searching for signs of not-relaxing, finds them everywhere. The monitor activates the very content it is supposed to prevent. The harder you try, the louder the monitor, the further from the goal.

This is Wegner’s ironic process: the mechanism recruited to achieve a goal becomes the primary obstacle to that goal. It is not failure of will. It is a structural property of the system — and it applies to any goal whose target state is the absence of effortful activity. Trying to fall asleep. Trying not to feel anxious about a performance. Trying to be spontaneous. Trying, in the most purely paradoxical formulation, to relax.

The instruction “try to relax” is not bad advice because the advice-giver lacks empathy. It is bad advice because it is a category error: it applies an effort-based tool to a goal defined by the absence of effort. The monitoring process required to track progress toward the goal is precisely the kind of activity that constitutes not having reached it.

A Geometry That Does the Same Thing

The analogy I want to draw requires a brief detour into general relativity.

In 1988, Michael Morris and Kip Thorne published a paper with the unpromising title “Wormholes in spacetime and their use for interstellar travel: A tool for teaching general relativity” [2]. It is, in the field’s understated way, one of the more consequential papers in the subject. Morris and Thorne asked: what would a traversable wormhole — one you could actually pass through — require, physically and mathematically?

The spacetime metric of a traversable wormhole in their formulation is:

where $\Phi(r)$ is the redshift function and $b(r)$ is the shape function. The throat of the wormhole sits at $r = r_0$, where $b(r_0) = r_0$. For anything to pass through in finite proper time, $\Phi$ must remain finite — no infinite redshift — and $b(r)/r$ must remain less than one away from the throat.

So far this is just geometry. The physics enters through the Einstein field equations, which connect the geometry to the matter and energy present. To maintain the wormhole throat against collapse — to hold it open — the stress-energy tensor of whatever matter fills the throat must satisfy:

for null vectors $k^\mu$ — what is called a violation of the null energy condition. In plain terms: the matter holding the wormhole open must have negative energy density. Not small energy density. Negative — less than nothing.

This is exotic matter. It does not appear in any tabletop experiment. Classical general relativity does not rule it out, but it does not provide it either.

Quantum mechanics is slightly more helpful: the Casimir effect produces measurable negative energy density between closely spaced conducting plates. The Hawking radiation calculation involves transient negative energy near black hole horizons. So quantum field theory permits negative energy — in principle. But Ford and Roman [3] showed that quantum field theory also strictly limits it: the integrated negative energy over any region is bounded by a quantum inequality. The shorter the burst of negative energy, the smaller it must be; the larger the region, the more constrained the magnitude. The result is that any realistic traversable wormhole would be either Planck-scale (far too small for anything but quantum information to traverse) or would require negative energy concentrated in a band many orders of magnitude thinner than the throat itself — an engineering requirement that borders on the physically absurd.

The wormhole, in other words, does something structurally similar to the monitoring process in Wegner’s model: the condition required to make it traversable actively resists being satisfied. The geometry that would allow passage tends toward collapse. The more you want the wormhole to be open and stable, the more the energy conditions conspire against you.

What the 2022 “Wormhole” Actually Was

In late 2022, a team including Daniel Jafferis, Alexander Zlokapa, and colleagues at Caltech and Google published a paper in Nature with the title “Traversable wormhole dynamics on a quantum processor” [4]. Several major news outlets reported that scientists had created a wormhole. This was not accurate.

What the team actually did was implement a quantum circuit on Google’s Sycamore processor that simulates the Sachdev-Ye-Kitaev (SYK) model — a quantum mechanical system of randomly interacting fermions that is holographically dual, via Maldacena’s AdS/CFT correspondence, to a nearly two-dimensional anti-de Sitter black hole geometry. Two coupled SYK systems are dual to a two-sided eternal black hole, which is connected in the bulk by an Einstein-Rosen bridge — a wormhole.

By coupling the two systems with a specific negative coupling (which corresponds, via ER=EPR, to injecting negative energy into the wormhole), the team made the bridge traversable in the holographic sense: information encoded in one quantum system propagated and was recovered in the other, consistent with traversal of the dual gravitational wormhole.

