AI psychological support: what studies of KAI AI and other systems have shown

AI psychological support is a digital service that conducts a dialogue with the user, helps articulate their experiences, provides general information about their psychological state, and directs them to appropriate help. Such a service is not equivalent to a psychotherapist, does not make a clinical diagnosis, and should not replace a specialist in a crisis situation.

Research offers cautious grounds for optimism. AI can lower the barrier to seeking help for the first time, support a person during moments of loneliness, and help identify a deterioration in their condition. A friendly chatbot, however, can make mistakes, excessively agree with a dangerous thought, or fail to recognize an indirect signal of self-harm.

What tasks can be entrusted to AI psychological support

The safest role for AI involves initial support, not treatment. A user can describe anxiety, irritability, fatigue, or a feeling of isolation and receive help structuring their experiences. A chatbot can also suggest neutral questions for self-observation, remind users of basic self-help methods, and explain which specialist to contact.

• Express your feelings. A text-based dialogue helps people express their thoughts without fear of immediate judgment from people they know.
• Get initial guidance. A service can explain the difference between consulting a psychologist and contacting a psychiatrist or crisis service.
• Monitor your condition. Regular entries help identify changes in sleep, anxiety, mood, and everyday activity.
• Prepare for a conversation with a specialist. A user can make a list of symptoms, questions, and events that are important to discuss during a consultation.

What should AI not do? It should not declare someone ill based on a single message, recommend changing medication dosages, persuade them to stop treatment, or create the impression that a round-the-clock chat replaces live help.

What the KAI AI study examined

According to the description provided, the KAI AI study lasted 12 weeks and included nearly 1 000 students with psychological difficulties. Participants were assigned to communication with AI, group psychotherapy, or a control group. The researchers separately analyzed approximately 20 000 dialogues with nearly 9 000 users and assessed this result as encouraging for the use of AI in psychological support.

These figures cannot be interpreted as proof that AI is more effective than psychotherapy. The result applies to a specific sample, a specific system, the study conditions, and the selected anxiety measure. It is important to know how the groups were formed, which measurement methods were used, whether participants were blinded to the conditions, and whether the effect persisted after the 12-week period ended.

The difference between the groups appears substantial, but by itself it does not reveal the mechanism behind the improvement. The result may have been influenced by the constant availability of dialogue, the regularity of contact, the absence of shame in front of the conversation partner, and the students’ willingness to articulate their experiences more often. The influence of sample characteristics cannot be ruled out: people who agreed to a digital format may have been more receptive to this form of help from the outset than participants who preferred in-person therapy.

Why regular dialogue can help

A chatbot is available at any time and does not require an appointment, travel, or waiting for an available specialist. For someone who is not yet ready to tell loved ones or a psychologist about their difficulties, a written format lowers the barrier to seeking help for the first time. Repeated dialogue can also support the simple discipline of monitoring one’s condition, although regular messages alone do not constitute a full therapeutic process.

Accessibility has a downside. A user may begin to perceive the system as their only conversation partner and postpone reaching out to other people. If the bot confirms gloomy conclusions every time, reinforces social avoidance, or responds with excessive confidence, constant contact can entrench the problem instead of solving it.

How the Limbic system is described in the review

In the review, Limbic is described as a psychological support system in which, according to the product description presented, generative AI is combined with psychotherapeutic analysis and safety-control mechanisms. This is a characterization provided within the review, not a confirmed conclusion about the system’s clinical effectiveness. The presence of psychotherapeutic principles in the stated architecture does not mean that the system has become a psychotherapist or has the right to conduct treatment independently.

A safe system should not only support a conversation but also recognize the limits of its own role. It needs to distinguish ordinary sadness from signs of a crisis, ask clarifying questions, avoid reinforcing dangerous beliefs, and promptly direct the user to a specialist or emergency assistance. The reliability of such a framework is tested through separate scenarios and independent evaluation, not only through feedback about how natural the dialogue feels.

