The stronger the mental health problems, the more people consulted generative AI: what a survey of 6,852 people shows

Contents
  1. What the study examined
  2. The study did not show that generative AI improved mental health
  3. Why people consult AI is the next research question
  4. Treating frequent use as an entry point to support
  5. References

Have you ever talked to generative AI about your worries? A study using a Japanese internet survey found that the stronger people's mental health problems, the more likely they were to use generative AI regularly to talk about worries and feelings1. The study is also summarized in Tohoku University's official announcement (in Japanese), "Stronger mental health problems are associated with regularly consulting generative AI about worries"2.

The survey covered 6,852 people without a regular doctor. The proportion who used generative AI a few times a month or more to talk about worries and feelings was 4.6% in the group with low mental health need, 9.4% in the elevated group, 16.3% in the high group, and 20.7% in the group with both serious psychological distress and a self-reported mental disorder. In the group where both indicators overlapped, about one in five used it regularly.

Level of mental health need and regular consultation with generative AI
4.6% 9.4% 16.3% 20.7% Low Elevated High Both indicators

JACSIS 2024–2025. Regular use means using it a few times a month or more to talk about worries and feelings. Proportions are survey-weighted estimates

What the study examined

A large proportion of people with mental health problems do not receive treatment, both worldwide and in Japan. People without a regular doctor have few contacts with medical institutions, so their mental health problems are hard to see from the health care side. However, few studies had followed people without a regular doctor to examine how their mental health relates to their later use of generative AI2.

The study included people who answered both the 2024 and 2025 waves of JACSIS, a nationwide internet survey in Japan, and who said in 2024 that they had no doctor they usually consult. Responses were weighted so that the sex and age composition of respondents approached that of the national population. Mental health need in 2024 was divided into four groups based on the K6, a six-item scale of psychological distress, and on self-reported current depression or other mental disorders. In 2025, respondents were asked whether they had used generative AI a few times a month or more to talk about worries and feelings1.

Four groups by mental health need in 2024
Low (3,804)
K6 0–4, no mental disorder
Elevated (1,887)
K6 5–12, no mental disorder
High (954)
Either serious psychological distress (K6 13 or more) or a mental disorder
Both indicators (207)
K6 13 or more and a mental disorder

The K6 is a six-item questionnaire scoring depressive mood and anxiety over the past 30 days from 0 to 24. A score of 13 or more indicates serious psychological distress

After statistically accounting for age, sex, income, education, employment, loneliness, consultation resources, past-year use of mental health services and other factors, the association remained: the stronger the mental health need, the more common regular use. An additional analysis of 6,424 people with a regular doctor showed an association in the same direction.

Ratio of the proportion of regular use, adjusted for age, income and other factors
Elevated
1.67 (95% CI 1.25–2.24)
High
2.61 (1.89–3.61)
Both indicators
2.71 (1.75–4.20)

Relative to the low-need group. 6,852 people without a regular doctor

The study did not show that generative AI improved mental health

What the study showed is an association between the strength of mental health problems and regularly consulting generative AI about worries. It cannot say that mental health problems caused AI use. Nor does it show that using generative AI improved symptoms, or that it was safe as a confidant.

The study did not confirm generative AI use before 2025. It therefore cannot tell when people started using it or how use and mental health influenced each other. The survey was self-reported, and it did not assess what people told generative AI, what responses they received, or whether their condition improved or worsened afterwards.

What the study can and cannot yet tell us
What it can tell us
Groups with stronger mental health problems more often consulted generative AI regularly
What it cannot yet tell us
Reasons for use, content of conversations, safety, effects on symptoms, and whether it can substitute for professional support

When thinking about safety, regular use of generative AI needs to be distinguished from emotional dependence. An earlier article, "When AI becomes a companion: which designs and which people make emotional dependence more likely," discussed how this differs from relying on AI as a tool, and which designs and user conditions make dependence more likely. Because this study did not measure dependence, regular use cannot simply be read as dependence.

The claims in lawsuits over companion AI that acts as a conversational partner, and the risks and benefits shown in research, are presented separately in "Lawsuits over Character.AI and research on the risks and benefits of companion AI." This survey did not target any specific service and did not directly test the safety issues discussed there.

Why people consult AI is the next research question

Generative AI responds whenever text is entered, at any time. Some people may first write to generative AI about worries that are hard to tell another person. At the same time, cost, access to services, and worries about stigma may make it difficult to consult people. The paper points out that consulting generative AI may indicate support needs that are hard to see from usual health care pathways, or digital help-seeking1.

However, this survey did not ask about reasons for use. We cannot assume that people use it "because it is easy," "because it is anonymous," or "because it is easier to talk to than a person." Under what circumstances people start using it, and what kinds of responses keep them using it, need to be studied through conversation content and users' experiences.

Treating frequent use as an entry point to support

It is worth noting that the main analysis covered people without a regular doctor. Regular consultation with generative AI may offer a clue to the support needs of people whom health care and counseling services do not know about.

In designing generative AI, it matters that people can move to human support when needed, rather than simply being kept in the conversation. If users wish, they should be able to choose among places to consult, such as health care, psychology, welfare, schools and workplaces. In situations that call for professional support, generative AI should not be left to carry the burden alone. The next task is to build such referral routes and to evaluate whether they are actually safe and helpful.

That generative AI is being used is different from generative AI being enough. The finding that people with stronger mental health problems more often consulted it regularly does not prove its effectiveness; it raises the question of how to respond to support needs that are hard to see.

References

  1. Koda M, Harada N, Matsui K, et al. Mental Health Needs and Emotional-Support Artificial Intelligence Use. JAMA Network Open. 2026;9(11):e2644315. doi:10.1001/jamanetworkopen.2026.44315
  2. Tohoku University. Press release (in Japanese). 18 November 2026.
English translation of the Japanese article.