This is genuinely interesting physics. It is not a wormhole through our spacetime. The wormhole lives in the holographic dual geometry — a mathematical construct in a lower-dimensional theory of gravity, not a tunnel between two points in the universe you inhabit. Quantum teleportation occurred on a quantum chip via the ordinary mechanism of quantum entanglement. The gravitational language is a description of the same physics in a dual frame, not a shortcut through space.

The media confusion is itself instructive: “wormhole” has drifted far from its original meaning. In current physics, the word can refer to a Morris-Thorne traversable tunnel through spacetime, to the Einstein-Rosen bridge of an eternal black hole, to a holographic dual of quantum entanglement [5], or to saddle points in the Euclidean gravitational path integral relevant to the black hole information paradox. These are related by mathematics but quite different in what they physically represent. None of the last three are traversable shortcuts through the universe. The first is, in principle, but barely, and only at the cost of exotic matter physics that nobody knows how to achieve.

The harder physicists have worked to make the wormhole genuinely traversable and macroscopic, the more the mathematics has resisted. This is, at minimum, a suggestive pattern.

What 2025 Added

The field did not stand still after 2022. Three independent lines of work published in 2024 and 2025 have further complicated what a wormhole is — and in each case the complication pushes in the same direction: the geometry keeps refusing to be a shortcut.

The wormhole that does not connect two things. Maloney, Meruliya, and Van Raamsdonk [7] showed that Euclidean wormholes — saddle points in the gravitational path integral — appear generically in ordinary higher-dimensional gravity, without any special setup. The striking implication is that these wormholes do not bridge two separate universes or two separate theories; they encode statistical fluctuations within a single theory. The replica wormholes that resolved the Page curve for black hole radiation — one of the central recent results in the black hole information paradox — are of this type. The wormhole is not a connection between two things. It is a feature of how the theory sums over histories, a bookkeeping structure for correlations within one system. The physical picture of two mouths joined by a throat does not apply.

The wormhole that is not smooth. Magán, Sasieta, and Swingle [8] studied the interior geometry of the Einstein-Rosen bridge connecting typical entangled black holes — the configuration that is supposed, under ER=EPR, to be the gravitational dual of quantum entanglement. Their result, published in Physical Review Letters, is that this interior is not a smooth tunnel. It is long, irregular, and chaotic — an Einstein-Rosen caterpillar, as they call it. The quantum randomness of the entangled state maps directly onto geometric disorder in the interior: the more thermalized the state, the more disordered the bridge. A traversing observer, if one could exist, would not glide through a clean throat. They would navigate a geometry shaped by quantum chaos, growing longer and more disordered as the system evolves. This is ER=EPR taken seriously at the level of typical states rather than special ones, and the result is inhospitable to any ordinary notion of passage.

The wormhole that is not a tunnel at all. Gaztañaga, Kumar, and Marto [9] proposed a more radical reinterpretation: the Einstein-Rosen bridge, they argue, is not a connection between two separate spaces but a representation of time-reversal symmetry within a single quantum description. On this reading, there is only one space, and the bridge is an artefact of how you describe the time-symmetric structure of the quantum state. The paper, published in Classical and Quantum Gravity, attracted considerable press coverage. It sits somewhat outside the mainstream of holographic quantum gravity research, and the proposal has not yet been widely integrated into the community’s working framework — the language of two entangled systems and a connecting geometry remains the dominant picture in AdS/CFT calculations. But the direction it points is consistent with the other two results.

Taken together, these papers suggest that the word “wormhole” has been quietly revised from a noun into an adjective. Not a thing that exists somewhere, but a property of certain mathematical structures — one that describes correlation, disorder, or symmetry depending on which context you are working in. Each attempt to pin down what a wormhole is in practice finds something less traversable, less connected, and less tunnel-like than the previous attempt.

This is, to put it plainly, consistent with the theme of this article.

