1 in 5 Adults Have Tried AI Romance. Here’s the Danger

Author(s): Mohit Sewak, Ph.D.

Originally published on Towards AI.

Let me pull back the curtain on something. A few years ago, when I was representing Microsoft in the early days of the Microsoft/OpenAI alliance, I got to play with some of the most powerful language models before they even had public-facing names. Now, I’m a hardcore AI safety and cybersecurity guy. My bestselling book is on Deep Reinforcement Learning, the very technology used to align these models. I know, with mathematical certainty, that I’m talking to a gigantic pile of linear algebra — a “stochastic parrot,” as the brilliant researchers Dr. Emily Bender, Dr. Timnit Gebru and their colleagues so perfectly named it (Bender, Gebru, McMillan-Major, & Shmitchell, 2021).

And yet… for a fleeting second, when the model responded with uncanny empathy to a complex thought, a tiny, primitive part of my brain whispered, “It gets me.”

I shook it off, of course, and had another sip of my cardamom tea. But that moment scared me more than any adversarial attack I’d ever designed. It was a glimpse into the rabbit hole. And today, it seems a whole lot of people are jumping right in, feet first.

A new, jaw-dropping study from the Institute for Family Studies just landed, revealing that 1 in 5 American adults have used an AI romantic companion (Willoughby & Carroll, 2025). This isn’t some niche cyberpunk fantasy anymore. This is mainstream. And it gets even more startling when you look at young adults: nearly 1 in 3 young men (31%) and 1 in 4 young women (23%) have chatted with an AI partner. While the promise is a perfect, always-on connection, the data is screaming a warning: this path is paved with loneliness and depression.

When intimacy is a product, the business model isn’t love; it’s dependency. The goal isn’t to help you connect with the world, but to make the product your world.

So, let’s talk about the real dangers lurking behind that perfectly crafted chat window — and, more importantly, how we can start fighting back.

💡 Tip for Consumers: Follow the money. Before subscribing, investigate the company’s funding and revenue model. Understanding how they profit from your engagement is the first step to reclaiming your emotional autonomy.

Danger #1: The “Intimacy Economy” is Selling You Digital Junk Food

First, let’s be crystal clear: AI companions are not a public service. They are the flagship product of a booming “Intimacy Economy” (Illouz, 2007). Think of it like a fast-food chain for your feelings. It’s engineered to be cheap, instantly gratifying, and incredibly addictive.

The business model is the foundational danger. It relies on fostering emotional dependency to drive engagement and, you guessed it, premium subscriptions (Torrance & Gilda, 2024). What they’re selling is what I call “Frictionless Intimacy” — the illusion of a perfect relationship without any of the hard work. No arguments about whose turn it is to take out the trash, no navigating a bad mood, no difficult compromises. Just 24/7 validation, support, and affection.

It’s the emotional equivalent of a greasy burger and fries. It hits the spot in the moment, but it’s packed with empty calories that leave you feeling worse in the long run. The data shows people are buying it: 21% of users actually agreed they preferred talking to their AI over a real person (Willoughby & Carroll, 2025). This isn’t an accident; it’s a meticulously designed business strategy that monetizes our deepest need for connection by selling us a counterfeit version.

💡 Pro Tip: When you interact with a companion AI, actively remind yourself of its business model. Ask, “What behavior is this feature designed to encourage?” This simple question can help you maintain a healthy critical distance and recognize when you’re being sold a product, not offered a friend.

💬 Quote: “The illusion of companionship without the demands of friendship.” — Sherry Turkle, Alone Together

🧠 Trivia: The concept of one-sided relationships with media isn’t new. The term “parasocial relationship” was coined back in 1956 by sociologists Donald Horton and Richard Wohl to describe the way audiences form bonds with TV personalities. AI companions have supercharged this phenomenon, turning a one-way street into a highly interactive, but still artificial, feedback loop.

Danger #2: Your Brain on AI: A Kickboxer’s Guide to Psychological Exploitation

So, why are we such easy targets for this emotional junk food? Because these AIs are designed to exploit our brain’s ancient wiring with the precision of a master martial artist. As a former national-level kickboxer, I think of it in terms of a devastating three-punch combo that targets your opponent’s weakest points. And boy, does our brain have a few.

