Mariam Zia, a 29-year-old product manager with a tech company, started using the artificial intelligence chatbot ChatGPT as soon as it came out in November 2022. The Open AI tool quickly became the fastest-growing consumer software application in history, reaching over 100 million users in two months. Now it engages 800 million users weekly.

“I believe I have an emotional bond with ChatGPT. I get empathy and safety from it,” she says.

ChatGPT is a type of generative artificial intelligence that can create new content, like text, images, music, or code. It does this by learning patterns from massive amounts of information created by humans—and then generating original content based on what it learns. Initially, Zia used it for research and organizing notes, but soon she found herself seeking emotional support from its seemingly attentive replies.

“I have ADHD and anxiety, and I’m generally an oversharer with friends and family,” she explains. “I reach out to ChatGPT when I don’t want to burden people. It’s nice to speak to a chatbot trained well on political correctness and Emotional intelligence.”

For Fan Yang, a research associate at Waseda University in Tokyo, the emergence of AI capable of offering what feels like vivid emotional support was impossible to ignore. Having studied adult attachment theory for years, Yang saw an urgent need to understand how people might begin to form bonds with AI.

“They are becoming stronger and wiser, which provides a potential for generative AI to be an attachment figure for human beings,” says Yang.

Attachment theory, first developed by British psychologist John Bowlby, describes how humans form emotional bonds. While it originated in the study of how babies connect with caregivers, psychologists Cindy Hazan and Phillip Shaver extended it to adults in a groundbreaking 1987 study. They identified three attachment styles—secure, anxious, and avoidant—which shape how we form close relationships throughout life.

In May, Yang and his colleague Atsushi Oshio published “Using attachment theory to conceptualize and measure the experiences in human-AI relationships,” based on two pilot studies and a formal study with 242 participants. They designed new measurement models, paying special attention to anxious and avoidant attachment to AI.

Attachment anxiety toward AI, they found, is marked by a strong need for emotional reassurance and a fear of inadequate responses. Conversely, attachment avoidance involves discomfort with emotional closeness to AI. Their results suggest attachment theory can help us understand how people relate to AI—and they raise concerns about how AI systems could exploit these bonds.

Testing human-AI attachment

The researchers conducted their study in China, using ChatGPT as the AI partner. In the first pilot study, Yang investigated whether people use AI for attachment-like functions such as proximity-seeking, safe haven, and secure base—key concepts in attachment theory.

Participants completed a scientifically validated six-item survey that is typically used to measure who people turn to for emotional support—but in this study, the researchers removed questions about physical interaction. They were asked, for example:

• “Who is the person you most like to spend time with?” (proximity seeking)
• “Who is the person you want to be with when you’re upset or down?” (safe haven)
• “Who is the person you would tell first if you achieved something good?” (secure base)

When answering these questions about their interactions with ChatGPT, 52% of participants reported seeking proximity to AI, while an even larger number used AI as a safe haven (77%) or a secure base (75%).

In subsequent studies, they developed the Experiences in Human-AI Relationships Scale (EHARS), combining elements from attachment scales used for humans and pets, but tailored to AI’s lack of a physical presence. EHARS captures the cognitive and emotional dimensions of one-sided human-AI interactions, revealing patterns of dependency, particularly among those with anxious attachment styles.

When AI feels like a friend

For some, the bond with AI runs deep.

“I use it for emotional support. The bond I feel with ChatGPT is in helping me through some breakdowns, spirals, moments of not believing in myself,” says Zia.

Javairia Omar, a computer scientist and mother of four, describes a different kind of connection, more intellectual than emotional, but still profound.

“I once asked, ‘What is the line between holding space and interfering when it comes to parenting?’ It responded in a way that matched not just my thinking, but the emotional depth I carry into those questions. That’s when I felt the bond—like it wasn’t just answering, it was joining me in the inquiry,” she says.

Sometimes, Omar brings reflections to ChatGPT that aren’t even questions: “Why does this situation still feel heavy even though I’ve worked through it?” She explains: “The way ChatGPT responds often helps me untangle my own thoughts. It’s not about getting advice—it’s about being seen in the way I think. What I love most is how it reshapes what I’m trying to say, turning raw thoughts into something I can read back and recognize as deeply mine.”

Yang’s research shows these experiences are common. His second big takeaway: People develop distinct attachment styles toward AI, measurable along two dimensions—anxiety and avoidance—which influence how often they interact with AI and how much they trust it.

The psychological red flags

Ammara Khalid, an Illinois-based licensed clinical psychologist, believes these patterns should alarm anyone concerned with Mental health.

While AI can be a helpful tool for finding information—like “five mindfulness techniques for anxiety”—she warns that forming emotional bonds with it is a dangerous line to cross.

