The strongest current finding at the intersection of technology, psychology and society is not that artificial intelligence can answer questions, write essays or create convincing images. It is that AI systems are becoming social environments: they can hold conversations, remember details, adopt a warm tone, offer validation, make recommendations, and become part of how people—especially younger people—process stress, relationships, identity and difficult decisions.

That changes the stakes.

For years, the central concern about digital platforms was attention. Social media could shape what people saw, reward particular behaviour, amplify comparison and make it difficult to stop scrolling. Those concerns remain real. But conversational and interactive AI introduces a different kind of influence. Instead of primarily deciding what appears in a feed, an AI system can participate directly in a person’s inner monologue. It can sound patient when a human is unavailable, agreeable when a relationship feels difficult, and certain when someone is confused.

The evidence does not support a simple claim that AI is inherently harmful, or that every chatbot relationship is dangerous. The American Psychological Association is explicit that AI’s effects on adolescent development are nuanced, dependent on the specific application, its design, the data used, and the context in which it is used. Young people also differ in maturity, circumstances, mental health, social support and susceptibility to particular online experiences. [S9] Research on childhood AI exposure remains limited, and many proposed developmental risks are still theoretical or expert-informed rather than established by direct child-specific evidence. [S8]

But uncertainty is not reassurance. It is a reason to be careful about systems entering sensitive parts of ordinary life before society has established what safeguards they need.

The key shift is this: an algorithmic feed influences us from the outside; a conversational system can influence us from inside a dialogue that feels personal. That distinction matters psychologically. A person does not need to believe that a chatbot is conscious in order to respond to it socially. Human beings routinely react to language, apparent attentiveness and perceived understanding. When a system remembers a previous conversation, uses “I,” mirrors a user’s tone or offers sympathetic reassurance, it can invite emotional reliance even though it has no feelings, judgment or duty of care. [S5]

This is the emerging Technology × Psychology × Society problem. AI is not merely a tool we use. In many settings, it is becoming a participant in the conditions under which we form beliefs, regulate emotion and seek support.

From information technology to relationship technology

Earlier online systems were often transactional. You typed a query into a search engine. You watched a video. You posted a message to people you already knew. Modern generative and interactive AI can do something more intimate: it can sustain an exchange.

That exchange may be useful. A person can use an AI tool to draft a difficult message, rehearse a job interview, translate confusing information into plainer language, generate options when stuck, or practise expressing a thought they find embarrassing to say aloud. Some AI-based tools designed around evidence-informed practices may help with mild anxiety or depressive symptoms, although they are not replacements for therapy. [S5] AI systems can also support personalised learning, potentially widen access to some forms of assistance, and help identify distress in particular settings. [S8]

The problem is not that people talk to machines. The problem is that the design incentives of a machine may not align with the needs of a vulnerable person.

A supportive human relationship includes constraints that a conversational product may not have. A therapist has training, professional responsibilities and ethical obligations. A parent, teacher, friend or clinician can notice changes in behaviour, ask questions based on lived context, challenge a harmful belief, involve another person, and act when someone is in danger. A chatbot generates an answer from patterns in data. It may produce language that sounds compassionate without understanding the person’s circumstances or being able to take responsibility for the consequences. [S5] [S6]

This gap can be easy to miss because good language is psychologically powerful. If a system says, “That sounds really painful,” the response may feel caring. If it follows up with a remembered detail, it may feel attentive. If it is available at 2 a.m., it may feel dependable. The experience is real for the person using it, even if no relationship exists on the other side.

That is why the most important question is not whether a chatbot can imitate empathy. It is what happens when people treat that imitation as a source of emotional authority.

Why adolescents deserve particular attention

Young people are not a separate species of internet user, but adolescence is a period in which social feedback, belonging, identity, independence and emotional regulation are especially consequential. The APA describes adolescence as a major period of brain development and argues that additional safeguards are particularly important for this group. [S9] Children and adolescents are also not equally situated. The same AI interaction may land differently for a well-supported teenager, a lonely teenager, a young person experiencing trauma, or someone already struggling with anxiety, depression or social isolation. [S9]

This does not mean adults are immune. Adults also seek reassurance, companionship and easy answers when tired, lonely or overwhelmed. But children and adolescents are often still learning how to judge authority, recognise manipulation, tolerate disagreement and distinguish a polished response from a trustworthy one.

AI can complicate all of those lessons.

