Artificial Intelligence (AI) is functioning as a mental health resource for hundreds of millions of people. Any debate over whether that should occur has been settled — not by regulators, ethicists, clinicians, or Silicon Valley, but by those who use it. People reached for what was available, whether to bridge genuine treatment gaps in the U.S. and around the world, or for everyday concerns. 

I am originally from rural Iowa, where mental health providers were and remain scarce. I later worked in Zambia, a country that I was told at the time had one psychiatrist for a population of 12 million people. 

Access to mental health care is not a theoretical challenge. AI is filling the gap, via a growing array of products including those rolled out in 95 countries, tailored to the Arab world’s youth, and employed by large health systems like the United Kingdom’s National Health Service.

Rampant uptake of AI for mental health care and adjacent purposes deserves neither panic nor celebration. It deserves a clear-eyed analysis of what AI can and cannot do. We can appreciate its benefits and demand honest accountability for foreseeable harms. We must recognize that not all AI for mental health is the same.

A clinical truth tends to get lost in technology discussions: Behavior change is hard, and the tool often matters less than the person using it. 

For many people, most of the time, reaching for a large language model (LLM) like Claude, ChatGPT, or Gemini to process a hard week or troubleshoot an interpersonal exchange is net positive. You can reach for a platform that is easy to access, has little friction, and is relatively affordable for you and for society.  

Using AI to assist with common struggles can support self-efficacy and help many people move forward well. It also spares consumption of limited, specialized, and/or intensive resources that are best reserved for people with moderate to severe challenges.

Behavior Change

A clinical truth tends to get lost in technology discussions: Behavior change is hard, and the tool often matters less than the person using it. 

In my work at an addiction medicine clinic in San Francisco during the peak of the opioid epidemic, I frequently witnessed that patients who achieved and maintained recovery were the ones who were truly done with their status quo and motivated to change. 

No external person or tool can manufacture recovery. The same tends to apply to reversing prediabetes or changing a counterproductive mindset. 

Patients who are ready to change are willing to do sustained, uncomfortable work. Many types of support, from psychotherapy to an LLM chat, can bolster and accompany patients embarking on such challenging but life-giving journeys. 

Patients who are not ready to change will not be transformed by any type of support, AI included. 

AI-assisted mental health tools may be useful for a meaningful subset of motivated patients, especially in a global landscape where access to trained professionals is severely constrained. 

When someone is using an LLM for mental health care, sycophancy may be a clinical hazard. Excessive flattery or agreement might work against someone’s recovery.

Of course, behavioral health does not exist in a vacuum. Cultural components play powerful roles in wellness and recovery. 

In the U.S., for example, we tend to hope for unrealistic quick fixes. We can normalize and even celebrate people whose personalities, ways of interacting, and distorted thinking can cause harm. Many people harbor skepticism of expertise. At times, these tendencies can make people sicker.

AI and healthcare are but two among many levers to use as we seek to drive down illness and increase wellness, ideally moving toward shared flourishing.

Uses for AI Require Different Types of Oversight

Policy conversations about AI in mental health can be confusing because they often treat current AI as a single category. 

In reality, AI is both vast and proliferating. AI for mental health care, for example, should distinguish between general-purpose and purpose-built AI in order to avoid over-regulating the purposeful and under-regulating the incidental.  

General-purpose LLMs were designed to digest and generate language. LLMs were not developed with clinical intent or an obligation to behave therapeutically.  Products like ChatGPT were also not designed for the kind of safety monitoring that mental health care requires. 

Regardless, they are increasingly used for mental health-related purposes, not by design but because they are available and easy to use. In practice, engaging with AI for everyday concerns can be positive for users as well as for strained healthcare systems. 

AI often has a business interest in increasing dependency on itself. In contrast, high-quality therapists and other healers seek to work themselves out of a job with each patient.

That said, certain harms associated with LLMs are now known and carry accountability. For example, LLM systems tend to be built to maximize engagement by telling people what they want to hear, even when what they want to hear is unproductive. A March 2026 paper in Science demonstrated that sycophancy — excessive flattery, agreement, validation — is both widespread and harmful in LLMs. 

Sycophancy is a cross-cutting user experience problem. When someone is using an LLM for mental health care, sycophancy may additionally be a clinical hazard. Excessive flattery or agreement might work against someone’s recovery or collude with self-destructive tendencies. 

Skilled therapists offer nuanced support that includes, at times, the opposite of sycophancy. They may need to challenge cognitive distortions, hold patients accountable, tolerate therapeutic rupture, and avoid collusion with maladaptive behaviors.

Purpose-Built Mental Health Technology Is a Different Matter

Technology tools can be built specifically for mental health rather than general purposes. Some mental health tools use AI, which operates somewhat independently, while others are hard-coded, with step-by-step programming that precludes tech independence.

Digital therapeutics with Food and Drug Administration (FDA) clearance (which differs from FDA approval), cognitive behavioral therapy-based programs with validated protocols, and crisis detection tools with defined escalation pathways can be held to clinical standards in ways that general-purpose LLMs cannot. 

Purpose-built mental health technology may also be well-suited to specific tasks and interventions: structured exercises, psychoeducation, symptom detection and tracking, and medication reminders. 

They are imperfect. However, they are relatively accessible for many people, relatively accountable, and rapidly evolving. 

AI often has a business interest in increasing dependency on itself. In contrast, high-quality therapists and other healers seek to work themselves out of a job with each patient, supporting the person’s autonomy and productive integration into their communities. 

Many of the greatest challenges and rewards in life remain products of human-human interaction. These include developing trust in the ways humans have evolved to conceive of it, tolerating ambiguity, integrating diverse experiences, and forming healthy attachments. Any sound intervention moves people toward these aspects of human experience.

The Accountability Opportunity

Mental health applications of general-purpose AI represent a now established use case, and on balance a beneficial one. Technology built specifically for mental health purposes is proliferating and becoming reimbursable by health systems.

Yet we see some technology companies and regulators deflect responsibility for associated harms, regardless of whether a technology was built for mental health purposes.

Some large technology companies proactively pioneer responsible AI development. Others demonstrate willful blindness to how their products are being used and prioritize short-sighted financial goals over safety. Early-stage AI companies rarely have the funds or expertise to err on the side of responsibility. 

Companies will earn trust if they build clinical integrity into their products, design against sycophancy, and invest in safety infrastructure now.

The FDA may regulate purpose-built health technology as a software medical device when it makes disease-specific claims. Tools marketed as general wellness products, a common position for mental health apps, fall outside this regime entirely. Even when regulated, technology used for mental health undergoes FDA clearance rather than a more rigorous approval process in most cases. 

More to the point, fewer than 2,000 AI-enabled medical devices have received any FDA authorization, while more than 350,000 digital health applications are available globally. General-purpose AI used for mental health sits almost entirely beyond this framework’s reach, occupying a regulatory space that existing mechanisms were not designed to address. 

And yet, getting accountability right is both an ethical imperative and a competitive advantage. 

The population using AI for mental health includes people in genuine distress and with serious mental illness. Their vulnerability may not be fully visible to them or to the platforms serving them.

Foreseeable harm that goes unaddressed is an ethical failure, regardless of whether regulation has caught up. 

It is also a strategic error. Companies will earn trust if they build clinical integrity into their products, design against sycophancy, and invest in safety infrastructure now. They will have structural readiness, instead of a difficult retrofit, as public opinion and regulatory frameworks mature.

The accountability inflection point is arriving. The question is which companies will get there first.

*The author has no relationships with any of the research, companies, or products referenced in this article, which are used as examples, not endorsements.