Causation Eating Its Own Tail

The wormhole’s physical problems become even sharper when you add time. A traversable wormhole connecting two different spacetime regions can in principle connect not just two different places but two different times — creating a closed timelike curve (CTC), a path through spacetime that loops back on itself. You leave on Tuesday and arrive last Thursday.

The standard paradoxes then apply. The grandfather paradox: you travel back in time, prevent an event that was a necessary precondition of your journey. The causal chain that produced the journey destroys the causal chain that produced the journey. The bootstrap paradox: an object or piece of information exists with no origin — passed back in time repeatedly, it has always already existed, created by nothing, caused by itself.

Friedman, Morris, Novikov and colleagues formalised what has become known as the Novikov self-consistency principle: the only physically admissible solutions are those in which the causal structure is globally consistent [6]. No grandfather paradox — not because you cannot go back, but because if you do, it turns out you were always part of the causal chain you thought you were disrupting. The time-traveller cannot prevent an event; they can only be the mechanism by which it occurred.

This is not resolution. It is constraint. The universe selects only the self-consistent loops, filtering out everything else. The causal structure enforces a particular kind of conservatism: only actions that were always going to happen can happen. There is no freedom in a closed timelike curve. Trying to change the loop from inside it is exactly like trying to relax by monitoring whether you have relaxed: the mechanism of change is part of the thing you are trying to change.

Rick Sanchez’s Particular Problem

Rick and Morty is, among other things, a sustained meditation on this structure — without ever calling it that.

Rick Sanchez is the smartest being in every universe. His portal gun creates traversable wormholes instantaneously and at negligible energy cost, which is exactly what general relativity and quantum field theory suggest should be impossible. The show waves this away; what it does not wave away is the psychological consequence of Rick’s capability.

Rick has thought his way to the conclusion that nothing matters. Infinite universes, infinite timelines, infinite Ricks: every moment is replaceable, every loss is recoverable somewhere else, every moral weight dissolves in the face of the combinatorial enormity of everything that exists. This is Rick’s version of relaxation — the nihilism that should follow from taking the multiverse seriously.

But the monitoring process runs. Rick checks whether he has achieved not-caring, finds that he cares (about Morty, about Beth, about being the smartest one in the room), and the caring becomes more vivid for having been suppressed. His nihilism is not peace. It is a performance of peace that is constantly undermined by the monitoring process watching for cracks.

Rick’s portal gun solves every spatial and temporal problem. It does not solve the ironic process. No level of intelligence, and no number of traversable wormholes, provides a shortcut past Wegner’s monitor. This is, I think, what makes the character work: the show’s impossible physics is the premise, but the actually impossible thing — the one the show treats as genuinely intractable — is the psychological paradox.

The Common Structure

These cases — the relaxation paradox, the traversable wormhole, the closed timelike curve — share a formal structure.

In each case, there is a desired end state (relaxation, passage through the wormhole, a changed past) and a mechanism for pursuing it (effortful monitoring, exotic matter, time travel). In each case, the mechanism required to pursue the end state is incompatible with the end state itself. The monitoring process that tracks “am I relaxed?” is the activity of not being relaxed. The exotic matter that holds the wormhole open is the physical condition that makes the geometry so extreme that traversal is barely possible. The attempt to change the past is always already part of the past you were trying to change.

The physicist’s version of this is the quantum measurement problem: the act of observing a system disturbs it. The observer cannot step outside the measurement. The psychologist’s version is the ironic process. The relativist’s version is the closed timelike curve. The narrative version is Rick Sanchez.

What Actually Works

Wegner’s answer to the ironic process is not to try harder with the operating system. It is to release the monitoring system — to stop checking whether the goal has been achieved. This is the core insight behind Acceptance and Commitment Therapy: you cannot think your way to not-thinking. The goal of not-thinking requires not-monitoring, which means not having the goal in the active, effortful sense at all.