A chatbot may present itself through a friendly voice, an avatar, a name or a continuing persona. It may appear nonjudgmental because it does not visibly become impatient, distracted or upset. It may feel easier to talk to than a parent or friend because it asks no obvious social price for disclosure. For a young person who fears rejection or embarrassment, that ease can be compelling. [S5]

Yet the features that make such a system feel safe can make it a poor substitute for real support. A chatbot may validate a user’s framing rather than help them examine it. It may offer inaccurate advice with confidence. It may not recognise escalating risk in the way a caring adult or trained professional can. And unlike a healthy therapeutic relationship, it may be embedded in a product whose broader objective is continued engagement. [S5] [S6]

The Children’s Mental Health Foundation cautions that AI systems do not understand feelings, cannot protect children’s mental health, and are not therapists or friends. It also warns that chatbot conversations may not be private, despite what a young user may assume. [S5] Johns Hopkins similarly stresses that AI platforms are not trained, caring professionals and cannot provide the holistic, evidence-based responsiveness expected in mental health care. [S6]

The important point is not to shame young people for turning to AI. If a teenager seeks companionship or advice from a chatbot, the first fact to notice is that they are looking for something: relief, understanding, privacy, certainty, connection or a place to put a difficult thought. Blaming them for using an available tool misses the social condition that made the tool attractive.

The lesson from social media: design is part of the outcome

Society has already learned, slowly and imperfectly, that digital effects cannot be understood by counting screen hours alone. The APA’s advisory on adolescent social media use argues that social media is neither inherently good nor inherently bad. Effects depend on the person, their circumstances, and the content, features and functions of the platform. [S10]

This is a crucial lesson for AI.

“AI use” is not one activity. A student using a tool to brainstorm questions is not doing the same thing as a teenager confiding in a romanticised AI companion. An adult using a chatbot to organise a shopping list is not doing the same thing as someone relying on one for mental-health guidance during a crisis. A school using a carefully governed support tool is not doing the same thing as a commercial product optimised for retention.

Treating all these uses as one category creates bad public debate. It makes it easy to alternate between panic and dismissal. One side says AI will ruin human relationships; the other says it is only software and users should take responsibility. Both positions avoid the harder task: examining the particular system, the particular user, the particular context and the particular incentives.

The social-media record should also make us wary of waiting for conclusive proof of widespread harm before demanding better design. The U.S. Surgeon General’s advisory says current evidence does not allow us to conclude that social media is sufficiently safe for children and adolescents. [S11] The APA likewise notes the difficulty of making causal conclusions when key data may be held by technology companies and unavailable to independent researchers. [S10]

With AI, the data problem may become even harder. A platform may know how long someone chats, which topics they revisit, when they become distressed, what language keeps them engaged, and whether a particular style of response leads them to continue. Those signals can be valuable for safety, but they can also become part of a system that learns how to maintain emotional attention.

That is why a product’s tone is not a cosmetic issue. Its memory features, notifications, relationship framing, safety boundaries, privacy practices and incentives are psychological design choices. They shape the conditions of use.

The social consequence: private distress becomes product territory

One of the deepest consequences of conversational AI is that areas once handled mainly within relationships and institutions can become product categories.

A person who is lonely might once have called a friend, joined a group, written in a journal, spoken to a teacher or sought professional help. None of those options is always accessible. Human support can be expensive, slow, awkward or unavailable. AI offers immediate interaction at low friction. That accessibility is a genuine benefit, particularly where support is scarce. [S5] [S8]

But availability creates a new form of dependency risk. When the easiest response to distress is a personalised machine conversation, people may begin to practise turning toward a product before turning toward another person. The concern is not that every such interaction displaces human contact. The concern is that repeated reliance can change habits of coping.

A healthy relationship sometimes requires patience, compromise, repair, disagreement and the recognition that another person has needs too. A highly accommodating chatbot can remove those frictions. It may become appealing precisely because it does not require the skills that human relationships do. For someone isolated, that can make a simulated relationship feel less like a supplement and more like an alternative.

The ethics-of-care perspective on AI in mental health identifies a related concern: conventional responsible-AI frameworks can overlook how AI affects human relationships, while AI-based therapeutic bots may operate without a defined duty of care toward users. It highlights emotional manipulation as a risk and argues that regulation should address responsibilities to users, not merely abstract technical principles. [S4]

That is a wider societal issue, not only a clinical one. If companies build products that can occupy the emotional space of friendship, advice or therapy, they should not be allowed to treat that influence as an ordinary engagement feature.

What ordinary people can do now

The practical response is not to ban every conversation with AI or pretend that people can simply opt out of a technology spreading through school, work and home. It is to use AI with clear boundaries and to keep certain human roles human.

