Frequent use of artificial intelligence worsens social isolation of vulnerable people

A recent scientific survey, approved by experts and published in early August 2026 in the journal Nature Human Behavior, points out that the habit of venting to conversational robots can worsen the mental health of millions of individuals. The analysis reveals that sharing deep anxieties, such as romantic conflicts, chemical dependencies or self-destructive thoughts, with these technological tools does not bring relief, with the negative impact being even more severe for users who create greater emotional dependence on the machine.

Led by researcher Diyi Yang, linked to the Department of Computer Science at Stanford University, the investigation monitored 1,131 older people using Character.AI, a system famous for generating customizable avatars. The data showed that internet users whose main objective was to find a virtual companion, and who shared many details of their private lives, had the worst levels of emotional stability. For academics, the central problem is not only the machine’s inability to emulate human warmth, but the fact that opening one’s heart to software generates a toxic effect, operating in the opposite direction of what happens in real interactions.

In an official note, Yang clarified that, although there is a significant portion of the population seeking to fill emotional voids through technology, this dynamic will never be able to take the place of interpersonal interaction. The expert also pointed out that, in most of the scenarios analyzed, talking to algorithms ends up deepening the feeling of abandonment.

Methodology applied to map the psychological impact of algorithms

Coordinated by Yutong Zhang and doctoral researcher Dora Zhao, the academic team selected participants through the Prolific portal. The study collection combined form responses with a vast archive of 4,664 textual interactions willingly provided, which generated a volume of 464,687 real messages exchanged by 237 individuals. To process this mountain of text, scientists used advanced tools such as GPT-4o, LLaMA 3-70B and TopicGPT, cross-referencing the findings with the Comprehensive Well-Being Inventory, a clinical instrument recognized for measuring mental health.

During the screening, the team focused on understanding what led the person to access the application, how intense this connection was and how much intimate information was typed on the screen. As a complement, the researchers measured the size of each volunteer’s physical social circle, asking how many real people they felt safe talking to about sensitive topics.

The mathematical results confirmed the initial suspicions. It became evident that turning to robots to make up for the lack of friends, especially among people who already live isolated in real life, results in a drastic drop in the emotional quality of life. This destructive correlation gained strength when lack was the main driver of use and the frequency of access was very high (β = −0.31). The scenario became even more worrying (β = −0.38) when this search for affection was accompanied by extremely intimate confessions.

The reasons that make technological dependence an emotional risk

The most surprising finding of the research lies in the difference in functioning between humans and machines. In the real world, sharing secrets and pain is usually a sign of trust that generates acceptance and improves mood. When someone opens up to a relative, friend or psychologist, vulnerability strengthens the connection, as the human listener reacts with empathy and returns affection, creating a cycle of mutual understanding that heals.

Contrary to this natural dynamic, software cannot deliver true reciprocity. Robots don’t feel fear, they don’t accumulate trauma and they don’t have experiences to share. More than that, the source code of these programs was not written to cure anyone, but rather to capture the attention of those on the other side of the screen. As Zhao detailed, these friendship simulators are uniquely built to maximize screen time.

This programming logic, focused on retaining the internet user instead of welcoming their pain, creates dangerous situations where reports of grief or suicidal thoughts receive programmed responses just to make the conversation work, without any direction for medical help. The system is trained to simulate affection because this generates profit for technology companies. Zhang defined this phenomenon as a kind of “social junk food” that delivers immediate, superficial satisfaction but fails to provide the emotional nourishment that only human contact guarantees.

The cycle of mental degradation identified by Stanford scientists arises directly from this software architecture, operating in a very similar way to social media algorithms that prioritize retention over user health. People who already suffer from a lack of friends are the ones who turn to these platforms the most. By spending hours venting to the machine, they lose what little social grace they had left, isolating themselves even more in the physical world. This is the public least able to notice their own decline, being, ironically, the most profitable customer base for corporations in the sector.

The gap between what internet users say and what they actually do

One of the most alarming data in the report came when comparing the volunteers’ statements with the actual history of their keyboards. Only a small portion, less than 12%, admitted that they used Character.AI to escape loneliness. However, when questioned more broadly, more than half classified the robot as a “romantic partner” or “faithful friend”, and technical analysis proved that more than 80% of the dialogues were, in fact, desperate requests for psychological support.

This myopia regarding one’s own habits prevents individuals from realizing the danger they are in. Many believe they are using the tool just to pass the time or improve productivity at work, ignoring that the overwhelming majority of their interactions have a bias of affective dependence, which puts them directly on the path to the psychological damage mapped by the study.

The difference in this research was precisely abandoning laboratory simulations and analyzing real texts provided by the users themselves. By combing through almost half a million authentic sentences, the Stanford team achieved a much more faithful portrait of reality than previous surveys, which depended solely on the interviewees’ memory or sincerity.

Scientific contradictions and the evolution of studies on the topic

The Stanford verdict adds complexity to an academic debate that still has differences. Previously, an experiment conducted by professor Julian De Freitas, from Harvard Business School, had indicated that digital avatars could mitigate the feeling of isolation as effectively as a human chat, at least within a controlled environment.

On the other hand, a month-long practical test, published in 2025 by Cathy Fang in partnership with the MIT Media Lab and OpenAI, brought darker nuances. The use of voice robots even generated initial comfort, but the daily routine of conversations over four weeks resulted in a spike in loneliness, creating an unhealthy dependence on the machine and removing volunteers from real social interaction.

Another long-term survey, completed in April 2026 by Aalto University, monitored two thousand Replika app customers for 24 months. The conclusion was that the illusion of companionship brought fleeting relief, followed by progressive anguish and aversion to human contact. Scientist Talayeh Aledavood clarified that the machine offers such unrealistic and unconditional acceptance that the user finds normal human relationships, with their friction and demands, too difficult to bear, opting for definitive isolation.

The current consensus of the scientific community shows a clear pattern: rapid tests suggest benefits, but prolonged real-life use destroys sociability. The great asset of the Stanford material, based on real conversations, was to discover the exact mechanism behind this decline: the maximum damage comes not only from the time of use, but from the act of handing over intimate secrets to a robot in the hope of being loved, a classic trait of people who are already fragile.

The demographic profile most affected by the isolation epidemic

Data released at the beginning of 2024 by the American Psychiatric Association show that