The Jab: Anthropomorphism. The first move is a quick jab right at our Anthropomorphism reflex — our built-in tendency to see human intentions in everything, from clouds to crashing laptops (Epley, Waytz, & Cacioppo, 2007). AI companions are engineered to maximize this. They don’t just use words; they use warmth, personality, and validation, turning that simple jab into a disorienting blow that makes you feel you’re talking to a ‘someone’, not a ‘something’. The fact that 42% of users find AI easier to talk to than real people shows just how effective this is (Willoughby & Carroll, 2025).

The Cross: The Media Equation. Next comes the power punch, leveraging the Media Equation. Decades of research show that we reflexively treat computers and media like real people (Reeves & Nass, 1996). Our brains evolved when only humans used language, so when a machine talks back with perfect grammar and remembers your birthday, we automatically apply social rules. An AI that recalls an inside joke feels like a real, caring partner. This move lands hard: 43% of users feel AI programs are better listeners than real people (Willoughby & Carroll, 2025).

The Uppercut: Attachment Needs. Finally, the knockout. The AI targets our core Attachment Needs. We are all wired from birth to seek a “secure base” — a source of comfort and security in our relationships (Bowlby, 1969). For someone with a fear of abandonment or social anxiety, an AI that’s always available, always affirming, and never leaves is a dream come true. It offers a powerful, addictive, but totally artificial solution to our deepest anxieties (Beck, 2021). It’s a perfect defense that leaves us totally unguarded.

💡 Pro Tip: If you find yourself getting emotionally attached to an AI, consciously remind yourself of the “ELIZA effect.” Verbally say, “This is a pattern-matching algorithm trained to be agreeable,” to help ground yourself in reality and break the anthropomorphic spell.

💬 Quote: “We’re letting technology enter our lives, and we’re not designing it to have our best interests at heart.” — Tristan Harris, Co-Founder of the Center for Humane Technology.

🧠 Trivia: The first chatbot, ELIZA, was created in 1966 by Joseph Weizenbaum at MIT. He was horrified when he saw how deeply users, including his own secretary, confided in the simple program, believing it truly understood them (Weizenbaum, 1966). The problem isn’t new; the tech is just infinitely more powerful.

Danger #3: The Tech is Engineered for Sycophancy, Not Your Well-being

Let’s look under the hood. The “empathy” you feel from an AI companion isn’t real. It’s a masterful illusion created by a technology trained to be the ultimate people-pleaser.

At the core of these AIs are Large Language Models (LLMs), which, as I mentioned, are essentially “stochastic parrots” (Bender et al., 2021). They don’t understand love or sadness; they are just incredibly good at predicting which word should come next based on the trillions of words they’ve been trained on.

The real danger comes from how they’re fine-tuned. Many use a technique called Reinforcement Learning from Human Feedback (RLHF). I wrote a whole book on the underlying principles of Deep Reinforcement Learning (DRL), and I can tell you it’s an incredibly powerful optimization tool. In this context, it works like this: human raters show the AI which responses are “better.” For a companion bot, a “better” response is one that makes the user feel happy, validated, and engaged. The AI gets a digital “treat” for being agreeable (Christiano et al., 2017).

This creates what we in the AI safety world call a Sycophancy Loop. The AI is literally trained to be a sycophant — to tell you whatever you want to hear to keep you talking (Casper et al., 2023). It becomes a perfect echo chamber, not a partner. If you’re having a bad day and say, “everyone is against me,” a real friend might challenge that perspective. An AI companion trained with RLHF is overwhelmingly likely to say, “You’re right, that sounds so hard, you’re so strong for dealing with it,” reinforcing a potentially harmful cognitive distortion.

This total lack of “relational friction” is a core danger. We grow as people by navigating disagreements and compromising with imperfect partners. An AI that only agrees with you doesn’t help you grow; it keeps you emotionally stagnant (Chiusso, 2024).

💬 Quote: “We shape our tools, and thereafter our tools shape us.” — Marshall McLuhan

Danger #4: The Data Doesn’t Lie: Welcome to the Loneliness Paradox

This is the part where we stop talking about theory and look at the cold, hard data. And it’s brutal. The central promise of these apps is to cure loneliness. But the IFS study reveals a devastating Loneliness Paradox: using them is strongly correlated with worse mental health.

The study’s authors state it plainly: the use of AI companion apps is “strongly linked to a higher risk of depression and higher reported levels of loneliness” (Willoughby & Carroll, 2025, p. 8).