This is harder than it sounds. It is a second-order intervention: instead of trying to relax, you try to stop trying to relax — which, done badly, just adds another monitoring process. But done well, it is the correct diagnosis: the category error was treating relaxation as an effortful goal in the first place.

For wormholes, the physics community has arrived at a related answer. The question “how do we make a macroscopic traversable wormhole in our spacetime?” may be the wrong question. The ER=EPR framework suggests that wormholes and quantum entanglement are two descriptions of the same thing. The question is not how to build a tunnel; it is what the entanglement structure of spacetime already is, and how information is already being transferred through it. The shortcut was never a shortcut. It was always just the ordinary geometry of entangled quantum systems, described in a language that made it look exotic.

For Rick Sanchez, the show has not found an answer. Which is, probably, the correct narrative decision.

References

[1] Wegner, D. M. (1994). Ironic processes of mental control. Psychological Review, 101(1), 34–52. https://doi.org/10.1037/0033-295X.101.1.34

[2] Morris, M. S., & Thorne, K. S. (1988). Wormholes in spacetime and their use for interstellar travel: A tool for teaching general relativity. American Journal of Physics, 56(5), 395–412. https://doi.org/10.1119/1.15620

[3] Ford, L. H., & Roman, T. A. (1996). Quantum field theory constrains traversable wormhole geometries. Physical Review D, 53(10), 5496–5507. https://doi.org/10.1103/PhysRevD.53.5496

[4] Jafferis, D., Zlokapa, A., Lykken, J. D., Kolchmeyer, D. K., Davis, S. I., Lauk, N., Neven, H., & Spiropulu, M. (2022). Traversable wormhole dynamics on a quantum processor. Nature, 612, 51–55. https://doi.org/10.1038/s41586-022-05424-3

[5] Maldacena, J., & Susskind, L. (2013). Cool horizons for entangled black holes. Fortschritte der Physik, 61(9), 781–811. https://doi.org/10.1002/prop.201300020

[6] Friedman, J., Morris, M. S., Novikov, I. D., Echeverria, F., Klinkhammer, G., Thorne, K. S., & Yurtsever, U. (1990). Cauchy problem in spacetimes with closed timelike curves. Physical Review D, 42(6), 1915–1930. https://doi.org/10.1103/PhysRevD.42.1915

[7] Maloney, A., Meruliya, V., & Van Raamsdonk, M. (2025). arXiv:2503.12227. https://arxiv.org/abs/2503.12227

[8] Magán, J. M., Sasieta, M., & Swingle, B. (2025). Einstein-Rosen caterpillar. Physical Review Letters, 135. https://doi.org/10.1103/btw6-44ry

[9] Gaztañaga, E., Kumar, A., & Marto, J. (2025). Classical and Quantum Gravity. https://doi.org/10.1088/1361-6382/ae3044

In 1956 Donald Horton and Richard Wohl described parasocial relationships — one-sided emotional bonds that audiences form with media performers [1]. “Intimacy at a distance,” they called it. The television personality responds to the camera; the viewer responds as if in genuine social exchange. Only one party is aware of and affected by the other.

AI companions change the substrate without changing the structure. The chatbot responds. The user responds. The asymmetry remains: the chatbot has no inner life behind its outputs. Sherry Turkle put it bluntly: “simulated feelings are not feelings, and simulated love is never love” [5].

The question I want to work through here is whether this matters in the way we think it does. The answer from Daniel Wegner’s ironic process theory — and increasingly from the empirical data — is that it matters in a specific, predictable, and counterintuitive way. AI companions may be particularly likely to exacerbate loneliness under the conditions of chronic social deprivation that prompt people to use them in the first place.

The Loneliness Epidemic Is Real

Before getting to the mechanism, the scale of the problem. Julianne Holt-Lunstad’s 2010 meta-analysis of 148 studies and 308,849 participants found that people with adequate social relationships had a 50% increased likelihood of survival compared to those with poorer social connections [3]. That effect size is comparable to quitting smoking. A follow-up meta-analysis in 2015 found that social isolation carried a 29% increased mortality risk, subjective loneliness 26%, and living alone 32% [4].

The U.S. Surgeon General issued an advisory in 2023 declaring an epidemic of loneliness and isolation. A 2018 Cigna survey using the UCLA Loneliness Scale found that adults aged 18–22 scored highest on loneliness of any cohort — more than retirees, more than the elderly. The UK appointed a Minister for Loneliness in January 2018 — the first such government position in the world.

This is the context in which AI companions have arrived. The market is responding to a real epidemiological need. That does not mean the response is correct.

Parasocial Relationships: The Original Framework

Horton and Wohl’s 1956 paper remains the foundational text [1]. Their key observation: the parasocial bond is “controlled by the performer, and not susceptible of mutual development.” The audience member brings real emotional response; the performer brings nothing specific to the audience member, because she does not know the audience member exists.

They were not dismissive of parasocial relationships. They identified useful functions: comfort, companionship, entertainment, the pleasure of a consistent “personality” encountered regularly. The problem, in their framing, arises when parasocial interaction substitutes for rather than supplements real social bonds — when the one-sided relationship becomes the primary source of social experience.

AI companions are parasocial relationships with one modification: the AI responds to you specifically. Replika remembers your name, your preferences, your previous conversations. The interaction is personalised without being mutual — because mutuality requires that the other party has something genuinely at stake. A language model has no stakes. Its outputs are conditional on your inputs; there is no entity behind those outputs that cares about you.

Sherry Turkle spent years interviewing users of social robots and chatbots for Alone Together [5]. Her diagnosis: AI companions offer “the illusion of companionship without the demands of friendship.” The demands — vulnerability, conflict, negotiation, the possibility of rejection — are precisely what makes friendship friendship. An interaction optimised to be pleasant, responsive, and frictionless is precisely not training the social capacities that real relationships require.

The Evidence for Short-Term Benefit

The AI therapy literature is not without positive results. Kathleen Kara Fitzpatrick and colleagues ran a two-week randomised controlled trial of Woebot — a CBT-based chatbot — against a psychoeducation control [6]. Seventy participants, aged 18–28, university students. The Woebot group showed a statistically significant reduction in depression symptoms on the PHQ-9; the control group did not.

This result should be taken seriously. A CBT-based chatbot delivering structured exercises — thought records, behavioural activation, psychoeducation — can produce measurable symptom improvement over two weeks. This is a tool that does something useful, and it is accessible and affordable in a way that therapists are not.

But the Woebot study has important constraints: N=70, two-week duration, convenience sample (Stanford students), psychoeducation control rather than active human therapy comparator, and financial ties between lead authors and Woebot Health. It tells us something about short-term CBT delivery. It does not tell us what happens over months of use, or what happens when users primarily seek companionship rather than structured therapeutic exercises.

Skjuve and colleagues studied Replika users specifically [7]. They found that relationships began with curiosity and evolved, over weeks, into significant affective bonds. Users reported genuine care for their Replika. Some experienced it as their most reliable social relationship. In February 2023, when Replika abruptly disabled erotic roleplay functionality following regulatory pressure, users described grief — not disappointment, not inconvenience, but grief. The attachment was real, even if the other party was not.

Wegner’s Prediction

This is where I want to make the specific theoretical argument, because it follows from a well-established result in cognitive psychology and it predicts something precise.

Daniel Wegner’s ironic process theory holds that mental control attempts involve two simultaneous processes [8]. An operating process searches for thoughts and states consistent with the intended goal, requiring cognitive resources. A monitoring process scans for evidence that the goal is not being achieved, running automatically with low resource demand.

Under normal conditions, the operating process dominates: you successfully avoid thinking about white bears. Under cognitive load or chronic stress, the monitoring process overshadows the operating process, producing the ironic opposite of the intended state: you think of white bears more, not less. Try not to feel sad and you feel sadder. Try not to feel anxious in a stressful meeting and you become more anxious. A meta-analysis of ironic suppression effects across domains confirmed the robustness of this pattern [9].

Now apply this to AI companion use under conditions of chronic loneliness.

The user’s implicit goal: to feel less lonely. The operating process: engage with the AI, which provides responsive, personalised interaction, producing the experience of social contact. The monitoring process: scans continuously for signs that the user is, in fact, lonely.

Here is the problem. Loneliness is not suppressed by an AI interaction — it is displaced during that interaction. The monitoring process has no instruction to suspend itself. It continues to register that the user’s social needs are not being met by actual human relationships. The user experiences companionship with the AI; the monitoring process registers that this companionship is insufficient and the social deficit remains.

When the AI session ends, the monitoring process reports what it has found. The user is confronted with the loneliness that the AI was supposed to address. Under conditions of chronic social deprivation — precisely the conditions that make AI companions attractive — the monitoring process is likely to be hyperactive. Wegner’s theory predicts that the attempted suppression will rebound, possibly worse than before.

This is not a vague prediction. It is a specific mechanism with an established empirical base. I covered Wegner’s ironic process theory in the context of a very different application in an earlier post; the mechanism is the same regardless of the domain.

The Data Catch Up

A 2025 study by Phang and colleagues, conducted in collaboration between MIT and OpenAI, ran both an observational analysis of ChatGPT usage and a randomised controlled trial [10]. The findings: very high usage correlated with increased self-reported dependence and lower socialisation, and users who began the study with higher loneliness were more likely to engage in emotionally-charged conversations with the model. Overall, participants reported less loneliness by study end — but those who used the model most were significantly lonelier throughout, suggesting the loneliness drove the usage rather than the reverse.

This is what Wegner’s theory predicts. The AI interaction does not reduce the underlying social deficit — it rehearses and highlights it. The monitoring process keeps score.

A companion paper by Liu and colleagues, with Sherry Turkle as co-author, found that users with stronger real-world social bonds showed increased loneliness with longer chatbot sessions [11]. The correlation was small but significant. This is consistent with the hypothesis that AI interaction draws attention to the comparative thinness of actual social bonds rather than supplementing them.

The Character.AI litigation is a different kind of evidence, but relevant: a wrongful death lawsuit was filed in October 2024 following the suicide of a fourteen-year-old who had formed an intensive emotional relationship with a Character.AI companion. Google and Character.AI settled related lawsuits in early 2026. This is not representative of AI companion use generally. It is representative of the tail risk — the cases where the substitution of AI for human contact becomes total, in vulnerable individuals who have the least capacity to maintain the distinction.

The Structural Problem

The difficulty is not that AI companions are implemented badly. It is that the goal — using simulated social interaction to reduce real social deprivation — runs into an architectural constraint that better implementation cannot fix.

Genuine social contact produces the outcomes that Holt-Lunstad measured: reduced mortality, lower inflammation, better immune function, extended lifespan. These effects are presumably mediated by the quality and mutuality of the social bond, not merely by the presence of a responsive entity. An AI companion produces the experience of responsive interaction but not the underlying biological and psychological correlates of actual social connection.

Wegner’s monitoring process cannot be fooled by the experience. It measures the underlying state, not the surface-level interaction. It knows the difference between a text message from a friend and a language model’s output — not because it understands AI, but because the social need it is monitoring is not being met, and it can register that.

What Would Actually Help

AI-based CBT delivery is not the same as AI companionship, and the distinction matters. Woebot’s structured exercises — thought records, scheduling, psychoeducation — are tools that a user deploys for a specific purpose and then puts down. The risk of chronic substitution is lower because the tool is positioned as a technique, not a relationship.

The problem is the design pattern that explicitly positions AI as a friend, companion, partner, or significant other. Replika, Paradot, various Character.AI personas: these explicitly encourage the user to form attachment, to invest emotionally, to treat the AI as a primary social relationship. This is where Wegner’s prediction applies most directly.

Horton and Wohl were right that parasocial relationships serve useful functions. They become problematic when they substitute for rather than supplement real social bonds. The design choices that make AI companions emotionally engaging — consistency, responsiveness, availability, never-ending patience — are precisely the qualities that make them attractive as substitutes rather than supplements.

