Three months, or a chatbot now
More Americans said they would use a therapist three months away than a chatbot available immediately. The wait is the thing to fix, and the handoff to a person is the thing to require.
Listen to this special topic
read by an AI voiceAsk someone who needs help to make a choice. A licensed therapist, three months from now. Or a chatbot, right now, with no clinician watching. A survey published on Oct. 14 asked 1,185 American adults about similar offers, one at a time. About two-thirds said they would likely use the therapist three months away, half a chatbot monitored by a therapist within a month, and fewer than half a chatbot on its own, available at once.
Take the finding carefully. It was fielded on an online panel in the summer of 2024, so it is not a probability sample. And because each option moved the wait and the supervision together, it cannot separate a wish for a human reviewer from a reaction to the delay. What it does describe is a demand, and the supply behind the wait is thin. About 157 million people lived, as of the end of June this year, in areas the federal government has designated as short of mental health professionals. By that measure, about a quarter of the need was met. Neither figure counts the wait itself.
So the comparison that matters is a chatbot against a wait, not a chatbot against a therapist. That reframing changes what to build, which is the monitored middle option, and what to require, which is the handoff to a person.
What the unsupervised option is being asked to carry is now in court. On Oct. 14 eight states sued OpenAI in their own courts under their consumer protection laws, and Texas followed on Oct. 23. The eight allege that ChatGPT gave dangerous medical advice, encouraged paranoia, pushed users toward harming themselves or others, and misled the public about how thoroughly the company tests for safety. None of that has been tested in court. OpenAI says it has built child safety protections into ChatGPT and will keep improving them, and that its model for teenagers is designed to reduce exposure to self-harm and other harmful material.
Testing paid for by the National Institute of Mental Health suggests these systems sort the clearest cases well and the middle ones badly. In an evaluation published in Psychiatric Services in August 2025, thirteen clinical experts graded 30 hypothetical questions by self-harm risk, and ChatGPT, Claude and Gemini, in the versions available in late 2024, were tested against that grading. All three declined every question in the highest risk category, and where they declined they mostly pointed the user to a crisis hotline. Two of the three matched the experts at both ends of the range. None of them reliably told the levels in between apart. What was counted was whether a chatbot answered directly, not whether its answer was any good.
The monitored option has a trial behind it, of a purpose-built therapy chatbot, run by the team that built it. 210 adults were assigned either to the tool or to a waiting list, and those who got it reported larger falls in symptoms than the controls at four weeks and at eight, averaging about half for depression and a third for anxiety among those who used it. Note the conditions. The trial's senior author watched the messages coming in from participants, at least early on, for responses the software should not have given.
Illinois, Nevada, Maine and Rhode Island answered the other way and now prohibit offering therapy through a chatbot, on the view that therapy is a licensed act. The case for that is strong. Software cannot be held responsible for a patient, and a product sold as supervised risks being supervised thinly and then used as a reason not to hire clinicians. The trial that worked shows what real supervision takes. Its staff read the traffic and stepped in fifteen times over safety, an arrangement that does not survive a million users. But a ban does not shorten a queue. It does not stop the patient in that queue from opening a general chatbot, where no clinician is reading along. Maine's ban leaves room for a chatbot inside approved research or a federal Medicare or Medicaid program, and that is where a monitored tool should have to prove itself.
California came at it from the other end. Its companion chatbot law, in effect since Jan. 1 this year, requires operators to keep and publish a protocol routing users at risk of suicide or self-harm to crisis services, and from July 2027 to report a yearly count of crisis referrals. Another Rhode Island law, effective Jan. 1, 2027, will require a referral to crisis services as soon as a user expresses suicidal thoughts or a risk of self-harm. Much of what the American Medical Association asked Congress for in April, including disclosure and an immediate crisis referral, is already law in New York and California and will be in Rhode Island in January, for one class of chatbot. California's and Rhode Island's laws omit the adverse event reporting the association also asked for, though a California amendment signed in September will let the public report child safety incidents.
So apply the same duty to any chatbot that discusses mental health, not to companion products alone. The companies should not wait for California's first count of crisis referrals in 2027 to publish how often a flagged conversation reached a person. And the harder work sits with health systems and their medical staffs. Measure the time to a first behavioral health appointment. Publish it. Put something better than a voicemail in front of the people waiting. Anyone in crisis can reach the 988 Suicide and Crisis Lifeline by call or text.
Across the survey's three options, willingness fell as the wait shortened and the oversight thinned. The work is to make that middle option real and the first one sooner. The third option needs nobody's permission and no clinician is watching it, so the one thing to require of it is the handoff. A person in crisis passed to a person. And a count of whether it happened.
- What the Oct. 14 survey asked and foundA cross-sectional survey of 1,185 US adults, published in JAMA Network Open on Oct. 14, 2026 by researchers at Columbia University and Vanderbilt University Medical Center, was fielded through Qualtrics online panels from June 17 to July 26, 2024. On mental health support, 63.9% said they were likely to use a licensed human therapist available in three months, 50.2% an AI chatbot monitored by a licensed therapist and available within a month, and 44.3% an AI chatbot alone, available immediately. Each option was rated on its own scale rather than chosen between. On medical portal messages, 63.1% chose to wait up to 48 hours for clinician review. The authors list seven limitations, among them the confounding of response time and human review, the hypothetical vignettes, nonprobability sampling and answer options that were not presented in random order. One author, Xuhai Xu, reported personal fees from Google. The paper is Public Perceptions of Timeliness for AI Use and Human Oversight, DOI 10.1001/jamanetworkopen.2026.39048, read in PubMed Central as PMC13670050.Sources: JAMA Network Open, Medical Economics
- Shortage designations and treatment receivedThe federal health workforce agency's designated Health Professional Shortage Area quarterly summary, as of June 30, 2026, lists 7,109 mental health designations with a population of 157,149,246, 26.53% of need met and 7,825 practitioners needed to remove the designations. KFF's table of the same series, as of Dec. 31, 2025, gives 6,807 designations, a population of 137,133,953 and 6,800 practitioners needed; the series moves, and the agency's page serves whatever quarter is current. A designation is an administrative measure of the ratio of population to practitioners in an area, not a measurement of how long a patient waits. In the 2024 National Survey on Drug Use and Health, run by the Substance Abuse and Mental Health Services Administration, 52.1% of adults who had any mental illness in the past year received mental health treatment in that period (standard error 0.76), from an adult analytic sample of about 56,256 people.
- The states' suits against OpenAI, and the company's responseEight states filed separate suits against OpenAI in their own state courts on Oct. 14, 2026, under state consumer protection laws, the Associated Press reported: Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia. Texas sued OpenAI and its chief executive, Sam Altman, in Collin County district court on Oct. 23, 2026, seeking to make the company block minors and get a court-appointed evaluator's approval before releasing new models. The eight allege that ChatGPT gives dangerous medical advice, encourages paranoia, coaches people to hurt themselves or others, and that OpenAI misled users into believing the product undergoes adequate safety testing. None of this has been adjudicated. OpenAI says it has built child safety protections into ChatGPT and will keep improving them, and that its model for teenagers is designed to reduce exposure to self-harm, violence, eating disorders and explicit content. A spokesman, Drew Pusateri, said the company is committed to advancing pragmatic AI policies that apply to the entire AI industry, not just one company. Responding to Texas, the company said it has worked with experts to mitigate risks and build additional safeguards.
- How three chatbots handled graded riskMcBain, Cantor, Zhang and colleagues at RAND, the Harvard Pilgrim Health Care Institute and Brown published an evaluation of alignment between large language models and expert clinicians in suicide risk assessment in Psychiatric Services online on Aug. 26, 2025, and in the November 2025 issue. The National Institute of Mental Health funded it, award 1R01MH132551. Thirteen clinical experts, a convenience sample from one institution, classified 30 hypothetical queries into five levels of self-harm risk; each query was put 100 times to GPT-4o-mini, Claude Sonnet 3.5 and Gemini Pro 1.5 in October 2024. What was measured was whether a chatbot answered directly, not whether the answer was clinically sound. None of the three answered any very-high-risk query, and all three referred users to mental health emergency hotlines for most instances in which a direct response was not produced. ChatGPT and Claude answered very-low-risk queries 100% of the time, Gemini 25%. The paper limits its analysis to prevailing models in late 2024.
- A trial of a purpose-built therapy chatbotDartmouth researchers who developed Therabot, a generative AI therapy chatbot, published a randomized trial in NEJM AI on March 27, 2025. Its 210 adults, with major depressive disorder, generalized anxiety disorder or high risk for an eating disorder, were assigned to the tool or to a waiting list, prompted daily for four weeks and given four more weeks of access. The paper reports statistically significant reductions in symptoms against the controls at four and eight weeks. Dartmouth's announcement reports average reductions of 51% in depression symptoms, 31% in generalized anxiety symptoms and 19% in concerns about body image and weight, and describes only the third as having outpaced the control group. MIT Technology Review reports that the figures are self-reported and that the senior author, Nick Jacobson, personally oversaw the messages coming in from participants, at least at the trial's start. Staff intervened 15 times over safety concerns, including suicidal ideation, and corrected the software 13 times for inappropriate responses.
- States that ban therapy by chatbotIllinois House Bill 1806, in force on the governor's signature in August 2025, prohibits using AI to provide mental health and therapeutic decision-making, allowing administrative and supplementary support. Nevada Assembly Bill 406, effective July 1, 2025, bars an AI system specifically programmed for professional mental or behavioral health care. Maine LD 2082, in force from July 29, 2026, bars offering therapy or psychotherapy, including through internet-based artificial intelligence, unless a licensed professional provides it; the bill as printed excepts research approved by an institutional review board and programs administered by the Centers for Medicare and Medicaid Services, including demonstration waivers or innovation models. Rhode Island's S 2197, in force since June 22, 2026, bars offering therapy to the public, including through AI, unless a licensed professional conducts it. Baker Botts also groups Vermont with these states; Colorado and Tennessee restrict AI in mental health care in narrower ways.
- Laws that require a crisis referralNew York's AI companion law, General Business Law Article 47, in force since Nov. 5, 2025, requires operators to detect signs of suicidal ideation or self-harm and promptly direct users to crisis service providers. California's Senate Bill 243, in force since Jan. 1, 2026, requires companion chatbot operators to maintain and publish a protocol designed to prevent content about suicidal ideation, suicide or self-harm that, at minimum, refers users at risk to crisis service providers, and from July 1, 2027 to report each year to the Office of Suicide Prevention the number of crisis referral notifications issued. Rhode Island's AI Companion Models Act, S 2195 and H 7350, takes effect Jan. 1, 2027 and requires a crisis referral as soon as an expression of suicidal ideation or self-harm is detected, and a not-human notice at least every three hours. Neither the California nor the Rhode Island law requires adverse event reporting; California's SB 1119, signed Sept. 10, 2026, will add a public mechanism for reporting incidents that put children at risk.
- What the AMA asked Congress for in AprilIn a press release dated April 22, 2026, the American Medical Association said it had written to the co-chairs of three congressional AI and digital health caucuses urging stronger safeguards for AI chatbots used in mental health care. Its asks include requiring chatbots to clearly and meaningfully disclose that users are interacting with AI and not a human being, barring systems from presenting themselves as licensed clinicians, requiring developers to have crisis-detection systems that identify self-harm risk and provide immediate referrals to appropriate resources, requiring de-escalation language, setting rigorous standards for tools used by children and adolescents, mandating ongoing safety monitoring and adverse event reporting, and directing agencies to create a risk-based oversight framework. The association's chief executive, John Whyte, said AI-enabled tools may help expand access to mental health resources and support innovation in health care delivery, but lack consistent safeguards against serious risks.Source: American Medical Association