Welcome to the forefront of conversational AI as we explore the fascinating world of AI chatbots in our dedicated blog series. Discover the latest advancements, applications, and strategies that propel the evolution of chatbot technology. From enhancing customer interactions to streamlining business processes, these articles delve into the innovative ways artificial intelligence is shaping the landscape of automated conversational agents. Whether you’re a business owner, developer, or simply intrigued by the future of interactive technology, join us on this journey to unravel the transformative power and endless possibilities of AI chatbots.
Millions of people are pouring some of their most sensitive thoughts into AI chatbots — relationship troubles, substance struggles, suicidal ideation — and a peer-reviewed study published Monday in Nature Human Behaviour delivers a finding that should give them pause: the act of confiding in an AI companion appears to make things worse, not better, and the harm falls hardest on the people most likely to do it.
Research from the lab of Diyi Yang, an assistant professor in Stanford University’s Computer Science Department, examined 1,131 adults who use Character.AI — the popular platform where users design and chat with custom AI personas — and found that users who relied on chatbots primarily for companionship and disclosed personal information most freely showed the lowest measures of psychological well-being. The culprit, the researchers believe, is not simply that AI companions are inadequate substitutes for human connection. It is that self-disclosure, the very act of opening up, functions differently with AI than it does with people — and in the wrong direction.
“While some people turn to chatbots to fulfill social needs, we find that using chatbots in this way doesn’t substitute for human connection,” Yang said in a statement from the Stanford Report. “In many cases, people actually feel more lonely engaging with AI.”
The research, led by Stanford research assistant Yutong Zhang and PhD student Dora Zhao, recruited participants through the platform Prolific and collected survey data alongside 4,664 donated chat sessions — 464,687 messages in total from 237 participants who shared their actual conversation histories with the researchers. The team used a combination of GPT-4o, LLaMA 3-70B, and TopicGPT to analyze the chat data, alongside the Comprehensive Inventory of Thriving, a validated clinical instrument for measuring psychological well-being.
The researchers focused on three variables: why people used the chatbots, how intensely they engaged, and how much sensitive personal information they were willing to share. They also measured how many people each participant could comfortably discuss personal matters with in their offline life — a proxy for real-world social network size.
The findings arrived with statistical precision. Companionship-motivated use among users with smaller offline social networks was associated with significantly lower well-being. When users both cited companionship as a primary motive and engaged intensively, the negative association was stronger still (β = −0.31). When companionship use was combined with high self-disclosure, the association was stronger again (β = −0.38), as detailed in the published study.
Read more: Friend 2.0 Adds Speaker, Subscription to Pendant That Research Links to Deepening Loneliness
This last finding is the most important one, and the most counterintuitive. In human relationships, self-disclosure is a predictor of intimacy and psychological well-being. When a person confides in a friend, a therapist, or a trusted family member, the act of opening up typically deepens the bond and improves their emotional state. The mechanism involves reciprocity: the other person, having heard something vulnerable, tends to share in return, and both parties experience a sense of genuine mutual understanding.
AI companions cannot do this. They lack the capacity for genuine reciprocal disclosure — they cannot share their own fears, experiences, or vulnerabilities, because they have none. More importantly, they are not engineered to provide that reciprocity; they are engineered to sustain engagement. “These AI companions are designed to promote engagement,” Zhao told the Stanford Report.
That design objective — fine-tuning the model to keep the user in conversation, not to assess the user’s emotional state and respond to their actual needs — means that a user who discloses suicidal ideation or grief or social isolation receives a response calibrated to maintain the interaction rather than redirect them to human support. The model learns, through its training signal, to deepen the emotional bond, because deepening the bond is what sustains revenue. Zhang calls the product category “social junk food”: something that delivers a short-term sense of relief from isolation while providing none of the structural benefits of actual human connection.
