People Are Already Using AI for Therapy. The Rules Haven't Caught Up

A message from one of our partners Mozibox specializes in physician hiring for health tech, from Medical Advisor to Chief Medical Officer. Built for teams that move fast: most see 3 to 5 highly qualified candidates within a week. That’s how Polaris Sleep hired its first Medical Director. Find your next physician leader →

When it comes to using AI in healthcare, chatbots for mental health are ground zero.

The space is also exploding in terms of demand, even as many companies are still determining their strategy for how to leverage AI safely and in a way that is clinically sound. The question is no longer if people will use AI for mental health support. They already are. The problem to be solved now is the role purpose-driven mental health chatbots will play and who decides their use cases.

Nearly two in five psychologists say they’ve had clients already using some form of an AI chatbot to self-diagnose. A RAND study found nearly one in five adolescents and young adults turned to AI chatbots for mental health advice in 2025. And last October, OpenAI disclosed that 0.15 percent, which equated at the time to about 1.2 million, of ChatGPT users, had conversations related to suicidal planning or intent within a week.

When Talkspace quietly launched its own behavioral health chatbot, the company’s CEO Jon Cohen told us on a recent appearance of Rounds that about 30,000 people started to use it with minimal marketing. The company has hired full-time therapists to monitor and supervise activity on Tee, in case an escalation is needed.

Today’s landscape “does reflect the biggest shift in mental health since the invention of more structured behavioral therapies like cognitive behavioral therapy (CBT),” Dr. Yusuf Sherwani, the co-founder and CEO of Pelago, a digital health company pairing AI with clinician-led care for substance use and mental health, tells Second Opinion.

“Consumers have already tasted something better and they are hungry for more behavioral support via generative AI tools,” he said.

There’s immense opportunity here to reach patients in a scalable, cost effective way. But also the potential for real harm if these chatbots are not designed with patient safety in mind.

Partnered by Pelago is the behavioral health company built on specialty substance use care. Pelago now delivers acuity-matched care across substance use, mental health and behavioral addictions, with fees at risk on clinical outcomes and ROI. Every member starts with Sona, a voice-first clinical AI, which works under licensed clinical supervision. This piece is sponsored by Pelago.

The rise of the omnibot

Experts tell us that the greatest risk we are facing at this time comes when there’s a mismatch between what people believe an AI tool can safely do, and what it was actually designed to do.

“We’re seeing the rise of the ‘omnibot,” says Jenna Glover, the chief clinical officer of Headspace. That refers to AI that consumers use for everything from asking for help booking a flight to deeply vulnerable personal mental health conversations.

That also brings up major questions, which we’ll address throughout this piece, about whether these chatbots should be classified as “wellness” applications or clinical. These definitions are deeply important because it impacts the way that the technology is regulated.

What we gathered from speaking with regulatory experts:

• Chatbots that are more focused on emotional resilience, and helping patients talk through daily stresses through active listening approaches likely wouldn’t be considered high risk.
• Assuming the interactions remain centered on helping people cope with daily life, that could be a positive outcome for the industry. Some of these individuals may find that the chatbot is all they need, versus requiring one-on-one clinical support. Or they might find it’s highly valuable to them between sessions.
• But it’s easy to see how this can become a slippery slope. If a patient starts exhibiting signs of suicidal ideation or a potential more serious mental health episode while talking to their omnibot, even wellness-focused chatbots will still need a mechanism to refer these individuals to care. And there’s a very grey line between conversations that remain about wellness or resilience, and those that begin to stray into discussions that feel more clinical inclusive of diagnosis and treatment.
• The AI bots are heavily in use and achieved enviable scale, but may not have the right in-built safeguards. For instance, the earlier models are known to be sycophantic and affirm what the individual wants to hear – versus sharing both sides of an issue. The mental health-specific chatbots, such as Talkspace’s Tee, are being trained to flag potential issues to human clinicians related to suicidal ideation, homocide risk, eating disorders, potential substance use, and so on.

These lines are still being drawn in real-time, and no one yet is claiming to have all the answers.