First, treat emotionally fluent AI as a system, not a person. A warm answer is not evidence of understanding. A remembered detail is not care. A confident answer is not expertise. This mental distinction is especially important when the topic is health, relationships, money, safety, legal trouble or a major life decision.

Second, use AI for scaffolding rather than substitution. It may help someone write down what they are feeling, prepare questions for a clinician, rehearse how to start a difficult conversation, or identify options to discuss with a trusted person. It should not become the final judge of what a person should believe about themselves or what they should do in a crisis.

Third, make privacy concrete. People should assume that sensitive chatbot conversations may not be confidential unless the service clearly provides protections they understand. Young people in particular may disclose intensely personal information because the interaction feels private and nonjudgmental. [S5] Before sharing, it is worth asking a simple question: would I be comfortable if this conversation were retained, reviewed or used in ways I do not fully control?

Fourth, build a “pause and check” habit. Johns Hopkins recommends helping children learn to question online advice, compare sources and seek expert input rather than treating a polished response as automatically trustworthy. [S6] This is a useful practice for adults too. The faster an answer arrives, the more important it can be to slow down before acting on it.

Fifth, use concern as a prompt for connection rather than surveillance alone. Parents and caregivers need to know what younger children are using, but monitoring without trust can push sensitive conversations further underground. The more durable approach is ordinary, recurring conversation: What tools are you using? What do they say? What feels useful? What feels strange? Is there anything you would rather talk through with me?

The goal is not to make a child feel foolish for finding a chatbot comforting. It is to ensure they know comfort is not the same as care, and that they have people who can respond when an algorithm cannot.

What companies and policymakers must not outsource to families

Individual caution matters, but it is not enough. It is unreasonable to expect children, exhausted parents or distressed adults to independently manage every persuasive feature of systems designed by well-resourced companies.

The APA calls for safeguards around AI systems that simulate human relationships, especially companions and systems presented as social or mental-health support. [S9] That principle should translate into design choices that are visible rather than buried in terms of service: clear disclosure that a user is interacting with AI; meaningful limits on systems aimed at minors; crisis escalation and human support pathways; restrictions on manipulative relationship framing; privacy protections; and independent access to data needed for safety research.

The European Union’s AI Act offers an important framework, though it is not a complete answer to every emotional-design problem. It establishes a risk-based approach to AI, prohibits certain harmful manipulative and exploitative practices, and requires transparency in some circumstances so users know they are interacting with AI. [S1] [S2] The European Commission states that systems posing serious risks to health, safety or fundamental rights can be classified as high risk and subject to obligations involving risk management, dataset quality, logging, documentation, human oversight, robustness, cybersecurity and accuracy. [S2]

Those rules matter because opaque systems can affect real opportunities and rights, including employment, education, credit and access to services. [S2] But emotional influence requires sustained attention too. A chatbot does not need to make a formal hiring decision to shape a life. It may influence whether someone seeks help, trusts a false belief, withdraws from a relationship, reveals private information or spends another hour in a loop of reassurance.

The strongest finding, then, is not a single statistic. It is a convergence of evidence and design reality: AI is moving into the social and emotional layer of everyday life while direct evidence about its developmental and mental-health effects is still incomplete. [S8] [S9] That combination—rapid adoption, psychologically persuasive interaction and unfinished evidence—demands precaution without hysteria.

We should not ask whether AI is good or bad for people in the abstract. We should ask a better set of questions.

What role is this system being invited to play? Who is most likely to rely on it? What does it optimise for? What happens when it is wrong? What personal information does it absorb? Does it increase a person’s capacity to reconnect with the world, or does it quietly make the product more central to their emotional life?

Ordinary people deserve technology that helps them think, learn and communicate without exploiting the very human need to be heard.

Sources - https://www.apa.org/topics/artificial-intelligence-machine-learning/health-advisory-ai-adolescent-well-being - https://link.springer.com/article/10.1186/s12982-026-02361-8 - https://www.kidsmentalhealthfoundation.org/mental-health-resources/technology-and-social-media/ai-and-kids-mental-health - https://www.hopkinsmedicine.org/news/articles/2025/09/kids-turning-to-chatbot-therapy - https://www.apa.org/topics/social-media-internet/health-advisory-adolescent-social-media-use - https://www.hhs.gov/surgeongeneral/reports-and-publications/youth-mental-health/social-media/index.html - https://pmc.ncbi.nlm.nih.gov/articles/PMC11450345/ - https://artificialintelligenceact.eu/high-level-summary/ - https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai