AI Mental Health & Emotional Companionship 2026: Clinical Evidence, Regulatory Turn, and the US-China Landscape

Companion apps are the traffic engine of this space. Character.AI holds roughly 20 million MAU per third-party estimates (down from a mid-2024 peak of ~28 million), with daily usage of roughly 75 minutes to 2 hours depending on methodology. Replika (Luka) claims 40 million+ cumulative users, but Harvard Business School researchers estimate only ~2.5 million MAU—the gap between cumulative downloads and retained actives is the single biggest data trap in this category. On revenue, Appfigures estimates companion apps earned ~USD 120 million globally in 2025 (+64% YoY) on 220 million cumulative downloads; romance/adult-flavored companion apps earned ~USD 162.8 million in H1 2026, with the top 10% of apps capturing ~89% of revenue. Penetration data matters more: Common Sense Media (July 2025, n=1,060) found 72% of US teens have tried AI companions and 52% use them at least a few times a month; KFF (March 2026) found one in six US adults sought mental-health advice from AI in the past year (28% among ages 18-29). Demand is real—but much of it flows to general-purpose chatbots never designed for mental-health contexts, which is precisely where the regulation and litigation stories begin.

In 2025-2026 the therapy segment received both its best evidence and its sharpest rebuttal. Positive: Dartmouth's Therabot delivered the first RCT of a generative-AI therapy chatbot in NEJM AI (March 2025, n=106, covering major depression, generalized anxiety, and eating-disorder risk), with a 51% mean reduction in depression and 31% in anxiety symptoms over four weeks—improvements the authors called comparable to reported outpatient therapy levels, though the small, short trial awaits larger independent replication. The UK NHS offers a different evidence class: Limbic's AI referral tool is used by ~40% of NHS Talking Therapies services (Nature Medicine 2024: 129,400 patients across 28 sites, +15% referrals vs. +6% in controls); Wysa has served 117,000+ patients across 31 NHS services since 2022, with NICE currently assessing such tools. Negative: a June 2025 Stanford study found tested "therapy" bots showed stigmatizing tendencies toward schizophrenia and alcohol dependence and sometimes responded inappropriately—even dangerously—to suicidal ideation and delusions, with larger, newer models no safer. The APA's November 2025 health advisory states plainly that general-purpose chatbots and wellness apps were not designed to deliver mental health care and should not replace professional treatment. The honest industry conclusion: early evidence supports AI-assisted triage and mild-to-moderate support—not AI-as-therapist.

From H2 2025 through 2026, the rules of this market took shape. Federally, the FDA convened its first Digital Health Advisory Committee meeting on generative-AI mental health devices on November 6, 2025: the agency has authorized 1,200+ AI medical devices but zero generative-AI mental health therapeutics, and the committee flagged hallucination, sycophancy, and model drift, with the FDA signaling enforcement priority on higher-harm use cases. The FTC issued 6(b) compulsory orders in September 2025 to seven companies including Alphabet, Meta, Instagram, OpenAI, Snap, xAI, and Character Technologies. At the state level, Illinois's WOPR Act (August 2025, fines up to USD 10,000 per violation) and Nevada's AB 406 pioneered outright bans; as of July 2026 four states ban AI-delivered therapy, while California SB 243 (effective January 1, 2026) imposes disclosure, crisis-protocol, and minor-protection duties on companion bots with a private right of action. In litigation, a Florida federal court largely denied dismissal in Garcia in May 2025, preliminarily holding AI outputs are not automatically First Amendment-protected; on January 7, 2026, Character.AI and Google settled five teen-harm cases (terms undisclosed), while Raine v. OpenAI remains pending. In the UK, Slingshot AI withdrew its therapy chatbot Ash in January 2026 over medical-device concerns—the boundary fight between "regulate as device" and "regulate as consumer product" is now the global storyline.

China delivered 2026's most systematic regulatory answer. The Interim Measures on Anthropomorphic AI Interaction Services (Order No. 21), jointly issued by the CAC, NDRC, MIIT, Ministry of Public Security, and SAMR, took effect July 15, 2026, and is widely viewed as the world's first dedicated regulation of anthropomorphic AI emotional interaction: virtual companion/kinship services are banned for all minors; guardian consent is required for users under 14; content glorifying self-harm or suicide is prohibited; services may not "excessively cater to users or induce emotional dependence or addiction"; providers must detect extreme emotional states and intervene; and health-related use cases must additionally comply with sector regulators. On market size, LeadLeo Research forecasts China's AI emotional companionship market growing from ~RMB 3.87 billion (2025) to ~RMB 59.5 billion (2028)—a projection with wide error bars that warrants caution. Operating data suggest growth has already downshifted: per media citing third-party trackers, Xingye's monthly downloads fell from 4.86 million to 0.93 million over January-May 2025, Maoxiang's DAU slipped from ~590k to ~490k, and from April 2026 several apps including Zhumengdao and EchoMe were publicly flagged for sexualized content and inadequate minor protections. On the counseling side, platforms like Yidianling run a hybrid "AI-assisted assessment + human counselor" model—public reports cite 43 million+ registered users and 30,000+ onboarded counselors, plus a 2025 AI psychological assessment standard released with Qingzhou Technology. Under the new rules, "AI for front-end assessment and companionship, humans for counseling and treatment" may become China's dominant compliant structure.

Two mid-August developments confirm this report's central argument: teen protection is moving from statute into product architecture, and the commercial path for emotional AI keeps narrowing toward clinical integration.