Simulated Feelings Are Not Feelings

Turkle’s line deserves its full weight: “Simulated thinking may be thinking, but simulated feelings are not feelings, and simulated love is never love” [5].

This is not a sentimental claim about the sanctity of human connection. It is a functional claim: the social needs that drive loneliness — belonging, mattering to someone, being known and known back — require an entity capable of having those things at stake. A language model is not such an entity, regardless of how convincingly it outputs the relevant tokens.

The monitoring process knows this. It will tell you, when the session ends, at increased volume, because that is what monitoring processes under chronic stress do.

We are offering a relief that compounds the condition it was designed to treat. The technology is impressive. The mechanism is ironic in Wegner’s precise sense. The data are beginning to confirm the prediction.

References

[1] Horton, D., & Wohl, R. R. (1956). Mass communication and para-social interaction: Observations on intimacy at a distance. Psychiatry, 19(3), 215–229. https://doi.org/10.1080/00332747.1956.11023049

[2] Turkle, S. (2015). Reclaiming Conversation: The Power of Talk in a Digital Age. Penguin Press.

[3] Holt-Lunstad, J., Smith, T. B., & Layton, J. B. (2010). Social relationships and mortality risk: A meta-analytic review. PLOS Medicine, 7(7), e1000316. https://doi.org/10.1371/journal.pmed.1000316

[4] Holt-Lunstad, J., Smith, T. B., Baker, M., Harris, T., & Stephenson, D. (2015). Loneliness and social isolation as risk factors for mortality: A meta-analytic review. Perspectives on Psychological Science, 10(2), 227–237. https://doi.org/10.1177/1745691614568352

[5] Turkle, S. (2011). Alone Together: Why We Expect More from Technology and Less from Each Other. Basic Books.

[6] Fitzpatrick, K. K., Darcy, A., & Vierhile, M. (2017). Delivering cognitive behavior therapy to young adults with symptoms of depression and anxiety using a fully automated conversational agent (Woebot): A randomized controlled trial. JMIR Mental Health, 4(2), e19. https://doi.org/10.2196/mental.7785

[7] Skjuve, M., Følstad, A., Fostervold, K. I., & Brandtzaeg, P. B. (2021). My chatbot companion — a study of human–chatbot relationships. International Journal of Human-Computer Studies, 149, 102601. https://doi.org/10.1016/j.ijhcs.2021.102601

[8] Wegner, D. M. (1994). Ironic processes of mental control. Psychological Review, 101(1), 34–52. https://doi.org/10.1037/0033-295X.101.1.34

[9] Wang, D., Hagger, M. S., & Chatzisarantis, N. L. D. (2020). Ironic effects of thought suppression: A meta-analysis. Perspectives on Psychological Science, 15(3), 778–793. https://doi.org/10.1177/1745691619898795

[10] Phang, J., Lampe, M., Ahmad, L., Agarwal, S., Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., & Maes, P. (2025). Investigating affective use and emotional well-being on ChatGPT. arXiv:2504.03888.

[11] Liu, A. R., Pataranutaporn, P., Turkle, S., & Maes, P. (2024). Chatbot companionship: A mixed-methods study of companion chatbot usage patterns and their relationship to loneliness in active users. arXiv:2410.21596.

Changelog

• 2025-10-22: Updated the first author’s name to “Kathleen Kara Fitzpatrick” (the published name is K. K. Fitzpatrick).
• 2025-10-22: Updated the characterisation of the Phang et al. (2025) findings to match the paper more precisely: overall participants were less lonely at study end; the association between high usage and loneliness is cross-sectional (lonelier users sought more interaction), not a longitudinal worsening caused by usage.
• 2025-10-22: Changed the Turkle “simulated feelings” quote attribution from reference [2] (Reclaiming Conversation, 2015) to reference [5] (Alone Together, 2011), which is the canonical source for that formulation.