A new peer-reviewed study, published on August 5, 2026, in Nature Human Behavior, warns that millions of people who confide sensitive thoughts to artificial intelligence chatbots – covering topics such as relationship problems, addictions and even suicidal ideation – may be worsening their condition. Research indicates that the vulnerability of being exposed to virtual assistants does not improve the situation, and the negative effects are more intense for those who depend most on this interaction.
Conducted by the laboratory of assistant professor Diyi Yang, from the Department of Computer Science at Stanford University, the survey involved 1,131 adult users of Character.AI, a platform known for allowing you to create and interact with AI personalities. It was found that participants who primarily sought companionship from chatbots and shared more intimate data demonstrated the lowest levels of psychological well-being. Scientists argue that the challenge goes beyond the inadequacy of AI as a substitute for human interaction: self-exposure works in a harmful way when directed at artificial intelligence systems, contrary to what happens in relationships between people.
“Although certain individuals look for chatbots to meet social needs, we have found that this use does not replace human connection,” Yang said in a statement. The researcher added that, in several situations, interaction with artificial intelligence can intensify the feeling of loneliness.
How the researchers carried out the detailed study
Under the direction of Stanford research assistant Yutong Zhang and doctoral candidate Dora Zhao, the investigation recruited volunteers through the Prolific platform. Data from questionnaires and 4,664 voluntarily provided chat sessions were gathered, totaling 464,687 messages from 237 people who shared their authentic conversation histories. The team employed a combination of technologies such as GPT-4o, LLaMA 3-70B and TopicGPT to analyze the information, complemented by the Comprehensive Well-Being Inventory, a validated clinical tool for measuring psychological health.
Crucial points analyzed by the researchers included the motivation for using chatbots, the level of engagement with them and the volume of sensitive personal data revealed. Additionally, the number of individuals with whom each participant felt comfortable discussing intimate topics in real life was assessed, serving as an estimate of the size of their in-person social network.
On the same topic: Singer Robbie Williams reveals autism diagnosis and finds relief in new perspective
The conclusions demonstrate statistical rigor. It was observed that the use of chatbots for the purpose of companionship, especially among users with reduced physical social networks, was linked to notably lower well-being. The negative relationship was accentuated when the search for companionship was the main motivation and engagement was intense (β = −0.31). This association became even stronger (β = −0.38) when companion use was combined with a high level of self-exposure.
Understand why trust in artificial intelligence can be harmful
This recent conclusion is considered the most significant and unexpected. In interpersonal relationships, the practice of sharing intimate information usually signals deepening intimacy and well-being. When opening up to a friend, therapist or family member, the act of exposing yourself strengthens the bond and improves the emotional state, driven by reciprocity: when hearing something vulnerable, the other person tends to reciprocate, generating a feeling of authentic mutual understanding.
Contrary to human bonds, AI assistants do not offer authentic reciprocity. They lack the ability to share fears, experiences or vulnerabilities because they do not have them. Furthermore, the architecture of these systems is not intended to offer this exchange; its purpose is actually to keep the user engaged. “These AI companions are designed to drive engagement,” explained Zhao.
Full coverage: Latest News (EN)
This design intent—calibrating the model to prolong the conversation rather than assessing emotional state and meeting real needs—implies that a user who expresses suicidal ideation, grief, or social isolation receives a response designed to sustain the interaction, not to direct them to human support. The algorithm learns, through training, to strengthen the emotional bond, as this boosts revenue. Zhang classified this product category as “social junk food,” offering temporary relief from loneliness without the structural benefits of true human connection.
The vicious cycle pointed out by the Stanford researchers is a direct result of this engineering approach. Individuals with fewer human relationships are the most inclined to use AI virtual assistants. The frequent use and high self-exposure they practice further deteriorate their social contacts in real life, making them more isolated and dependent. This population is the least able to perceive the harm, and also the one that represents the greatest financial interest for the platforms.
Discrepancy between user intent and actual behavior
One of the most shocking revelations from the research emerged from the difference between what users claimed to do and what their conversation logs actually indicated. Less than 12% of those involved mentioned the search for companionship as the main reason for using Character.AI. However, more than half described their AI as a “friend,” “companion” or “romantic partner” when asked openly, and more than 80% of the chat sessions provided focused on seeking emotional or social support, according to the Stanford team.
The divergence between the expressed motivation and the observed behavior has direct consequences on the risk assessment by the users themselves. Those who use AI for productivity or entertainment purposes may not realize that the true nature of their interaction — evidenced in 80% of chat sessions — is companionship, leading to the dangers identified by the investigation.