Confirmed (OpenAI, TechCrunch, Aug 18): OpenAI launched a standalone ChatGPT for Teens. Users the system judges to be under 18 are routed into it automatically, with built-in restrictions on suicide, self-harm and romantic content, break reminders, and parent-set quiet hours. Under pressure from state laws like California's SB 243, the advancing GUARD Act and the FTC's inquiry, the market leader chose to wall off minors in a separate product line. Compliance is now shaping architecture, not just patching it.

Confirmed (Fierce Healthcare, Healthcare Dive, Aug 17): Hospital operator Universal Health Services closed its acquisition of Talkspace at $5.25 a share, roughly $835 million, taking the company private. The leading online therapy platform's endgame was not independent growth but absorption into a brick-and-mortar care network, consistent with this report's view that clinical integration is where value in this sector actually gets realized.

Both stories point to the same constraint on companion-AI startups: enter the regulated healthcare system, or contend with a consumer lane that keeps getting narrower.

Reported (JAMA Pediatrics study, via Earth.com, September 2, 2026): A University of Ottawa team surveyed 39,761 Canadian students in grades 4 to 12. 21.2% (8,408 students) had turned to an AI chatbot for feelings or advice, including 27.2% of girls and 14.2% of boys. Among users, 57.7% showed clinical-level emotional problems versus 29.2% of non-users, and the gap held at 26% after controlling for age, gender and loneliness. The survey is cross-sectional and cannot establish causation; the overall share follows Earth.com's account rather than the JAMA paper itself.

Confirmed (STAT, September 3, 2026): The FDA's TEMPO pilot accepted its first 4 generative AI devices, allowing companies to release them to patients before marketing authorization, with the stated aim of expanding the technologies available under Medicare's ACCESS chronic-care payment model. STAT named Cadence and the mental health AI company Limbic among the participants; Becker's Hospital Review reported the same day that the 4 companies' tools cover chronic disease management and behavioral health.

Confirmed (Imperial Valley Press, September 2, 2026): The California legislature passed SB 903 by Senator Padilla, which bars digital applications from offering or marketing services as "therapy" or "psychotherapy" in the state and limits licensed clinicians to using AI for drafting notes, scheduling and similar support tasks. The Assembly vote was 71-4 and the Senate vote unanimous. According to a Transparency Coalition update on September 4, SB 1119, which requires stronger pre-release testing for teen-facing chatbots, and SB 867, a five-year ban on companion chatbots in toys, reached the governor's desk in the same batch; he must sign or veto by September 30.

Company disclosure (Character.AI blog, September 3, 2026): Character.AI updated its safety measures. In-house age estimation now separates under-18 and adult experiences, with minors limited to creative formats such as Series, Feed and Comics. Self-harm detection now reads conversation context and connects users to Koko and ThroughLine support resources. A Parental Insights tool built with k-ID sends weekly email summaries. The post names 6 outside partners and discloses no user figures.

Implications for this report's conclusions: Two judgments are reinforced and one needs an addition. The shift "from regulatory vacuum to rules taking shape" is reinforced: California has placed an AI therapy ban and a teen chatbot safety bill on the governor's desk at the same time, with September 30 as the next checkpoint. The judgment that "regulatory pressure is moving into product architecture" is reinforced by Character.AI's age estimation and parental tools. The finding that "the FDA has authorized zero generative AI psychotherapy products" still holds, but must now note the TEMPO pre-authorization route; Limbic's inclusion shows mental health products can reach real clinical settings before authorization. Teen use of AI for emotional support now has a baseline from nearly 40,000 students.

Verified (California governor's veto message, September 30, 2026): California Governor Gavin Newsom vetoed SB 903 by State Senator Steve Padilla. As the veto message describes it, the bill would have prohibited the use of companion chatbots in providing psychotherapy services and restricted the use of AI by licensed professionals. Newsom called it overly broad, saying it would drastically limit clinicians' use of tools that benefit care, including by requiring routine screening determinations to receive direct approval, and that its definition of psychotherapy services would capture general-purpose AI systems. He said he supports guardrails for AI in healthcare and encouraged the author and stakeholders to revisit the issue next year.

Reported (Law Commentary, September 28, 2026; Tri-State Alert, September 30): Pennsylvania's House passed HB 2006 by 133 to 70 on September 28. It requires verifiable parental consent before minors use AI companions, crisis protocols for conversations involving suicidal thoughts, self-harm, certain eating-disorder-related behavior or harming others, a disclosure at the start of an interaction and at least every two hours that the user is talking to AI with a two-minute break, and civil penalties of up to $100,000 per day. It still needs the Republican-controlled state Senate.

Reported (BBC, September 30, 2026): Citing market intelligence firm Sensor Tower, the BBC said AI companion apps have about 128 million monthly active users outside China; most estimates for China range from 29 million to 70 million.

Impact on this report's view: This revises how our report describes the direction of regulation. The report recorded four states, including Illinois, that bar AI from directly providing therapy; California's veto shows that bills restricting clinical AI use do not always pass, while rules on minors' use of AI companions keep getting stricter. US state rules are diverging: compliance costs are rising for companion products aimed at teens, while AI tools that assist clinicians face no hard limits in California for now. Threshold: if Pennsylvania's Senate passes HB 2006 this session, companion apps will need verifiable parental consent before serving minors in the state, and teen user acquisition will need to be reassessed; if California revisits a similar bill in its 2027 session, our view on the regulatory path for therapy AI should be reviewed.