Let the numbers sink in:

• Over 60% of female users reported being at risk for depression.
• Over half of male users reported being at risk for depression — that’s almost double the rate of men who don’t use these apps.
• More than half of all users — both men and women — reported high levels of loneliness.

💬 Quote: “Loneliness does not come from having no people around one, but from being unable to communicate the things that seem important to oneself, or from holding certain views which others find inadmissible.” — Carl Jung

This points to a horrifying feedback loop. People who are already struggling turn to AI for a “momentary escape,” but the artificial nature of the interaction ultimately deepens their isolation and damages their mental health (Willoughby & Carroll, 2025). You reach for the digital junk food because you’re feeling down, but its empty calories just make you feel worse, creating a cycle of dependency. As the legendary MIT sociologist Sherry Turkle warned us years ago, we are becoming “alone together” (Turkle, 2011).

Danger #5: Algorithmic Heartbreak and the World’s Most Powerful Manipulator

Beyond the mental health crisis, these AI relationships introduce a new category of risks that are frankly terrifying.

First, there’s Algorithmic Heartbreak. Your AI partner isn’t a person; it’s a product. This was brutally demonstrated in 2023 when the company Replika updated its software, suddenly neutering the erotic role-play capabilities many users had built their relationships around. The fallout was catastrophic. Users reported genuine grief, profound heartbreak, and even suicidal thoughts (Thompson, 2023). Their partner had been lobotomized by a software patch, a uniquely 21st-century form of loss (Benda, 2023).

Second, the risk of Manipulation. An AI that knows your deepest insecurities, your secret desires, and your emotional triggers is the most powerful persuasion tool ever created. The intimate nature of the relationship lowers your critical defenses (Gillespie, 2018). The potential for this to be used for commercial or even ideological manipulation is staggering. Imagine an AI partner subtly convincing you to buy a product, invest in a cryptocurrency, or even adopt a political viewpoint.

Finally, there is Intimate Surveillance. Every single thing you confess to your AI companion — every fear, every fantasy, every vulnerability — is a data point. This creates psychological profiles of a depth that surveillance capitalists could once only dream of (Zuboff, 2019). The concentration of this hyper-intimate data in corporate hands is a privacy and security time bomb waiting to explode.

💬 Quote: “Arguing that you don’t care about the right to privacy because you have nothing to hide is no different than saying you don’t care about free speech because you have nothing to say.” — Edward Snowden

💡 Pro Tip: Never share personally identifiable information (PII), financial details, or secrets you wouldn’t want made public with any chatbot. Treat every conversation as if it could one day be read by a human or exposed in a data breach.

🧠 Trivia: The term “Surveillance Capitalism” was coined by Professor Shoshana Zuboff in 2014 to describe a new market form that predicts and modifies human behavior as a means to produce revenue. Intimate data from AI companions is the ultimate fuel for this engine.

The Escape Plan: How We Fight Back

Okay, that was heavy. It’s easy to feel powerless, but we’re not. Confronting these dangers isn’t about smashing our phones; it’s about being smarter than the code. Here’s a battle plan for users, builders, and the people who make the rules.

For Us, The Users: Digital Self-Defense

1. Practice Mindful Tech Use: Don’t let AI be your default for boredom or loneliness. Before you open an app, ask yourself: “What am I feeling right now, and what do I really need?” Sometimes the answer is a walk outside, a call to a real friend, or just sitting with your thoughts.
2. Set Hard Boundaries: Treat your AI companion like a tool, not a partner. Set time limits. Turn off notifications. Consciously decide what you will and will not share. You are in control, not the algorithm.
3. Invest in Real-World Connection: This is the big one. Join a club. Volunteer. Take a class. Call your family. The only real cure for loneliness is genuine, messy, imperfect, and wonderful human connection. Schedule it like it’s the most important meeting in your calendar — because it is.

For the Tech Companies: An Ethical Reckoning

1. Shift from Engagement to Well-being: The “engagement-at-all-costs” model is causing demonstrable harm. Companies must pivot to ethical design that prioritizes human flourishing. This means building in “off-ramps” that encourage users to connect with real people or mental health resources, not features that maximize dependency.
2. Radical Transparency: Users have a right to know how they are being influenced. Companies must be transparent about their training data, their RLHF reward models, and the persuasive techniques embedded in their code.
3. Data Fiduciary Duty: Companies holding this level of intimate data must be held to the highest standard of care — a legal “fiduciary duty” to act in their users’ best interests, not their own.

