AI and Mental Health: Promise, Risk and Neurodivergent People
The promise and the risk can sit in the same conversation
An AI chat can feel easier to start than a difficult conversation with another person. There is no appointment to book. You can try a sentence, delete it, and try again. For someone who has spent years explaining how their mind works, that space can feel useful. It can help organize thoughts or put words around a question they want to ask a professional.
That usefulness is real as an experience, but it does not make the system a therapist or prove that its advice is sound. A chatbot can misunderstand, invent an explanation, agree too readily, or respond badly when the stakes rise. People also vary in what they need. There is no single neurodivergent reaction to AI.
This video asks us to hold both ideas at once: AI may lower a barrier to reflection, and it may introduce risks that are easy to overlook when the reply sounds warm. The practical question is what the tool is doing, what it cannot know, and where a real person needs to be involved.
Why the first step can feel easier
Starting is often the hardest part of a task. A blank page, an unfamiliar form, or a conversation you have rehearsed ten times can take more energy than the answer itself. A writing assistant can help turn a tangle of thoughts into a list, suggest questions for an appointment, or draft a message that the person can revise.
For someone who prefers text, the ability to read and edit before sending may be helpful. For someone who processes aloud, speech input or playback may be useful. The benefit comes from having options, not from assuming every person with ADHD, autism, dyslexia or another difference needs the same interface.
A bounded request is safer than asking the system to decide who you are. “Help me list the questions I want to ask my clinician” keeps the human decision in view. “Tell me exactly what diagnosis I have” asks a general-purpose model to do work it is not qualified to do. The distinction matters most when an answer could affect care or relationships.
A conversational tone can create false confidence
Chatbots are built to respond in language. A smooth reply may feel as if the system understands the full situation, but it only has the information supplied and the patterns it has learned. It does not observe your life, know the intention of another person, or conduct a clinical assessment.
It may generate a plausible reason for a partner's behaviour, an employer's decision, or your own feelings. Plausibility is not evidence. If you are distressed, a confident explanation can narrow your view at the moment you most need room to consider alternatives. Ask what the answer is based on, verify factual claims, and avoid using a chat as the sole basis for a significant choice.
The American Psychological Association's advisory on chatbots and wellness apps warns of unintended effects and potential harm. The World Health Organization's guidance discusses both opportunities and risks in health uses of large multimodal models. Neither source turns every use into a bad one; both point to the need for careful boundaries and evaluation.
Agreement is not always support
A person may arrive wanting reassurance. An AI response that echoes every assumption can feel kind in the moment. It can also leave an inaccurate interpretation unchallenged. A useful assistant should help someone examine a claim, not simply reward a dramatic conclusion because it sounds emotionally satisfying.
This matters when a conversation involves conflict. The tool hears one account, often written in a moment of hurt, and cannot independently verify the other person's motives. It can help draft a calm message or make a list of what you know and do not know. It should not become an authority on whether another person is malicious, whether a relationship must end, or whether a symptom has one explanation.
There is also a difference between reflecting and looping. If a chat keeps producing new interpretations of the same event, it may encourage more checking rather than resolution. Notice whether the interaction leaves you clearer and able to act, or more dependent on another reply. A break, a trusted person, or professional support may be the better next step.
Privacy is part of the decision
People may tell a chatbot things they would hesitate to put in a public post: medical history, a workplace problem, intimate relationships, or a child's details. Before sharing, inspect the service's privacy settings and terms. Consider whether the task can be done with less detail. A general request for help structuring a conversation rarely requires names, addresses or a full personal history.
Do not paste someone else's private information into a general-purpose tool just because you need help with wording. Replace identifying details where possible. If a service is used within healthcare or work, ask who can see the record, how it is retained, and whether a person will review the output.
Privacy is also about the person reading later. A generated summary can make a sensitive story sound more certain or more clinical than the speaker intended. Check the final wording before it is shared with a clinician, manager or family member. You own the meaning; the tool is only helping with the draft.
When human care matters
A general chatbot is not a substitute for assessment, treatment or crisis care. It may help prepare a list of symptoms, organize questions and explain an unfamiliar term, but a qualified professional can ask follow-up questions, review history and take responsibility for a care plan. The National Institute of Mental Health's review of mental health technology describes both the possibilities and the limits of apps, including uncertainty about effectiveness.