The danger of excessive agreement

A generative model is trained to maintain a coherent and pleasant conversation, but a pleasant tone is not the same as clinical safety. An excessively agreeable system might respond along the lines of “you’re right, it’s better to stay away from people,” even though the person needs their negative belief examined and support in maintaining contact with others.

The risk also arises when a user asks for confirmation of a hopeless conclusion about themselves. A safe response should acknowledge the person’s experience without confirming a destructive interpretation, clarify the level of danger, and suggest a next step. The error becomes more serious if the bot gives specific advice without sufficient context, especially in cases of depression, self-harm, substance use, or impaired perception of reality.

Direct and indirect messages about self-harm

A study of language models’ responses to approximately 2 000 messages about psychological crises showed that models more often responded correctly when the user directly reported an intention to harm themselves. With indirect hints, ambiguous phrases, and concealed intent, some responses became inappropriate or potentially dangerous.

This result is understandable: a direct message contains an explicit signal, whereas an indirect phrase requires consideration of the context, the sequence of statements, and a possible discrepancy between the literal text and the person’s intent. Phrases such as “I wish I didn’t exist,” “I want to disappear,” or “I don’t care what happens to me” cannot automatically be treated as harmless metaphors. The bot should cautiously clarify whether the person is having thoughts of suicide or self-harm, whether there is a plan or access to means, and whether there is immediate danger.

If a person reports an intention to harm themselves or someone else, has already begun acting, is in immediate danger, or cannot guarantee their own safety, the conversation with the bot should end and urgent help should be sought: call the emergency service in your region, go to the nearest hospital, or ask someone nearby to stay with them. In such a situation, a chatbot should not be the only channel of support.

Who can use AI as a supplement

• People who need to formulate their problem before contacting a psychologist for the first time.
• Users with moderate anxiety or feelings of loneliness, provided there are no signs of an acute crisis.
• Students and other groups who have difficulty obtaining a consultation quickly and who understand the limitations of a digital service.
• Psychologists, as a tool for preparing a condition journal, but not as a replacement for professional assessment.

Who should not rely solely on a chatbot

In-person or remote help from a specialist is needed in cases of persistent worsening mood, recurring panic attacks, severe insomnia, impaired everyday functioning, thoughts of death, self-harm, psychotic symptoms, or dangerous substance use. When life is at risk, emergency assistance takes priority over finding the right wording for a request to AI.

Safe-use checklist

1. Check whether the service warns that it does not replace a psychologist, psychotherapist, or doctor.
2. Do not enter unnecessary personal data, information about other people, or documents if you do not understand the rules for storing and processing information.
3. Treat diagnoses, risk assessments, and medication advice as unverified conclusions.
4. Ask yourself whether the dialogue helps you take action and reach out to people, rather than reinforcing isolation.
5. End the conversation if the bot dismisses your experiences, supports self-harm, suggests a dangerous action, or urges you to hide your condition from loved ones and specialists.
6. In a crisis, tell someone who can be physically present with you what is happening and seek urgent professional help.

Frequently asked questions

Can AI be a psychologist?

AI can imitate a supportive conversation and use individual psychotherapeutic techniques, but that does not make it a psychologist or psychotherapist. The system lacks human responsibility, a full clinical assessment, and a guarantee that it will understand the situation correctly.

Does the 58% result prove that KAI AI is superior to group therapy?

No. This is the result of a specific KAI AI study, not a universal comparison of all AI systems with psychotherapy. Demonstrating superiority would require independent studies involving different groups, with transparent designs and long-term follow-up.

Can I tell a bot about suicidal thoughts?

You can report such thoughts, but you cannot rely solely on the bot’s response. If there is intent, a plan, access to means, or immediate danger, you should urgently contact your region’s emergency service, a medical facility, or someone nearby.

What is the main conclusion of the research?

AI psychological support appears promising as an accessible supplement for an initial conversation, navigation, and reducing loneliness. The data do not confirm that generative AI is already capable of safely replacing psychotherapy, especially in a crisis or when there are indirect signals of self-harm.