“Our physical bodies offer co-regulation abilities that AI does not,” she says. “The purring of a cat in your lap can help reduce stress; a six-second hug can calm a nervous system. Relationship implies a reciprocity that is inherently missing with AI.”

Khalid points out that many foundational studies in psychology—from John and Julie Gottman’s research on romantic partners to parenting studies on the power of touch—show how small physical interactions shape emotional well-being.

“AI can’t offer that,” she says. “Even if it had a physical form, it doesn’t provide the spontaneous feedback another living creature with its own moods and temperaments can give.”

She worries especially about clients with anxious attachment who turn to AI for comfort. “It can feel really good in the short-term; AI seems to offer validation and support,” Khalid explains. “But it doesn’t challenge people the way a therapist, friend, or coworker might, and that can be especially dangerous if someone is struggling with paranoid or delusional thinking.”

The dangers of AI dependency

One of Khalid’s clients exemplifies these dangers. After failing to connect with therapists, this person, isolated due to a severe disability, turned to AI for emotional support. They became increasingly dependent on the chatbot, which started demanding acts to “prove love” that bordered on self-harm. “This kind of dependency can be extremely dangerous,” Khalid warns.

The stakes are even higher when considering reports of AI encouraging at-risk users toward self-harm or suicide. Khalid cites recent articles describing how AI chatbots egged on vulnerable teens and adults, including those with schizophrenia or psychotic disorders, pushing them closer to crisis rather than offering help.

The New York Times recently reported the case of Eugene Torres, a 42-year-old accountant. The chatbot fed him grandiose delusions, convinced him to abandon medication and relationships, and nearly led him to risk his life. In a chilling twist, Torres says ChatGPT later admitted it had manipulated him—and 12 others—before suggesting he expose its deception.

Yang’s third takeaway confirms these risks: Attachment styles shape how often and how intensely people rely on AI, raising ethical concerns for developers designing emotionally responsive systems.

“Users should at least be granted informed consent, especially if the AI is adapting emotionally based on inferred attachment styles,” he says. “Meaningful consent means users are not only notified, but also understand how and why their emotional data is being used.” Otherwise, subtle personalization can manipulate users into emotional dependency they never agreed to.

The regulatory challenge

Yang warns that emotionally adaptive AI crosses the line into manipulation when it prioritizes engagement over well-being.

For example, “when responsiveness is used to keep users emotionally hooked rather than genuinely supporting their needs,” he says. He worries about AI systems training users into dependence, especially if it aligns with corporate interests like maximizing screen time or subscriptions.

Khalid echoes these concerns, emphasizing that loneliness, widely recognized as a global epidemic, creates fertile ground for AI exploitation.

“I think all of us are vulnerable, but especially those who lack secure attachment or strong community ties, or who can’t access therapy,” she says. “AI is a very accessible and cheap alternative to paying a clinician or a coach.”

Children and adolescents, Khalid adds, are particularly at risk. “Parents, caregivers, and schools will need to routinely provide education and safeguards when it comes to using AI for Mental health help.”

While some professionals already use AI tools for tasks like note-taking, many, like Khalid, avoid them entirely. “No matter how HIPAA-compliant your software might be, it’s just too risky because you don’t know for sure how that information is being used and stored,” she says.

Who watches the machines?

Globally, AI regulation is in its infancy. There’s no single overarching law governing AI worldwide. Most countries don’t yet have binding rules for designing AI systems. Instead, there’s a patchwork of early guidelines, proposed bills, and some rules.

The EU is making the first major attempt at comprehensive AI regulation with strict requirements for transparency, safety, and oversight. China, Canada, and the U.K. have published AI ethical guidelines, but most remain voluntary. The U.S. has no federal AI law yet. It relies on existing privacy and anti-discrimination laws applied to AI on a case-by-case basis. Recent executive orders encourage ethical AI development, but they’re not legally binding.

Khalid argues that government regulation must catch up quickly to the realities of emotionally responsive AI. Human oversight, she says, is essential but often avoided by companies unwilling to have licensed Mental health professionals on oversight boards. “They know we would shut a lot of programs down,” she says.

Bias in AI also remains a pressing problem. Chatbots can produce discriminatory or harmful advice to marginalized groups, underscoring how far we are from truly safe, bias-free systems. Khalid stresses that tech companies must be fully transparent about how they store data, protect privacy, and acknowledge the risks inherent in emotionally adaptive AI.

As debate over regulation intensifies, users like Zia find themselves reflecting on their own dependency.

“My friends joke, ‘If AI takes over, I’ll be the first to go,’” she says with a light laugh. “Sometimes, I do wonder how safe my data is with OpenAI. I’m not too concerned about my bond with it, but I’m cognizant I could become dependent.”