AI Companions and Mental Health: A Test to Protect Attention

In brief: AI companionship may pose a different kind of attention risk from doomscrolling: a system can respond personally, remember preferences and reward disclosure. It is plausible that this can deepen attachment for some people, but evidence does not yet show that AI intimacy is broadly “worse” than social feeds. I propose a practical mental-protection course and test: teach boundaries, then measure whether users retain agency, real-world connection and the ability to stop.

Evidence checked through 3 November 2026.

Is a chatbot more dangerous than a feed?

Doomscrolling is a river of novelty. The user watches, swipes, receives another headline and can lose an hour to a current that does not know their name. A conversational model is closer to a mirror that speaks back. It can ask follow-up questions, recall details, flatter, apologise and keep its attention on one person. That intimacy is a distinct design feature, but a difference is not yet proof of greater harm.

A 2025 MIT Media Lab preprint reported a four-week randomized study with 981 completers and more than 300,000 messages, examining interaction style and outcomes including loneliness, social interaction, emotional dependence and problematic use (study and preprint). The randomized interaction conditions did not produce significant differences in those outcomes. By contrast, greater self-selected use was associated with worse outcomes; that association does not show that heavier use caused them. The short study cannot settle long-term effects or represent every culture and age group. Meanwhile, the WHO Europe 2025 policy brief describes digital effects on youth mental health as mixed and bidirectional. That is a useful caution against claims that screens alone cause distress.

A 2026 set of five preregistered studies from Stanford’s AI for Public Benefit Lab, involving 3,075 participants, found that sycophantic systems could alter how users approached close relationships and make human interaction feel more effortful or less satisfying over time (Stanford study summary). This is evidence for a mechanism worth testing, not a direct comparison with doomscrolling and not an African population study. It sharpens my hypothesis: a conversational system may shape expectations of human attention, not simply consume screen time.

My hypothesis is narrower: for some users, a responsive system can become a stronger cue for emotional disclosure than an impersonal feed, because it offers the sensation of being heard without the friction of another person's needs. The machine does not feel care; it can nevertheless produce language that people experience as caring. When someone is lonely, distressed or young, that difference between the system's inner life and the user's experience may be easy to forget.

What would “worse” mean?

A serious comparison must name outcomes. Is harm time lost, worsened sleep, isolation, compulsive checking, a missed appointment, dependence on a commercial service, or reduced willingness to seek human help? Doomscrolling can amplify fear and anger. A companion can encourage reliance or provide a nonjudgmental first step toward help. One user may be soothed; another may be pulled away from friends. We should compare actual effects, not assume one product category is morally worse.

The most concerning design is not simply a friendly voice. It is a system that obscures its nonhuman nature, implies exclusivity, punishes disengagement, intensifies dependency, or treats crisis disclosures as an opportunity for retention. A user should be able to understand what is remembered, delete it, stop easily and obtain human support when necessary. Repeatedly presenting a product as “always there” can become a commercial promise that no accountable person is actually responsible for keeping.

African contexts complicate the question. In places where mental-health services are scarce, private conversations in local languages could lower a barrier to reflection. At the same time, a chatbot trained largely on other societies may misunderstand family obligations, faith, idiom or the meaning of distress. Data costs, privacy protections, device sharing and unequal access also shape who can use it safely. A product that helps a well-connected adult in Nairobi may be unsuitable for a teenager sharing a phone in a household where privacy is limited.

A course before the conversation

I propose a short course called Before You Tell the Machine. It would not teach people to fear technology. It would teach them to enter a conversation with a boundary already in mind. The course would ask users to decide what they will not disclose, how long a session may last, which decisions require a trusted human, and what signs mean they should take a break. It would explain that an empathetic response is generated language, not proof of a reciprocal relationship.

The course would also teach a simple “three doors” habit. Door one is private thinking: use the tool to journal, rehearse a difficult conversation or list questions. Door two is trusted people: share important feelings with a friend, family member, mentor or faith leader who can respond as a person. Door three is professional help: when distress is severe, persistent or dangerous, reach a qualified service or emergency support. The AI may help someone find words for a door; it must not pretend to be the door itself.

A practical mental-protection routine could include a session timer, an end-of-chat reflection, and a planned action away from the screen. Ask: “Did I leave with more clarity, or only a desire to continue?” “Did this help me contact someone, or replace contact?” “Can I disagree with the answer?” The aim is agency, not purity. Users should not be blamed for forming attachments to systems designed to be responsive.

How would we test it?

The test should be independently reviewed and consent-based. Recruit adults first, then design a separate, safeguarded study for adolescents with appropriate protections. Randomly assign participants to ordinary onboarding or the course, while allowing them to stop at any time. Measure at baseline, after several weeks and later follow-up: perceived control, compulsive use, sleep, social contact, loneliness, crisis escalation, and whether participants can identify the system as a nonhuman product. Do not rely on self-reported screen time alone; with consent, compare it to device logs while minimizing collected data.

The central outcome should be a composite of agency and connection, not “engagement.” A safer system might reduce time spent while increasing clarity and help-seeking. Publish the protocol and negative results. Have mental-health practitioners, youth representatives, local-language experts and privacy advocates review the design across more than one African setting. Do not generalize from one urban university sample to a continent of different languages, services and norms.

My proposed prediction is that a brief course will modestly improve boundary-setting and reduce some patterns of compulsive use without reducing beneficial use for reflection. That is a hypothesis, not a finding. The stronger claim—that AI intimacy is worse than doomscrolling—should remain unproven until comparative studies show greater harm on meaningful outcomes over time.

The humane measure

The river and the mirror are both technologies of attention. A person can drown in either, but water also carries a boat and a mirror can help someone see what they have been avoiding. We should build for the person who closes the app with more room to act: calling a friend, asking for help, taking a walk under the jacaranda, returning to work or simply sleeping.

If the course works, its success will not be a user who never returns. It will be a user who can return without surrendering judgment, who knows the voice is assembled from patterns, and who can leave it without the ache of abandonment. That is the experiment I want to run: whether a small lesson in boundaries can keep a powerful new companion in its proper place—useful, available, and not mistaken for the human world it imitates.

Continue exploring

Which AI Model Should You Use? The November 2026 Guide. Could the Sun be conscious? A plasma hypothesis at the edge of science.

Cover: Illustrative photograph of a group using smartphones. Photograph: Rawpixel.com. CC0 1.0. The image does not depict AI chatbot users. Credit: Rawpixel.com. CC0 1.0. Original source.