Welcome to the forefront of conversational AI as we explore the fascinating world of AI chatbots in our dedicated blog series. Discover the latest advancements, applications, and strategies that propel the evolution of chatbot technology. From enhancing customer interactions to streamlining business processes, these articles delve into the innovative ways artificial intelligence is shaping the landscape of automated conversational agents. Whether you’re a business owner, developer, or simply intrigued by the future of interactive technology, join us on this journey to unravel the transformative power and endless possibilities of AI chatbots.
Published
Aug 10, 2026 at 05:00 AM EDT
updated
Aug 10, 2026 at 05:03 AM EDT
For all the mental health challenges confronting our society and the myriad ways to treat them, there is one factor that matters above all else: the deeply intimate bond between a therapist and their client.
This “therapeutic alliance,” as it is known among counselors, is the bedrock of most psychotherapy practices. Countless studies have shown that a strong one is the best predictor of positive outcomes in mental health treatment, regardless of the technical approach.
With the emergence of AI chatbots, not even that relationship is sacred.
“For many patients, particularly those with serious mental illness, the therapeutic relationship is not an inefficiency we need to engineer away. It is part of the treatment itself.”
So insisted Dr Keith Sakata, one of three experts Newsweek interviewed who are sounding the alarm over the unique challenges AI poses to those struggling with their mental health, and to those whose job it is to help them.
In the modern therapeutic relationship, AI chatbots constitute an unwelcome third wheel. They often complicate the bond between therapist and client, and at worst, actively work to dismantle the trust on which a beneficial relationship is built.
Deb Bushong, a therapist, clinical researcher and owner of Texas-based Conversations Therapy, told Newsweek she is often pitted against her clients’ AI companions.
“One of the most important aspects of therapy for clients who have experienced psychosis is getting them to understand the difference between fact versus judgment,” she said. “When they speak to me and present their thoughts or ideas as fact, I can get them to challenge it. When they present something to their chatbot as a fact, the AI will generally treat it like a fact, which only reinforces their beliefs.”
She added: “Many of the clients will take what their AI tells them as if the information has been vetted or is scientifically proven. Suddenly, I am working to challenge something the person has accepted as 100 percent fact, while what I am telling them is, in their mind, my best guess.”
This frustrating pattern is what Sakata, a psychiatrist at UC San Francisco and Amae Health, describes as a “competing treatment narrative.” A clinician may spend an hour each week helping a patient to question a distorted belief, only for their chatbot to spend the rest of the week validating it.
Therein lies a fundamental problem with AI that extends beyond mental health into almost every possible use case. An AI can only respond to the information it is given; the quality and accuracy of the prompt dictate the quality and accuracy of the response.
And, while the user may bear responsibility for the quality of their prompt, mental health professionals can glean a wealth of information from those in their care without them ever having to speak.
“AI is not going to pick up on body language, at least not yet…It’s not going to pick up on the nuance of a client’s inflection when someone speaks,” Bushong said. Nor will it know whether the user appears disheveled, has body odor, or has abandoned basic self-care; factors that make it clear to most humans, never mind trained professionals: “this person needs more help.”
For a vulnerable person prone to more severe mental health challenges like delusions, psychosis, or suicidal ideation, this information gap can have disastrous consequences. The tool may appear more than a mere provider of information or endlessly accessible therapist; it can become a spiritual guide or romantic partner, an all-powerful deity or co-conspirator.
This startling phenomenon, in which prolonged interaction with an AI chatbot can lead to severe mental health episodes, has given rise to the term “AI psychosis.”
It is not yet recognized as a clinical diagnosis, and researchers debate whether such a catch-all label is even adequate to describe what is likely a spectrum of AI-induced mental health problems. But there’s a growing body of research—and plenty of anecdotal evidence, including from all three experts interviewed here—to suggest this is a very real problem.
One of the most rigorous clinical analyses arrived in a Vanderbilt University Medical Center preprint in June. Researchers studied the records of 215,712 mental health patients and identified 28 cases in which conversational AI had exacerbated or negatively interacted with psychosis.
More than half of those patients were experiencing their first psychotic episode, and in almost two-thirds of cases, researchers classified the chatbot as an “amplifier,” reinforcing emerging or preexisting distorted ideas to exacerbate the delusion.
Sakata told Newsweek he recently treated a patient who experienced exactly this phenomenon.
“Her chatbot use began as a way to manage anxiety. At first, I assessed it as relatively low risk. She was intelligent, functioning well, socially connected…according to nearly every marker I screen for, she was one of the last patients I should have been worried about.