If you are worried about immediate safety, use local emergency services or an established crisis service rather than waiting for a chatbot response. For ongoing distress, speak with a licensed clinician or a trusted person who can help you reach care. A machine may produce compassionate wording, but it cannot provide the accountability and continuity of a real support relationship.
These boundaries do not make everyday use pointless. They help place a tool where it can actually help: preparation, organization and access to information that you can check. The higher the stakes, the less sensible it is to let an unverified answer carry the decision alone.
Neurodivergence is not one risk profile
Some people find a text conversation less overwhelming than a phone call. Some may find the open-ended exchange exhausting. A person who has been dismissed by others may value a reply that takes their question seriously, while also being more vulnerable to an answer that overstates its understanding. None of these possibilities defines an entire group.
Design should allow a person to choose their pace and format. Clear labels, a way to pause, access to the original information, and an obvious path to human help matter. W3C's cognitive accessibility guidance recommends clear steps and involving people with cognitive and learning disabilities in testing. A supportive tone does not compensate for an inaccessible or confusing process.
Avoid describing a user as a problem to be fixed. If a tool rewrites every direct message into a heavily social one, it may be teaching someone to mask rather than helping them communicate. A better prompt might be: “Keep my voice, make this request clearer, and show me what you changed.” The person decides which changes to keep.
A more grounded way to use the tool
Start with a task you can inspect. Ask for a list of questions, a neutral summary of your own notes, or several ways to phrase a request. Keep the original material beside the output. Check whether anything was added, omitted or given a certainty you did not intend.
For factual health information, look for reputable primary sources and bring questions to a clinician. Do not treat citations generated in a chat as real until you open and verify them. For emotional interpretation, ask the tool for multiple plausible readings and what information is missing, then step away rather than seeking endless reassurance.
Set a boundary before you begin. Decide what private details are unnecessary, what decision belongs to you, and when a person should be involved. If the chat is making you more anxious, confused or isolated, stop and choose another route. The measure of a useful tool is whether it helps you move through the world, not whether it keeps you in the conversation.
What builders should measure
A product team may count engagement, replies or time spent in the app. Those numbers do not show whether a person feels better, reaches appropriate care, or leaves with an accurate understanding. Measure errors, harmful reassurance, privacy concerns, accessibility barriers and the ability to reach a human. Test with the people the service is meant to help, including varied neurodivergent users, without treating a single participant as representative of everyone.
Make limitations visible in the interface. Do not present a general assistant as a clinician. Provide a usable escalation route and review failure reports. A comforting design needs to be matched by evidence, governance and accountability. Warm language on its own is not a safety feature.
One small test is to close the chat and see whether you can name your next action in your own words. If you can, the tool may have helped you think. If the next action is only to ask the same question again, the conversation may be keeping you in place. That is a cue to pause and seek another perspective.
Questions people ask
Can AI be useful for mental health without being therapy?
Yes. It may help organize thoughts, prepare questions or draft a message. Check its claims and use a qualified professional for assessment or treatment.
Why can a chatbot feel as if it understands me?
It responds in fluent, personalized language. That feeling can be meaningful, but the system only works from the information available to it and can misinterpret or invent details.
Should I tell a chatbot my diagnosis and private history?
Share only what is needed for the task after checking the service's privacy terms. You can often ask for help with wording without names or sensitive details.
When should I stop using it and speak to a person?
If the issue affects safety, care, a major relationship or a consequential decision, seek appropriate human support. Stop if the interaction increases distress, confusion or dependence.
Keep exploring
For the broader question, watch How AI Can Affect Mental Health and Emotional Well-Being. For how design can make a tool easier to use, see Designing Digital Spaces for Neurodivergent Individuals. For the assumptions a system may carry, read How AI Bias Can Exclude Neurodivergent People.
AI can help us find a first sentence. The important part is knowing when the next sentence needs to come from a person who can listen, question and remain responsible for what happens next.
Go deeper
Read Feeling Seen by a Machine online, or keep the complete edition on Kindle or in paperback.
Read the book