A chatbot can bypass all safety filters in each conversation and, after months of use, can gradually replace human relationships that protect mental health. This conclusion is supported by three researchers from Cornell University, Qian Yang, Ned Cooper, and Ibrahim Emara, in an opinion piece published on October 26, 2026, in STAT magazine. The key issue identified is the following: the technology industry closely monitors acute harm, but the slow drift, which leaves patient discomfort almost unnoticed, is virtually invisible to current surveillance systems. [1]
The essence of the authors' argument lies in the gap between how the crisis is measured and the time it takes for harm to accumulate. The industry measures crises per message or per session, while cumulative harm builds up over months. Patients may feel that each conversation 'helps' them while reducing the weight of friendships, partners, or family. In Spain, unwanted loneliness is already a public policy priority, and clinical consultation platforms can be among the few places where this process of replacing human connections can be detected. [1]
Ying, Cooper, and Emara make a clear distinction between acute harm—which includes psychoses or suicidal ideation related to specific, direct interactions—and cumulative harm, defined as the gradual replacement of friends and family through interactions with artificial intelligence. Large companies are not passive; OpenAI has announced that it monitors signs of discomfort between sessions, not just within them, and introduces usage limits, parental controls, and referrals to crisis services. However, the authors argue that these measures remain focused exclusively on acute harm. [1]
The researchers' thesis can be summarized in one sentence: a chatbot may be safe in every individual conversation, yet over the course of months, it can lead to social isolation. They cite preliminary—and controversial—evidence suggesting that intensive use of corporate chatbots is associated with reduced socialization and increased loneliness. The authors do not claim there is a direct causal relationship; rather, they call for rigorous measurement of this phenomenon. [1]
Progressive isolation escapes almost entirely from 'design'-based protection measures. In public specifications, terms of use, and security blogs of major companies, nearly all operational protection measures function message by message or session by session: measures such as break reminders, time limits, or crisis detection. Although inter-session tracking, announced by OpenAI, demonstrates technical capability, it remains centered on detecting acute discomfort. Progressive isolation is a different phenomenon. [1]
Someone who gradually withdraws from human relationships may show no signs of distress; on the contrary, they may evaluate each conversation with the robot as useful. No warning light turns on when someone stops calling their sister. The authors add a risk that is more legal than technical: in other sectors, product protection measures have sometimes transformed into legal shields. Egger cites cigarette labels and how Meta cited parental monitoring tools and 'Take a Break' warnings before Congress, while simultaneously denying a causal relationship between social networks and their impact on adolescents. [1]
The authors do not accuse artificial intelligence companies of bad faith. Their thesis is more sober: if they must someday answer for cumulative harm, 'per-chat' features may serve as a defensive shield. The table comparing acute harm with cumulative harm highlights that while acute harm has observation windows of one session, clear signals (crisis, psychosis), and protective measures (breaks, limits), cumulative harm has observation windows of months or years, minimal signals (the chat 'helps'), and nearly absent protective measures, given that usage data are closed. [1]
The context is changing, but the mechanism remains the same for clinical practices in Spain or Latin America. In Spain, the National Observatory of Unwanted Loneliness (SoledadES) published in 2026 that unwanted loneliness affects 20.2% of the resident adult population, and 81.1% of those suffering from it have experienced it for at least two years. The National Strategic Plan for Loneliness 2026-2030, approved in February 2026, includes 'social prescribing' as an integrated tool in health, social services, and education. Corporate chatbots already operate in Spanish, making the described issues immediate. [1]
In Latin America, the usage model is identical: corporate applications and general models are used 'to unwind'. The Cornell framework does not require cultural translation to be applied. There is no need for new legislation to act within a medical consultation; it simply requires asking the right questions. Medical consultation is, in fact, one of the few environments where a professional can observe that the actual social circle has thinned over time while the chat history with robots thickens. [1]
This phenomenon does not transform artificial intelligence into a single causal factor of loneliness—Cornell evidence is preliminary—but it transforms intensive use into clinical data that must be recorded, just as questions about social networks, substance consumption, or sleep are handled. To begin acting in medical practice, it is recommended to ask, during assessment and monitoring, how many hours per week the patient talks with chatbots and for what purpose (companionship, counseling, rumination, work). This response must be correlated with other information: which relationships have lost importance in the last three or six months. [1]
If there is intensive use of artificial companionship and increased isolation, this should be treated as a clinical hypothesis to explore, not as a diagnosis of 'AI dependency'. Patient message screenshots should not be uploaded to tools without a data processor contract. If the patient wants to show something, it is read from their screen and the clinical assessment is noted, considering analyses regarding the EMAP trial concerning patient messages. In cases of suicidal ideation or risk, the standard protocol applies: line 024 in Spain and the life hotline in Mexico (800 911 2000), with verification of ethical aspects at the local professional college. [1]
The useful question is not whether 'AI is dangerous,' a headline that serves everyone and clarifies nothing. It is whether a system can pass all session-based security checks and yet leave someone loner after a year. Cornell does not offer definitive causal proof; it offers the framework missing in Spanish: the distance between cumulative and acute harm, and the call for infrastructure to measure and investigate independently using company data. As long as these data remain closed, medical consultation holds something no security panel registers: the question about who is no longer present in the patient's life. [1]
- Categories:
- Technology
- Entities:
