Why Are So Many Americans Turning to Bots for Therapy?

Last week, FDA’s Digital Health Advisory Committee heard testimony about A.I.-enabled digital mental health medical devices. It was a marathon session with comments from entrepreneurs, professional societies, academics, and many more. The topics ranged from vocal biomarkers to safeguards for autonomous vehicles. After 8 hours listening to every possible stakeholder, I was expecting a Replika or Character AI bot to weigh in. That did not happen.

What struck me was not only the breadth of topics but the thin evidence base. Unlike FDA meetings on drugs or medical devices, there was little hard data on either the benefits or the risks of using Gen-A.I. for mental health care. Speaker after speaker lamented the nation’s worsening access to treatment—and worried aloud about the growing number of Americans turning to general-purpose A.I. chatbots for support.

Their concern isn’t misplaced. A recent survey of chatbot users found that nearly half—48.7 percent—had sought mental health help from a bot, and more than a third said they found the bot more helpful than traditional therapy. If we were running a national restaurant chain and tens of millions of hungry people were eating elsewhere, we would not blame the customers. We would ask: What’s failing in our system?

But I want to pose a slightly different question: If people are already seeking therapy from chatbots that were never built for clinical use, is it possible to build a model that is designed for mental health care—one that is safer, more effective, and purpose-built to help people change? That is the premise behind Ash, a bespoke large language model (LLM) created by the company Slingshot AI. Unlike general-purpose systems such as ChatGPT, Claude, or Gemini, Ash was trained on large volumes of psychotherapy dialogues and engineered with guardrails specifically intended to detect and mitigate suicide, psychosis, and other high-risk clinical situations.

Yesterday, a team from NYU and Univ of Washington working with Slingshot AI released the first (non-peer reviewed) results based on 305 adult users. This is not a randomized clinical trial, there was no comparison treatment, and the results are short-term (10 weeks); but just as a description of real-world use, I find the results surprising. A few topline points:

• Improvements in depression and anxiety symptoms were comparable to treatment studies (roughly 75% report improvement on standard scales).

• Improvements were found in social support and social engagement (roughly 72% reported a decrease in loneliness on a standard scale and about half reported an objective increase in social support such as a new person or new social activity in their lives).

• Improvements were seen in personal goal attainment (roughly one-third successfully reached a personal goal).

• No adverse events were noted even though there were some high-risk conversations detected (the guardrails and escalation protocols worked essentially 100% of the time).

It’s easy to dismiss these results because they have not yet been subject to peer review, few users had high scores on mood and anxiety scales, and the company that developed Ash was involved. I may be biased (I am an advisor to Slingshot AI), so, consider my comments accordingly. Several insights seem worth taking seriously.

• First, a bespoke LLM can be built for relational bridges rather than relational substitutes. As the authors state, “an AI model that repeatedly reminds users of existing social resources, and helps builds skills to access them, can meaningfully shift patterns of isolation.”

• Second, safety might not be the barrier we think it is. Imagine an LLM trained to not only detect suicidal risk but treat suicidal ideation with an approach that mitigates adverse outcomes. In truth, we are not great at doing this in clinical care. I’m guessing that an LLM could be better at saving lives.

• Third, for people with clinical levels of mood or anxiety disorders, a bespoke LLM looks comparable to what we would expect with real world application of evidence-based traditional psychotherapy. These results are consistent with findings from a recent report using Therabot, another Gen-A.I.-based therapy bot tested against a waitlist control.

We are at the beginning of what may be a historic shift in how mental health care is delivered. It is plausible—indeed likely—that within a few years a substantial share of people seeking therapy will begin with a chatbot rather than a human clinician. Already, about one-third of users of systems like ChatGPT say they prefer the bot to traditional therapy. I find that both troubling and telling: troubling because these foundational models were never meant for therapeutic use and telling because the public is voting with its feet.

Therapists often insist that the human relationship—the empathy, humility, and embodied presence—is indispensable. Decades of research suggest they are right: the therapeutic alliance predicts outcomes more powerfully than the specific modality of treatment. And yet we now face an uncomfortable question: What if something that is essentially an autocomplete algorithm can, at least for some people, approximate that alliance?

Ted Chiang writing “Why A.I. Isn’t Going to Make Art” noted, “It’s by living our lives in interaction with others that we bring meaning into the world. That is something that an autocomplete algorithm can never do.” I want to believe that. Most of us do. Even ChatGPT would likely say the same.

But belief is one thing. Evidence is another. The next generation of mental-health A.I. systems may force us to confront whether our faith in the irreplaceability of the human therapist is based on principle—or on habit.