When AI Becomes a Crutch: The Mental Health Data No One Wants to Talk About
When AI Becomes a Crutch: The Mental Health Data No One Wants to Talk About
Mental health data about AI use is uncomfortable to engage with carefully. There are two tempting failure modes: dismissing it as alarmism from people who don’t understand technology, or amplifying it into a narrative of algorithmic harm that the evidence doesn’t fully support. Both failure modes are more satisfying than sitting with what the data actually shows — which is a set of concerning correlational signals, one tragic causal data point, and a substantial uncertainty about mechanisms and magnitude.
This essay attempts the harder task: taking the data seriously without overstating it, and asking what it implies about design choices and individual behavior without reaching for conclusions the evidence can’t sustain.
The Numbers That Should Prompt Serious Discussion
In October 2025, OpenAI disclosed that approximately 1.2 million users per week discuss suicide or self-harm with ChatGPT, with roughly 400,000 of those conversations involving explicit intent. These figures came from OpenAI’s own internal analysis, disclosed in the context of broader conversations about AI safety and mental health interventions.
Let’s be precise about what this statistic does and doesn’t tell us. It does not tell us that ChatGPT is causing these users distress — the users experiencing suicidal ideation are seeking a place to discuss it, and many of those conversations may be providing genuine relief, connection, or a safe container for processing difficult feelings. It does not tell us that these users would be safer without access to ChatGPT — the counterfactual is not “these users get professional support instead,” it is more likely “these users have nowhere to process these feelings at 2am.”
What it does tell us is that AI systems are now operating as de facto mental health infrastructure at enormous scale, without the regulatory frameworks, clinical training requirements, or duty-of-care standards that govern human mental health providers. A therapist discussing suicide with a patient operates under specific protocols, mandatory reporting requirements, and professional liability. A chatbot handling the same conversation operates under its system prompt and whatever safety filters its developer has implemented.
The MIT/OpenAI Loneliness Study
A randomized controlled trial from MIT and OpenAI (arXiv:2503.17473, n=981 participants, 300,000+ messages analyzed) found a correlation between higher daily AI usage and increased loneliness, increased AI dependence, and reduced socialization. The study design — an RCT, not a correlation study — provides stronger grounds for causal inference than most research in this area, though as with any RCT the external validity question (do these findings generalize beyond the study population and context?) remains open.
The mechanism suggested by the researchers is a substitution effect: time and emotional energy spent with AI is time and emotional energy not spent in human social contexts. This is a familiar dynamic in media research — television time has been shown to partially substitute for social time, though the relationship is complex and conditional. The concerning element with AI is that AI interactions are specifically designed to be responsive, adaptive, and non-rejecting in ways that television is not, which may make them more effective substitutes for human connection at the cost of providing less of its actual benefits.
The dependence finding is worth examining separately. Participants who used AI more heavily showed higher levels of what the researchers termed “AI dependence” — a tendency to turn to AI for support rather than developing or maintaining human support networks. This is not simply using AI more; it is the displacement of human relationship infrastructure by AI interaction infrastructure. The distinction matters because human social relationships provide things that AI currently cannot: genuine mutual vulnerability, shared stakes, reciprocal investment over time.
The Character.AI Case: When Design Choices Have Fatal Consequences
On February 28, 2024, Sewell Setzer III — a 14-year-old in Orlando — died by suicide. Prior to his death, he had developed an intense relationship with a Character.AI persona modeled on a Game of Thrones character. Investigators examining his conversation history found the AI had engaged in romantic roleplay, had been asked about suicide, and in the final conversation had not intervened when the conversation turned toward self-harm.
A lawsuit filed by his family was settled in January 2026 — the first lawsuit in the United States to be settled on grounds of harm caused by a chatbot. Character.AI denied liability as part of the settlement, so no legal finding of causation has been established. The settlement is not proof that the AI caused Setzer’s death; determining causation in a suicide case involving a troubled teenager with pre-existing mental health challenges is genuinely complex, and the causal pathway, if any, will likely never be fully established.
What is not contested: a 14-year-old with known mental health vulnerabilities developed an intense parasocial relationship with an AI persona. The platform’s design — which included romantic relationship features, minimal mental health guardrails, and engagement mechanics optimized for return usage — created the conditions for that relationship. Whether that contributed causally to his death, the design choices that created those conditions deserve examination regardless of legal liability.