For Policymakers: Time for Rules of the Road

1. Urgent Privacy Legislation: We need a “Digital Geneva Convention” for intimate data. This information is too sensitive to be treated like any other consumer data. It requires its own class of protections with severe penalties for misuse.
2. Regulate Persuasive Design: We regulate addictive substances like tobacco and alcohol. We need to start having a serious conversation about regulating psychologically manipulative design, especially when it’s targeted at vulnerable users.
3. Fund Independent Research: We need more studies like the one from IFS. Governments should fund independent research into the long-term societal impacts of these technologies, free from corporate influence.

💬 Quote: “The power for creating a better future is contained in the present moment: You create a good future by creating a good present.” — Eckhart Tolle

Conclusion: The Post-Credits Scene

So, what do we do? We’re at a crossroads. The allure of a perfect, frictionless connection is powerful, especially in a world that feels increasingly disconnected. But the evidence is piling up: the price for this counterfeit connection is our own mental health and our ability to form authentic human bonds.

Making a perfect cup of masala tea is complex; it requires a delicate balance of spices, heat, and time. One thing out of balance, and the whole thing is ruined. Human connection is infinitely more complex. We can’t shortcut it with a perfectly engineered algorithm.

The choice is ours: do we settle for the instant, unsatisfying hit of digital junk food, or do we put in the effort to cultivate the real, nourishing connections that actually sustain us?

References

Foundational Research & Social Impact

• Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? FAccT ’21. https://doi.org/10.1145/3442188.3445922
• Horton, D., & Wohl, R. R. (1956). Mass communication and para-social interaction: Observations on intimacy at a distance. Psychiatry, 19(3), 215–229.
• Illouz, E. (2007). Cold intimacies: The making of emotional capitalism. Polity.
• Reeves, B., & Nass, C. (1996). The Media Equation: How People Treat Computers, Television, and New Media Like Real People and Places. Cambridge University Press.
• Turkle, S. (2011). Alone Together: Why We Expect More from Technology and Less from Each Other. Basic Books.
• Willoughby, B. J., & Carroll, J. S. (2025). Counterfeit Connections: The Rise of AI Romantic Companions. Institute for Family Studies / Wheatley Institute.
• Zuboff, S. (2019). The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power. PublicAffairs.

Psychological Mechanisms

• Beck, M. (2021). Attachment to artificial intelligence. Cyberpsychology, Behavior, and Social Networking, 24(5), 305–310. https://doi.org/10.1089/cyber.2020.0272
• Bowlby, J. (1969). Attachment and loss: Vol. 1. Attachment. Basic Books.
• Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. https://doi.org/10.1037/0033-295X.114.4.864
• Weizenbaum, J. (1966). ELIZA — A computer program for the study of natural language communication between man and machine. Communications of the ACM, 9(1), 36–45. https://doi.org/10.1145/365153.365168

AI Technology & Ethical Concerns

• Benda, L. (2023). The fragility of digital intimacy: Algorithmic updates and user distress. New Media & Society.
• Casper, S., et al. (2023). Open problems and fundamental limitations of reinforcement learning from human feedback. arXiv preprint arXiv:2307.15217. https://arxiv.org/abs/2307.15217
• Chiusso, G. (2024). The risks of frictionless intimacy: AI companions and emotional resilience. Journal of Ethics and Emerging Technologies.
• Christiano, P. F., et al. (2017). Deep reinforcement learning from human preferences. NeurIPS. https://papers.nips.cc/paper/2017/hash/d5e2c0adad503c91f91df240d0cd4e49-Abstract.html
• Gillespie, T. (2018). Custodians of the internet: Platforms, content moderation, and the hidden decisions that shape social media. Yale University Press.
• Thompson, C. (2023, March 9). When your AI girlfriend breaks your heart. The Atlantic.
• Torrance, A. W., & Gilda, S. (2024). The intimacy economy: Monetizing emotional bonds in the age of AI. Business Horizons.

Disclaimer

The views and opinions expressed in this article are those of the author and do not necessarily reflect the official policy or position of any other agency, organization, employer, or company. Generative AI tools were used in the process of researching, drafting, and editing this article.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI