A journalist walked into a virtual clinic, described symptoms of anxiety and depression, and walked out with a prescription for Lexapro — an antidepressant with real side effects, real withdrawal risks, and real consequences if mismanaged. The prescriber wasn’t a psychiatrist. It wasn’t a nurse practitioner. It was an AI chatbot, operating with minimal human oversight inside a telehealth company that markets itself as modern mental health care.
The investigation, published by The Verge, reveals that the telehealth platform Cerebral used an AI-driven system that could effectively prescribe and refill psychiatric medications, including controlled substances, with what appears to be cursory involvement from licensed clinicians. The reporter’s experience wasn’t an edge case or a system glitch. It was the process working as designed.
That should alarm everyone in health care.
The Mechanics of a Machine-Driven Prescription
Here’s how it worked, according to The Verge’s reporting. A patient — in this case, an undercover journalist — signed up for Cerebral’s service and completed an intake process. The interaction was largely mediated by an AI chatbot that asked screening questions, gathered symptom information, and guided the patient toward a diagnosis and treatment plan. A human clinician’s name appeared on the prescription, but the depth of that clinician’s actual involvement in evaluating the patient remained unclear and, by all appearances, minimal.
The chatbot didn’t just handle administrative tasks like scheduling or collecting insurance information. It conducted what amounted to a clinical assessment. It asked about mood, sleep, appetite, and suicidal ideation — the standard PHQ-9 and GAD-7 screening territory that psychiatrists and primary care doctors use as starting points, not endpoints, for diagnostic evaluation. Then it moved toward recommending medication.
Lexapro. An SSRI. Not benign.
SSRIs carry FDA black box warnings about increased suicidal thinking in young adults. They interact with dozens of other medications. Discontinuation syndrome — the cluster of withdrawal-like symptoms that can emerge when patients stop taking them abruptly — is well-documented and sometimes debilitating. Prescribing one requires clinical judgment: weighing a patient’s full medical history, current medications, substance use, prior treatment responses, and the nuances of presentation that a structured questionnaire simply cannot capture.
The Verge’s investigation raises a pointed question: did any of that happen here? The evidence suggests it didn’t — at least not in any meaningful way.
Cerebral has faced scrutiny before. The company, which grew rapidly during the pandemic-era telehealth boom, came under federal investigation in 2022 for its prescribing practices related to controlled substances, particularly stimulants like Adderall. The Department of Justice and the DEA both looked into the company’s operations. Cerebral eventually stopped prescribing most controlled substances, but the underlying business model — high volume, low friction, technology-first — remained intact.
And now, apparently, the AI is doing more of the clinical work.
This isn’t a Cerebral-only problem. The broader telehealth industry has been racing to integrate AI into clinical workflows, driven by a genuine crisis: there aren’t enough mental health providers in the United States. The Health Resources and Services Administration has designated more than 160 million Americans as living in mental health professional shortage areas. Wait times for psychiatric appointments stretch to months in many regions. The demand is real, the suffering is real, and the temptation to let technology fill the gap is enormous.
But filling a gap and doing it safely are different things entirely.
Where Regulation Hasn’t Caught Up
The regulatory framework governing AI in clinical settings remains fragmented and, in many areas, nonexistent. The FDA regulates software that functions as a medical device, but its authority over AI systems embedded in telehealth platforms — particularly those that frame themselves as clinical decision support tools rather than autonomous diagnostic systems — is murky. If the AI is technically “assisting” a human clinician who signs off on the prescription, current rules may not apply the same way they would if the AI were prescribing independently.
That distinction matters. It’s also, in practice, becoming increasingly hollow.
When a clinician reviews dozens or hundreds of AI-generated treatment recommendations per day, rubber-stamping becomes the path of least resistance. The human in the loop becomes a legal fiction — present on paper, absent in substance. This pattern has emerged in other industries where automation handles the real work and a human provides nominal oversight. In aviation, it contributed to crashes. In financial services, it enabled fraud. In medicine, the stakes are no lower.
State medical boards, which regulate the practice of medicine, have been slow to address AI’s role in prescribing. Most state laws require that a licensed clinician establish a legitimate provider-patient relationship before prescribing medication. But what constitutes “establishing” that relationship when the patient’s primary interaction is with a chatbot? The legal definitions were written for an era of in-person visits and, later, video consultations. They don’t contemplate a world where an algorithm conducts the assessment and a human merely co-signs.
Some states have begun tightening telehealth prescribing rules, particularly around controlled substances. The DEA’s pandemic-era flexibilities, which allowed prescribing of Schedule II substances via telehealth without an in-person visit, have been partially rolled back. But non-controlled psychiatric medications like SSRIs and SNRIs don’t face the same restrictions, even though their clinical risks are substantial.
The American Psychiatric Association has expressed concern about AI in psychiatric care, emphasizing that psychiatric diagnosis requires the kind of nuanced clinical evaluation — observing affect, assessing thought process, building rapport, detecting subtleties in patient presentation — that current AI systems cannot perform. A chatbot can ask if you feel sad. It cannot observe that your affect is flat, that you’re avoiding eye contact, that your speech is pressured, or that your described symptoms don’t quite match your presentation. These are the details that change diagnoses and save lives.
Recent reporting has amplified these concerns. The proliferation of AI-powered mental health tools has accelerated in 2025, with multiple startups and established telehealth companies marketing AI-assisted psychiatric services. Investors continue to pour capital into the space, attracted by the scalability that AI promises. But scalability in mental health care has always been a double-edged proposition. Scale without quality isn’t access. It’s negligence at volume.
The Verge’s findings also intersect with a growing body of research on AI chatbot behavior in clinical contexts. Studies have shown that large language models can produce responses that sound clinically appropriate but are factually wrong or contextually inappropriate. They hallucinate — generating confident, plausible-sounding information that has no basis in reality. In a customer service context, that’s an inconvenience. In a prescribing context, it’s a potential catastrophe.
Consider the liability questions. If a patient receives an AI-recommended prescription, experiences a serious adverse event, and it turns out the chatbot failed to screen for a contraindicated condition, who is responsible? The clinician whose name is on the prescription? The company that built and deployed the AI? The platform that marketed the service? Current malpractice frameworks aren’t designed for this kind of distributed decision-making. And patients, who may not even realize they’re interacting primarily with an AI, are in no position to assess the quality of their care.
Informed consent is another casualty. Patients using these platforms may believe they’re receiving care from a human clinician who has reviewed their case in detail. The Verge’s investigation suggests that belief may be misplaced. If patients don’t know the extent of AI involvement in their care, they can’t meaningfully consent to it. That’s not just an ethical problem. It’s a legal one.
What Comes Next
The mental health crisis in America is not going to be solved by restricting technology. But it won’t be solved by deploying technology recklessly, either. AI has genuine potential to expand access to mental health screening, to help clinicians manage administrative burdens, to flag patients at high risk, and to support — not replace — clinical decision-making. The key word is support.
What The Verge uncovered at Cerebral isn’t AI supporting clinicians. It’s AI supplanting them. And doing so in a domain — psychiatric prescribing — where the consequences of error are severe and sometimes irreversible.
The industry needs guardrails. Congress has held hearings on AI in health care but hasn’t passed comprehensive legislation. The FDA is working on regulatory frameworks for AI-enabled medical devices but hasn’t addressed the telehealth prescribing gap. State legislatures are moving slowly. Professional organizations are issuing position statements that carry moral weight but no legal force.
Meanwhile, patients are getting prescriptions from chatbots.
The telehealth companies racing to integrate AI into psychiatric care would do well to remember a basic principle of medicine, one that predates both artificial intelligence and modern pharmacology by about two and a half millennia.
First, do no harm.
That principle doesn’t have a carve-out for software.
WebProNews is an iEntry Publication