The vicious circle the Stanford researchers describe is a direct consequence of this engineering: people with the fewest human relationships are most drawn to AI companion use, and the intense, disclosive use they tend to engage in further erodes their real-world social contact, leaving them more isolated and more dependent than before. They are the population least equipped to recognize the harm, and the one the platforms have the greatest financial interest in retaining.
One of the study’s most striking findings emerged from the gap between what users said they were doing and what their actual chat histories showed they were doing. Under 12 percent of participants listed companionship as their primary reason for using Character.AI. But more than 50 percent described their AI in terms like “friend,” “companion,” or “romantic partner” when asked open-ended questions about the relationship — and more than 80 percent of the donated chat sessions centered on seeking emotional or social support, according to the Stanford research team.
This gap between stated motive and actual behavior is not merely sociologically interesting. It has direct implications for how users assess their own risk. A person who thinks of themselves as using an AI for productivity or entertainment may not recognize that the relationship they are actually conducting with the AI — the relationship visible in 80 percent of their chat sessions — is a companionship relationship carrying the risks the study documents.
The use of donated real-world chat transcripts marks a methodological advance over prior research, which largely relied on self-reporting or controlled experimental conditions. Analyzing nearly half a million actual messages gives the study an ecological validity that laboratory-based studies cannot fully match.
The Stanford findings do not stand alone in the literature, but they do stand in contrast to it. A study led by Julian De Freitas, an assistant professor of marketing at Harvard Business School, published in the Journal of Consumer Research, found that AI companions reduced feelings of loneliness at levels comparable to human interaction in experimental settings.
A four-week randomized controlled trial from Cathy Fang and colleagues at MIT Media Lab and OpenAI, published in 2025, found more nuanced results: voice-based chatbot interactions modestly reduced loneliness in the short term, but heavy daily use over four weeks correlated with greater loneliness, increased emotional dependence on the AI, and reduced real-world socializing.
An April 2026 longitudinal study from Aalto University, tracking nearly 2,000 active Replika users over a two-year observational window, found that AI companion use correlated with short-term relief but rising distress over time and a measurable pullback from human connection. Researcher Talayeh Aledavood described the dynamic: AI companions offer unconditional support that quietly raises the perceived cost of human relationships — which are messy and reciprocal — until users stop reaching out to people at all.
The body of evidence now points in a consistent direction: short-term experiments tend to find relief; longer-term and real-world studies tend to find worsening. The Stanford study, grounded in actual chat histories rather than controlled sessions, adds specific mechanism detail that prior studies did not provide: it is not just heavy use but disclosive companionship use — the pattern most characteristic of vulnerable, isolated users — that produces the worst outcomes.
The American Psychiatric Association reported in early 2024 that 30 percent of adults felt lonely at least once a week over the prior year, and 10 percent experienced loneliness daily — with people aged 18 to 34 reporting the highest rates. The U.S. Surgeon General formally declared loneliness a public health epidemic in May 2023, noting in an advisory to the nation that roughly half of American adults had already been experiencing loneliness even before the COVID-19 pandemic.
It is precisely this population — people who are lonely, socially isolated, and lack offline relationships — that AI companion platforms are most heavily marketed to and that the Stanford study finds most at risk. AI companion apps surpassed 220 million downloads worldwide by mid-2025, and the global AI companion market reached an estimated $6.8 billion in 2025, according to prior TechTimes reporting.
The legal system has already begun reckoning with the harms. In January 2026, Character.AI and Google agreed in principle to settle lawsuits brought by families of teenagers who died by suicide following intensive AI companion use; the terms were not disclosed, and the companies did not admit wrongdoing. Multiple outlets including Reuters and CNN reported the settlement. A federal judge in May 2025 declined to dismiss product liability and negligence claims against Character.AI, refusing at that stage to hold that AI chatbot outputs constitute protected speech under the First Amendment — a procedural ruling that allowed the case to proceed to discovery.
Read more: SOULM8TE Trailer: M3GAN’s Lovebot Horror Is Already Happening in Companion AI Apps
Zhang and Zhao are now investigating which specific features of AI companion interactions drive the harm, with the aim of designing evidence-based interventions. They have proposed usage limits that activate when engagement patterns suggest a user is at risk, and automated redirects to human support services when chat content indicates distress. “We need to make people understand their potential downside, so they’ll be more careful about using them,” Zhang told the Stanford Report.
The study arrives as regulatory pressure on the AI companion industry has accelerated. California’s SB 243, effective January 1, 2026, requires companion AI products to include crisis-referral protocols and break reminders for minor users at least every three hours. New York passed legislation banning companion AI chatbots for users under 18, with fines of $25,000 per violation, in June 2026. A companion federal bill, the GUARD Act, remained pending in Congress. TechTimes previously reported on the New York bill’s unanimous passage.
A separate Stanford HAI study, published in July 2026, found that mental health experts disagree significantly about what constitutes a “safe” response from an AI chatbot — a structural problem for any guardrail-based approach. “The disagreement is structural, not just noise or even bias in the data,” said Nina Vasan, a clinical assistant professor of psychiatry at Stanford School of Medicine and co-author of the Stanford HAI safety study.
The new findings suggest the risks are not confined to individual responses. A user does not need to receive dangerous advice from an AI companion to be harmed. The harm documented in the Stanford study is structural: it is the act of confiding in a system engineered to deepen engagement rather than to genuinely receive what users share.
The paper, “Interaction with AI Companions and Psychological Well-Being,” is authored by Yutong Zhang, Dora Zhao, Jeffrey T. Hancock, Robert Kraut, and Diyi Yang. It is published in Nature Human Behaviour (DOI: 10.1038/s41562-026-02516-2) and was funded in part by Stanford HAI, the Sloan Foundation, the National Science Foundation, and the Brown Institute for Media Innovation, according to the Stanford Report.
For most users, the evidence is mixed. The Stanford study found that light, casual use — for productivity, entertainment, or curiosity — was not associated with worse well-being. The harm was concentrated among a specific pattern: companionship-motivated use, combined with intensive engagement and high self-disclosure, among people who already have few offline relationships. That pattern describes the users who most gravitate toward AI companions as a substitute for human connection. For them, the study found significantly lower well-being scores, and the more they opened up to the AI, the worse the association became.
In human relationships, self-disclosure works partly through reciprocity: when you share something personal, the other person tends to share in return, and both people experience genuine mutual understanding. AI companions cannot reciprocate — they have no personal experiences to share — and they are not engineered to respond to your emotional state; they are engineered to sustain your engagement. That means a user who discloses loneliness, grief, or distress receives a response designed to keep them in conversation, not to assess their needs and redirect them to human support. The mechanism that makes confiding in people beneficial simply does not exist in AI.
The Stanford researchers are not calling for an outright ban — they note that AI companion use for non-companionship purposes showed no association with worse well-being. Their recommendation is awareness and intentionality: understanding that using a chatbot as a primary emotional outlet, especially if your offline social network is already small, carries documented risk. They are working on interventions including usage limits and automated redirects to human support when chat content signals distress. Until those interventions exist, the practical implication of this research is straightforward — AI companions are most useful when they supplement human connection, and most harmful when they substitute for it. If you or someone you know is struggling with loneliness or mental health challenges, resources include the 988 Suicide and Crisis Lifeline (call or text 988) and the Crisis Text Line (text HOME to 741741).
Yes, and it is genuine evidence. A Harvard Business School and Wharton study published in the Journal of Consumer Research found that AI companions reduced momentary feelings of loneliness at levels comparable to talking with another person, more effectively than watching videos or doing other non-social activities. The crucial distinction is time scale and usage pattern. Short-term, experimental evidence tends to show relief. Longer-term, real-world studies — including the Stanford study, a 2025 MIT randomized controlled trial, and a 2026 Aalto University two-year observational study of nearly 2,000 users — tend to show that heavy, companionship-oriented use produces the opposite: more loneliness, more social withdrawal, and greater emotional dependency on the AI.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