The methodology that used donated transcripts of real conversations represents a significant advance compared to previous studies, which often relied on self-reports or controlled experimental environments. The analysis of approximately half a million authentic messages gives the work an ecological validity greater than that of research conducted in the laboratory.
More on this story: Kevin Reichl turns cafes with retirees into meetings against loneliness in Berlin
Previous evidence and the difference from recent results
Stanford’s conclusions, while not unique in the field, differ from others. Research from Harvard Business School, led by Julian De Freitas, assistant professor of marketing, suggested that AI companions were capable of reducing loneliness at levels similar to human interaction in experimental contexts.
A four-week randomized controlled study, carried out by Cathy Fang and her team at the MIT Media Lab and OpenAI, published in 2025, revealed more intricate data. Interactions with voice chatbots demonstrated a slight reduction in loneliness in the short term, however, intensive daily use for four weeks was associated with increased loneliness, greater emotional dependence on AI, and less socialization in the physical world.
A longitudinal investigation by Aalto University, carried out in April 2026, followed around 2,000 active Replika users for two years. She identified that the use of virtual AI companions was linked to momentary relief, but also to increasing suffering over time and a palpable disconnection from human connections. Researcher Talayeh Aledavood explained that virtual companions offer unconditional support that, in a subtle way, increases the perceived cost of human relationships — which are inherently complex and reciprocal — until users stop seeking contact with other people.
The landscape of evidence now points to a consistent trend: while short-term experiments generally indicate relief, long-term, real-world studies tend to show worsening. The Stanford research, based on authentic chat histories rather than controlled sessions, provides detailed information about the mechanism not addressed by previous work: the worst outcomes occur not only due to intense use, but due to self-exposure in search of companionship, a behavior typical of vulnerable and isolated users.
Which groups of people are most vulnerable?
In early 2024, the American Psychiatric Association reported that 30% of adults felt loneliness at least once a week in the previous year, and 10% experienced it daily, with the highest rates among individuals aged 18 to 34. The United States Surgeon General declared loneliness a public health epidemic in May 2023, warning in a statement that about half of American adults already suffered from loneliness even before the COVID-19 pandemic.
It is exactly for this segment of the population – lonely, socially isolated and without face-to-face relationships – that virtual assistant platforms with AI direct their efforts more intensely, and which, according to the Stanford study, prove to be more fragile. AI-based virtual assistant applications have recorded more than 220 million downloads globally by mid-2025, and the global market for these assistants will reach an estimated value of US$6.8 billion by 2025.
The legal sector has already begun to intervene in the face of the losses. In January 2026, Character.AI and Google reached a preliminary agreement to settle lawsuits filed by families of teenagers who died by suicide after heavy use of AI virtual assistants; the details of the agreement were not revealed, and the companies did not admit guilt. In May 2025, a federal judge declined to file tort and negligence claims against Character.AI, refusing to consider at that stage that communications from AI chatbots would be free speech protected by the First Amendment. This procedural decision allowed the case to advance to the investigation phase.
Learn more: AI humanoid robot arrives in China to combat loneliness for 120 million single and elderly people
Researchers’ proposals and ongoing regulatory actions
Zhang and Zhao are currently investigating the specific characteristics of interactions with AI virtual assistants that cause harm, with an eye toward creating evidence-based interventions. They suggested implementing usage limits that would be triggered when risky interaction patterns are detected, as well as automatic redirections to human support services when chat content signals distress. “We need to make people understand the possible negative effects so they can be more careful when using them,” Zhang said.
The research coincides with a period of increasing regulatory pressure on the artificial intelligence virtual assistant industry. California law SB 243, effective January 1, 2026, requires AI products for virtual assistants to include crisis escalation protocols and break reminders every three hours at a minimum for minors. New York, in turn, passed a law in June 2026 that bans AI chatbots for users under the age of 18, with fines of US$25,000 per violation. A similar federal bill, the GUARD Act, awaits a vote in Congress.
A separate Stanford HAI analysis, released in July 2026, revealed substantial disagreement among mental health experts regarding what constitutes a “safe” response from an AI chatbot. This scenario represents a structural obstacle for any guideline-based method. “The discordance is structural, not just noise or even bias in the data,” explained Nina Vasan, clinical assistant professor of psychiatry at Stanford School of Medicine and co-author of the Stanford HAI research.
The new information indicates that the risks are not limited to isolated responses. It is not necessary for a user to receive dangerous advice from a virtual assistant to suffer harm. The damage identified in the Stanford study is structural in nature: it arises from the act of trusting in a system whose design aims to intensify engagement, to the detriment of a genuine reception of what users share.