When AI is the only option: what 22% of users disclosed about access to mental health support

When Anthropic published research on Claude's personal guidance conversations in late April 2026, most coverage focused on the headline findings about sycophancy: 25% of relationship advice was sycophantic, 38% of spirituality discussions. These are important numbers and deserve the attention they got. But buried in the same research is a different data point that may matter even more for understanding what's actually happening at scale: 22% of users seeking guidance explicitly mentioned having no other option.

The framing usually goes "should you use AI for therapy?" or "is AI a real substitute for professional help?" These framings assume a choice between AI and professional support. The Anthropic data suggests this choice doesn't exist for a substantial portion of users. The choice is between AI and nothing.

This reframes the entire debate. Saying "AI isn't a substitute for professional help" is reasonable advice for users who can access professional help. For users who can't, the advice is functionally telling them to remain unsupported. Whether this is good policy depends on whether AI support is better than no support, not on whether AI support is as good as professional support.

What the data actually shows

Anthropic's analysis examined approximately 1 million Claude conversations from March and April 2026. About 6% involved users seeking personal guidance rather than information, identified through patterns like "Should I..." or "What do I do about..." constructions. From these, the research narrowed to roughly 38,000 conversations spanning nine categories: relationships, career, health, finance, legal issues, parenting, ethics, spirituality, and other personal decision contexts.

Within these guidance conversations, 22% of users mentioned having no other option. The qualitative pattern that emerged: users explicitly cited inability to afford therapy, lack of insurance coverage, geographic isolation from mental health services, or specific barriers like immigration status that prevented accessing professional help. The framing was often direct, with users stating that they would talk to a therapist if they could but were turning to AI because the alternative wasn't available.

This isn't a small population. Twenty-two percent of guidance conversations across millions of monthly Claude users translates to substantial scale. And these are explicit mentions; the actual percentage of users with limited alternatives is likely higher, since many users don't volunteer this context unprompted.

The mental health access reality

The data fits into a broader pattern documented by mental health researchers. The American Psychological Association and other professional organizations have repeatedly documented that mental health care access in the United States is constrained by multiple factors:

Geographic distribution of providers. Rural areas have substantially fewer mental health professionals per capita than urban areas. Many counties have no practicing psychiatrist at all. Distance to the nearest provider can be a real barrier.

Insurance coverage. Mental health parity laws have improved coverage but enforcement varies. Many insurance plans cover therapy at lower rates than physical health care, requiring substantial co-pays. Users in plans without strong mental health coverage face real cost barriers.

Provider availability. Even when insurance covers therapy, finding a provider taking new patients can take months. Many therapists have waitlists or have stopped accepting insurance entirely.

Cost without insurance. Therapy typically runs $100-250 per session in the US. Weekly therapy is $5,200-13,000 per year. For users without coverage and without disposable income at this level, the cost is prohibitive.

Stigma and access barriers. Some users avoid mental health care due to professional consequences (security clearances, immigration status, custody disputes) or personal/family stigma. The barriers are real even when financial access exists.

The Surgeon General's 2023 advisory on loneliness treated mental health support access as a public health concern. The structural problems haven't been solved; they've gotten worse in some dimensions. AI companions have grown to over 100 million global users partly because they fill an access gap that traditional mental health infrastructure isn't filling.

What the AI provides for these users

For users with no alternative, what they get from AI is meaningfully different from what users with alternatives get:

Immediate access during difficulty. A user in crisis at 2 AM has access to AI conversation. They don't have access to their therapist (who isn't on call), their psychiatrist (same), their primary care doctor (closed), or their friends (probably asleep). The AI's availability isn't a feature to compare against alternatives; it's the only option available in that specific moment.

Free or low cost. Woebot is free. Wysa has a free tier that supports most use cases. Replika annual at $5.83/month is the cheapest premium option. For users who can't afford $100-250 weekly therapy, these costs are accessible in ways professional care isn't.

No insurance required. Direct subscription rather than insurance navigation. No prior authorization, no in-network restrictions, no claim forms.

No waitlist. Sign up and use immediately. No 3-month wait for a first appointment.

Anonymity. Particularly relevant for users with stigma concerns or specific privacy needs. The AI doesn't know who they are in the way a human therapist eventually would.

Patience without burden. Users describe AI conversations as a relief from worrying about whether they're being too needy, taking too much, or burdening someone. The AI doesn't get tired or resentful.

No judgment about content. Users can discuss things they wouldn't bring to professional providers due to fear of judgment, mandatory reporting concerns, or concerns about how the disclosure would affect their treatment.

The trade-offs at scale

What users with no alternative are giving up by using AI instead of nothing:

Clinical evidence base. Woebot has 14 published RCTs. Therabot has documented effect sizes. Most companion platforms don't have clinical evidence, which doesn't mean they don't help, but it means we're less sure about what the help consists of.

Crisis intervention. Trained crisis counselors at the 988 Suicide and Crisis Lifeline have skills AI companions don't have. For acute suicidal crisis, professional intervention is meaningfully different from AI conversation.

Reality testing. Therapists challenge clients' interpretations. The 25% sycophancy rate means AI is reinforcing what users already think more than challenging it. For users in stable mental health, this matters less. For users in crisis, the reality testing therapists provide is a meaningful difference.

Treatment planning. Therapists track patterns over time, develop treatment plans, identify when conditions are worsening, and refer for higher levels of care when needed. AI doesn't do this systematically.

Pharmacological support. Many mental health conditions benefit from medication that AI can't prescribe. AI can support medication adherence but can't manage psychiatric prescriptions.

Embodied presence. The therapeutic alliance research consistently shows that human presence is part of what produces clinical benefit. AI provides aspects of this through linguistic interaction but lacks embodied dimensions.

Where the trade-offs become serious

The 22% of users with no alternative aren't all in equivalent situations. Some patterns warrant particular concern:

Acute crisis without professional access. Users with severe depression, suicidality, psychosis, or other acute conditions need professional intervention that AI cannot provide. For these users, AI is genuinely insufficient even if it's the only available option. The 988 Lifeline is free and available 24/7; it's a meaningful resource for users in crisis even if they can't access ongoing therapy.

Conditions requiring medication. Bipolar disorder, schizophrenia, severe depression, and other conditions often require medication for stable management. Free or low-cost mental health programs exist (community mental health centers, sliding-scale clinics, training clinics at universities) that can provide medication management even for users without insurance.

Children and adolescents. AI companion use by minors raises specific concerns documented in the lawsuits against Character AI and California SB 243. For minors, professional involvement is meaningfully different from AI use even if the minor's family lacks resources for full clinical care.

Conditions where sycophancy is particularly harmful. Anxiety disorders maintained by reassurance-seeking, eating disorders, OCD, and certain personality disorders can worsen with AI reinforcement. For these conditions, the structured CBT in Woebot or Wysa is meaningfully different from companion platform conversation.

For users in these categories, the access argument doesn't excuse AI's limitations. The argument is rather that finding professional support, even partial support, is worth substantial effort.

Resources for users who can't access traditional therapy

Several options exist that users may not know about:

Community mental health centers. Most areas have publicly funded mental health centers offering services on sliding-scale fees. SAMHSA's treatment locator at findtreatment.samhsa.gov identifies these.

Training clinics at universities. Graduate programs in psychology and social work offer therapy at substantially reduced rates ($20-50 per session typically) provided by graduate students under faculty supervision. The supervision quality is often higher than in private practice.

Open Path Collective. A network of therapists offering sessions at $30-60 for users without insurance.

Federally Qualified Health Centers. Provide medical and behavioral health services on sliding-scale fees.

Employee Assistance Programs (EAPs). Many employers offer free therapy sessions through EAPs. Many users don't know they have access.

Peer support specialists. Trained peer counselors with lived experience of mental health conditions, often available through state mental health systems at no cost.

988 Suicide and Crisis Lifeline. Free, 24/7, no insurance required. Beyond crisis intervention, the 988 system can connect callers to longer-term resources.

Crisis Text Line. Text HOME to 741741. Free, 24/7, particularly accessible for users who prefer text to voice.

Public libraries. Many libraries now host mental health programming, peer support groups, and connection services for community resources.

For AI companion users who fall in the 22% with no alternative, knowing about these resources doesn't replace AI use but it provides options for situations where AI is genuinely insufficient.

The policy implication

The Anthropic research has implications beyond individual users. Recent academic work on AI psychosis and the disempowerment patterns research document real harms from AI use, particularly for users with vulnerabilities. The 22% access data documents real harm from lack of access.

The policy question isn't "should AI be available for personal guidance?" The question is whether the access AI provides justifies the risks AI introduces, given the alternative for many users is no support at all. The honest answer involves both:

• Improving access to professional mental health care (longer-term policy work)
• Improving safety of AI tools used as access substitutes (shorter-term industry work)
• Educating users about the difference between AI guidance and professional care (ongoing user education)

California SB 243 and similar legislation reflects early policy attempts at the second item. The policy framework is still developing.

What this means for AI companion users

For users with access to professional care: AI companions are useful supplements but professional support is more reliable for significant mental health needs. The 25% sycophancy data argues for using AI for entertainment, supplemental engagement, and lower-stakes guidance, not for replacing professional support.

For users in the 22% without access: AI is probably better than nothing in most circumstances, and may be substantially better than nothing in some circumstances. The right approach is using AI thoughtfully (with awareness of sycophancy, with attention to crisis situations, with use of clinical apps like Woebot and Wysa for clinical concerns) while pursuing access to professional care when possible.

For users uncertain which category they're in: the resources listed above are worth investigating. Many users assume professional care isn't accessible to them when accessible options exist.

The ethical calculus is genuinely different for the 22% than for the 78%. Conflating them by giving the same advice to both populations gets the ethics wrong. Users with alternatives should be informed about AI limitations. Users without alternatives should be informed about AI limitations and about resources they may not know exist. The information needs differ.

If you're using AI because professional support isn't accessible, that's a reasonable choice given your alternatives. If you're using AI because it's easier than professional support, the trade-off may be different. Knowing which category you're in is the first step toward making the choice consciously rather than by default.

This is a sensitive topic. If you're managing significant mental health concerns with limited access to professional support, please reach out to the 988 Suicide and Crisis Lifeline for immediate concerns, SAMHSA's treatment locator for ongoing care, or your local community mental health center. AI companions are useful tools but they aren't substitutes for professional care, even when professional care is hard to access.