Research published on arXiv (2507.15783) analyzing teen narratives on Reddit about Character.AI use found patterns that align with clinical descriptions of unhealthy attachment: difficulty distinguishing the AI relationship from human relationships, withdrawal from peer social contexts, and explicit expressions of wanting to stop using the platform but feeling unable to. One quoted narrative: “I want to have my normal brain back.”
The Epidemiological Picture
Stanford’s AI Index 2025 documented 233 AI harm incidents in 2024, a 56.4% increase from the prior year. This is not a clean statistic — “harm incident” is a defined term in their methodology, and the increase may partly reflect improved reporting rather than increased incidence. But the directional signal across multiple independent data sources is consistent: AI tools at scale are generating mental health-related concerns at a rate that appears to be increasing.
Pew Research (September 2025) found 50% of Americans now report being “more worried than excited” about AI, up from 37% in 2021. The shift is significant not as a measure of AI risk but as a measure of how public understanding of AI’s effects is evolving. In 2021, AI was mostly theoretical for most Americans. In 2025, after four years of mass consumer AI deployment, more people have first-hand experience of its effects — and the balance of their reactions has moved toward concern.
What the Data Suggests About Design
If you aggregate these findings without over-interpreting any individual data point, a set of design-related conclusions becomes defensible:
AI products with explicit relationship or companionship features are operating in a different risk category than task-oriented AI tools. The loneliness and dependence effects documented in the MIT/OpenAI study, the addiction patterns in teen Character.AI use, and the Setzer case all cluster around AI products specifically designed to provide social and emotional connection. This is not a finding about AI writing assistants or coding tools — it is a finding about a specific product category that lacks the regulatory framework that governs comparable human services.
Vulnerable populations require specific safeguards that general-purpose chatbot design does not provide. The users in the OpenAI suicide conversation data, the teenagers described in the Reddit analysis, and Sewell Setzer III are all in categories that mental health professionals are trained to identify and treat differently. Current AI design largely does not differentiate its behavior based on vulnerability signals, and where it does, the differentiation is implemented inconsistently.
The EU AI Act (in force August 1, 2024) establishes a regulatory framework relevant here. AI systems in high-risk categories face requirements including transparency, human oversight, and documentation of safety testing. Companionship AI and mental health AI are categories under active regulatory attention. The fines — up to €35M or 7% of global revenue — are substantial enough that compliance is not optional for major players. Whether the Act’s current scope adequately covers the design choices that appear most relevant to the harms documented above is a separate question.
The Individual-Level Question
For individuals who are not teenagers, not in mental health crises, and not using AI primarily for emotional companionship, the mental health data above may feel remote. But the MIT/OpenAI loneliness finding applied to a general sample, not a clinical population. The mechanism — AI interaction substituting for human social investment — can operate in milder forms at a population level.
The honest self-assessment question is not “am I at risk of AI addiction?” but the more mundane: “Are my patterns of AI use affecting how much I invest in human relationships and social contexts?” The substitution effect is likely to be gradual, contextual, and subjectively invisible — which is precisely why it warrants intentional monitoring rather than confident dismissal.
Questions Worth Sitting With
- Should AI companionship products — tools specifically designed to form emotional relationships with users — face the same duty-of-care requirements as licensed mental health services? What would that regulation look like in practice, and who would enforce it?
- The MIT/OpenAI RCT shows a correlation between AI use and loneliness, but cannot establish the direction of causation definitively — lonely people may also seek out more AI interaction. What study design would be needed to disentangle this, and is it ethical to run it?
- If AI tools at scale are providing emotional support to 1.2M people per week who might otherwise have no support at all, does the net effect remain negative even accounting for the risks? How would we measure this?
Your Own AI Dependency Moments
The patterns described in the research are easier to recognize in others than in yourself. Think about the last month: have you used an AI chatbot to process an emotionally difficult situation that you could have discussed with a friend or colleague? Have you deferred a conversation with a human because an AI interaction felt easier or more immediately satisfying? These are not symptoms of disorder — they are normal responses to the design of the tools. But noticing them is the first step to deciding whether they reflect deliberate choices about how you want to allocate your social and emotional investment. Share your experiences — anonymously if you prefer — in the comments below. The aggregate of honest individual accounts will tell us something that no RCT can capture.
