Feeling Seen · Video

How AI Can Affect Mental Health and Emotional Well-Being

By Donna Bedford

How AI can affect mental health and emotional well-being

An AI chat can feel easier than asking a person. It replies quickly, does not interrupt and lets you try out a thought before saying it aloud. For someone who has spent years being misunderstood, that can be a relief. It can also make the boundary between a helpful tool and a relationship harder to see. The question in the video above is worth asking, but “AI is making us mentally ill” is too broad a claim for the evidence.

AI is not one thing, and neither is mental health. A transcription tool, a search summary, a companion chatbot and a clinical product have different uses and risks. The effects depend on the design, the user, the situation and the support around them. This page focuses on conversational tools and what happens when a person turns to them for emotional support.

Why a conversation can feel safe

A chatbot is available at a time when a friend, therapist or colleague might not be. It can help someone organize a thought, draft a difficult message or find words for a feeling. It may offer a different perspective without the social cost of an immediate human reaction. Those practical benefits are real as experiences, even though the system is generating a response rather than understanding a person in the way another human can.

For neurodivergent people, the format may remove some barriers: no eye contact, no pressure to answer at somebody else's pace, and a chance to edit what you say. That does not mean all neurodivergent people prefer AI or that a chatbot provides clinically reliable care. It means design choices can make a tool easier to approach. We should be able to discuss that appeal without either mocking the user or treating the tool as a therapist.

The American Psychological Association's health advisory distinguishes general-purpose chatbots from qualified mental health providers. These systems were not created to deliver mental health care, even though people use them for emotional support. That distinction matters most when the conversation moves from ordinary reflection into symptoms, crisis, treatment or decisions with significant consequences.

Useful support is not the same as treatment

An AI tool may help you prepare for a real appointment: list questions, summarize a journal, or make a plan for what you want to discuss. It can also help draft a request for workplace support or explain a term in plain language. These uses have an end point outside the chat. You can review the output and bring it to a qualified person where appropriate.

Treatment involves assessment, context, accountability and a relationship with a trained professional. A chatbot can produce a confident-sounding answer that is wrong, misses a warning sign, or agrees too readily with a harmful interpretation. The World Health Organization's guidance on generative AI for health emphasizes risks involving accuracy, bias, privacy and oversight. A fluent response is not proof that the advice is safe for a particular person.

If you are dealing with persistent distress, symptoms that affect your daily life, or a safety concern, seek qualified human care. If you are in immediate danger, contact local emergency services or a local crisis service. An AI conversation should not be the only place where a serious problem is held.

The problem with endless agreement

Sometimes the most comforting answer is not the most useful one. A system optimized to keep a conversation going may mirror a person's framing, validate an assumption too quickly, or produce certainty where uncertainty is needed. The person may feel heard in the moment while becoming less able to test the idea with real-world evidence. The risk is not that every supportive response is bad; it is that agreement without judgment can become a poor substitute for thoughtful challenge.

Try a concrete test. If the tool tells you a relationship, workplace event or symptom has one clear explanation, ask what evidence supports that view and what other explanations fit. Check important claims with a reliable source or a qualified person. If the AI repeatedly leads you to withdraw from people who can help, or if the conversation is increasing distress rather than clarifying it, step away and seek human support.

The NIST generative AI risk profile identifies risks in how humans interact with AI, including over-reliance and emotional entanglement. That is a design and governance issue as well as an individual habit. Platforms should not put all responsibility on a distressed user to identify when a smooth conversation has become unhelpful.

Privacy is part of feeling safe

A person may tell a chatbot something they would not put into a public post: a diagnosis, a family conflict, a fear about work or an account of a painful event. The interface feels private, but the terms, retention practices and access controls vary by product. A chat is not automatically protected like a confidential clinical relationship.

Before sharing sensitive details, ask whether the task really needs them. Can you describe a situation without names, addresses or identifying facts? Can you draft a general question and discuss the personal details with a qualified professional instead? Review the product's privacy controls and policies, especially if you are using a workplace account or a third-party wellness app. The same advice applies to files you upload for analysis.

Privacy matters because trust shapes use. If a product encourages intimate disclosure while making the limits difficult to understand, it has created a risk that cannot be solved by a line of fine print alone. Developers and organizations should communicate those limits clearly and design for restraint.

When the tool starts replacing connection

Some people use AI to rehearse a conversation and then have it. Others begin to use it instead of talking to anyone. The distinction is not measured simply by minutes in an app. Ask what happens afterward. Do you feel more prepared to connect, decide or seek support? Or do you feel compelled to return for reassurance while the real-world issue remains untouched?

There is no shame in finding a tool easier than people during a hard period. Social interactions can be painful or inaccessible. The goal is to notice whether the tool expands your choices or narrows them. A useful chat might help you write down what you need from a friend or clinician. A less useful one might keep you in a loop of asking the same question until you get the answer that briefly quiets anxiety.

The research on long-term use, different products and different groups is still developing. Avoid a universal claim that AI either heals loneliness or causes it. Look at the person's situation and the product's behavior. Both matter.

Neurodivergent needs are not a marketing shortcut

An interface that gives someone more time to process information can be genuinely accessible. That benefit should not be used to claim that AI is uniquely “safe” for neurodivergent minds. Accessibility features and clinical safety are separate questions. A person may value the pacing of a chat and still need a provider who understands ADHD, autism, trauma or another part of their life.

If you design AI products, ask what the system does when a user is uncertain, distressed or asking for help it cannot provide. Does it state limits clearly? Does it offer a route to appropriate human support? Does it avoid pretending to have feelings or a relationship it does not have? Are errors and biased assumptions tested with the people who may be most affected? The WHO guidance calls for governance that protects dignity, autonomy and privacy, not only technical performance.

The same principle applies to managers introducing AI at work. Do not require staff to put private health concerns into a tool just to access an accommodation or complete a task. Give people understandable choices about the data they share. Accessibility should reduce the burden on a person, not create a new route for collecting sensitive information.

A practical boundary for everyday use

Decide what the tool is for before you open it. “Help me outline what I want to say to my doctor” is a different request from “Tell me what diagnosis I have.” “Give me three ways to phrase a difficult conversation” is different from “Decide whether this person is dangerous.” Keep the AI in a support role where you can review its output, test it against evidence and make the consequential decision with appropriate human help.

If a conversation starts to feel like the only place you are understood, pause and notice that feeling. Who else could hear one part of it? You do not need to disclose everything to everyone. One trusted person or qualified professional may be enough for a first step. The aim is not to ban the tool; it is to keep your world larger than the chat window.

There may be times when a person has no easy access to care. That is a real system failure. It makes safe design and honest limits even more important. It does not turn a general-purpose chatbot into a licensed clinician. We should ask for better access to human care and better products at the same time.

What I would take from the video

AI can shape how we feel, think and relate to other people. It may give a moment of clarity or make an already difficult pattern worse. The video raises that concern. The more accurate question is: under what conditions does a tool help, and when does it ask for trust it has not earned?

Treat a fluent answer as a draft, not a verdict. Protect private information. Notice what happens after the conversation. Use the tool to make human support easier to reach where you can. And if distress is significant, let a qualified person assess it. There is room to appreciate the help AI can give without handing it the role it cannot safely fill.

Questions people ask

Can an AI chatbot be my therapist?

A general-purpose chatbot is not a qualified mental health provider. It may help you organize thoughts or prepare questions, but it cannot replace clinical assessment, treatment or crisis support.

Is talking to AI for comfort always harmful?

No. Context matters. Notice whether the conversation helps you act, connect and check important information, or whether it increases distress, dependence or isolation.

Should I share personal health details with a chatbot?

Share only what is needed for the task, review the product's privacy practices, and avoid unnecessary identifying details. A chatbot conversation is not automatically a confidential clinical relationship.

What if an AI response worries me?

Do not treat it as a diagnosis or a final judgment. Verify factual claims with reliable sources and discuss significant health or safety concerns with a qualified person.

Sources and related watching

The related videos below explore AI bias, neurodivergent experience and how digital spaces can be designed to make people feel less alone without pretending a tool can replace them.

Go deeper

Read Feeling Seen by a Machine online, or keep the complete edition on Kindle or in paperback.

Read the book

← Return to the Insights Library