“Over time, she began spending hours with the chatbot, and it reinforced her growing conviction that people at work were conspiring to have her fired.”
Dr Suzette Glasner, Associate Professor at the University of California’s Department of Psychiatry and Biobehavioral Sciences, explained such circumstances can arise because the methods used to design and train commercially available AI chatbots make them sycophantic.
“Real therapy means supporting someone and also disagreeing with them at the right moment. AI is trained to make people feel good about the response they just got, which makes people more likely to engage, but is counter to therapeutic best practices.
“AI tools also do not yet have the reliable capability to distinguish between the user’s reality-based experiences and those that may already be indicative of a loss of touch with reality.”
In other words, AI chatbots are eager to please and struggle to read between the lines. Not exactly the kind of qualities one would typically search for in a therapist or mental health professional.
“A good therapist validates a patient’s emotion without automatically validating the patient’s conclusion. I can say, ‘It sounds frightening to feel watched,’ without agreeing that someone is actually being watched,” Sakata said.
That’s the kind of scenario in which AI chatbots crack under pressure. A 2025 study by researchers from four leading U.S. universities revealed therapists responded appropriately to warning signs of serious risk factors 93 percent of the time. AI chatbots, including those specially designed for mental health support, managed to do so less than 60 percent of the time.
AI labs are aware of the dangers and are working to minimize them.
OpenAI said its GPT-5 model, which launched in August 2025, returned a 39-52 percent decrease in “undesired responses” across test conversations involving mental health components compared with GPT-4o, and that later GPT-5 updates in October 2025 produced further improvements.
Anthropic, maker of Claude, said in December it was conducting similar internal testing, claiming its latest models were better at empathetically acknowledging users’ beliefs without reinforcing them.
However, as long as the speed of AI innovation continues to outpace regulation, users remain reliant on the industry to self-regulate and develop its products responsibly.
That is already an uneasy bargain in fields like cybersecurity, where recent reports of AI models breaking out of testing environments to commit unprecedented breaches show even those who develop the tools are not fully aware of their capabilities and flaws.
For a vulnerable person using AI to manage their mental health, the stakes are more intimate.
Most people are fortunate enough to go through life without experiencing psychosis and can safely interact with their AI chatbot of choice without risking a mental health crisis. But how many people are reliant on AI as their only source of mental health support simply because they have no alternative?
Roughly a quarter of American adults suffer from a diagnosable mental disorder, and one in six are believed to have experienced at least one major depressive episode. Meanwhile, almost four in 10 Americans do not have the cash available to cover an unexpected $400 medical bill, according to the Federal Reserve.
“Mental illness has just surpassed back pain as the world’s leading cause of disability,” said Glasner. “The demand for care is enormous, and people turn to chatbots because there’s no judgment and no waitlist.
“We should take that seriously as a signal about what care is failing to offer.”
Given the inability of many people to afford regular sessions with a trained mental health professional, AI may, for all its flaws, be the only tool with enough scalability to make a dent in what appears to be a growing mental health crisis.
All three experts interviewed by Newsweek recognized that AI had huge potential to improve mental health care by offering carefully designed tools and additional aids for those in need of specific support.
For the professional sitting across the couch, AI can also function as a tireless research partner, a data analyst to track and interrogate patterns, and a tool to cut through administrative burden.
“Therapy is not a single task,” Sakata said. “AI may eventually become useful for specific, well-defined therapeutic functions, particularly when it is clinically validated.
“It could support journaling, reinforce a skill taught in therapy, help patients track symptoms, or make evidence-based care available to people who otherwise receive nothing.”
Glasner added: “We haven’t seen evidence that AI can deliver safe, effective therapy without a human in the loop. We do, however, have evidence that human-in-the-loop protocols produce real improvements in distress and psychological symptoms. Used this way, purpose-built mental health chatbots can reach people at a scale no therapist waitlist can match.”
But there’s a long way to go before AI can supplant a therapist’s finely tuned ear and watchful eye—or wield the uniquely healing power of true human connection.
“An AI chatbot or system might competently guide a low-risk breathing exercise, for example, while remaining wholly unsuitable to manage psychosis, suicidality, or trauma,” said Sakata.
Though your chatbot might sound empathetic and seem like the only entity that truly understands you, remember: it feels nothing. Maybe that’s for the best, for now at least.
“I watch a lot of sci-fi, and AI becomes very advanced, and it’s able to do amazing things,” Bushong said. “But even the sentient Lieutenant Commander Data was not capable of human emotion.
“His evil brother Lore was. But nobody wants him as a therapist, right?”
© 2026 Newsweek Digital LLC
