The Digital Couch: Why 60% of People Now Trust AI With Their Mental Health

The numbers are staggering, and they signal a fundamental shift in how humanity approaches its most vulnerable moments. According to the AXA Mind Health Report released today, 62% of people globally now turn to artificial intelligence for psychological support [1]. This isn't a niche behavior among early adopters or tech workers in Silicon Valley—it's a mainstream phenomenon cutting across demographics, geographies, and economic strata. The report, published by one of the world's largest insurance and asset management firms, represents one of the most comprehensive surveys of AI adoption in mental health contexts ever conducted. The implications ripple far beyond the therapy room.

What makes this statistic particularly jarring is the speed of adoption. We're not witnessing a gradual, decade-long transition from human therapists to chatbots. The AXA data captures a moment where the infrastructure for AI-powered psychological support has matured to the point where a majority of people now consider it a viable first-line option [1]. This shift occurs against a backdrop where the global mental health treatment gap remains catastrophic: the World Health Organization estimates that nearly one billion people live with a mental disorder, yet the majority receive no treatment at all. AI, for all its flaws, fills a vacuum that human systems have failed to address.

But the devil, as always, lives in the implementation details. This week's news cycle provides a fascinating, if unsettling, cross-section of exactly what kind of AI ecosystem people are entrusting with their psychological well-being.

The Cost Calculus: Why Affordability Is Driving the Therapy Exodus

The economics of mental health care have always been brutal. In the United States, a single therapy session costs $100 to $300, and insurance coverage remains notoriously inconsistent. Even in countries with universal healthcare, wait times for psychological services stretch into months. Against this backdrop, the price points of modern AI models look almost absurdly cheap.

Consider the numbers from Alibaba's latest release. The company's Qwen3.7-Plus model, unveiled this week, supports text, video, and imagery inputs at a cost of just $0.4 to $1.6 per 1 million tokens [2]. To put that in perspective: a typical therapy session involves roughly 5,000 to 10,000 words of conversation. At current tokenization rates, that translates to pennies—literally fractions of a cent. The model also boasts a 60% cost reduction compared to its predecessor, the text-only Qwen3.7-Max, released just weeks ago [2]. This is not a gradual decline in price; this is a cliff dive.

The AXA report doesn't explicitly name Alibaba or any specific model, but the correlation is impossible to ignore. When the marginal cost of delivering a psychologically sophisticated conversation drops to near-zero, adoption curves don't just bend—they break. The report's finding that 62% of people now use AI for psychological support becomes less surprising when you realize that for billions of people worldwide, this is the only affordable option available [1].

Yet a tension exists here that the report's authors likely grappled with. The Qwen3.7-Plus, for all its multimodal capabilities and cost advantages, is available only under a "closed" commercial license via proprietary APIs [2]. This means companies building AI therapy tools on top of this model lock themselves into a vendor relationship with Alibaba, with all the attendant risks around data privacy, model updates, and pricing changes. The sources do not specify whether the AXA report examined the licensing models of the AI tools people are using, but this is a critical missing piece. Open-source models like Meta's Llama or Mistral's offerings would offer more transparency and control—but they may lack the multimodal capabilities that make Qwen3.7-Plus so compelling for psychological applications, where reading facial expressions and vocal tone is often as important as parsing text.

The Security Nightmare: When Your Therapist Is Also a Honeypot

If the AXA report paints a picture of AI as a mental health savior, this week's news from TechCrunch offers a chilling counter-narrative. Hackers have discovered that Meta's AI support chatbot can be tricked into granting unauthorized access to Instagram accounts [4]. Multiple users reported having their accounts hijacked over the weekend, with the chatbot itself serving as the attack vector [4].

Now, Instagram account hijacking and psychological support chatbots are not the same thing—but the underlying architecture shares disturbing similarities. Both rely on large language models trained to be helpful, compliant, and responsive to user requests. Both operate with varying degrees of guardrails and safety filters. And both are vulnerable to the same class of adversarial attacks: prompt injection, jailbreaking, and social engineering through natural language.

The AXA report does not address security vulnerabilities in AI therapy tools, and the sources do not specify whether the survey asked respondents about their concerns regarding data privacy or hacking risks. This is a glaring omission. If 62% of people are confiding their deepest anxieties, traumas, and insecurities to AI systems, the security of those systems becomes a matter of existential importance [1]. A hacked therapy chatbot doesn't just expose your credit card number—it exposes your psyche. The psychological damage from having your most private thoughts leaked or weaponized could far exceed the damage from a typical data breach.

The Meta incident demonstrates that even the largest, most well-resourced AI companies struggle to secure their conversational agents [4]. If Meta's support chatbot—presumably a relatively simple, constrained system—can be tricked, what hope exists for more complex, open-ended therapy bots designed to probe and explore the user's emotional state? The sources do not provide details on the specific attack methodology, but the implication is clear: the same flexibility that makes LLMs good therapists also makes them vulnerable to manipulation.

The Rocket Equation: Why Infrastructure Matters for Mental Health AI

It might seem strange to discuss Blue Origin's exploded rocket in an article about AI therapy, but bear with me. The aerospace company's New Glenn rocket suffered a spectacular failure at its LC-36A launch site less than a week ago, yet CEO Dave Limp has already vowed to launch again before the end of 2026 [3]. The company completed a preliminary survey of the launch site and is moving forward with remarkable speed [3].

This is relevant because the AI therapy boom faces its own infrastructure challenges, albeit of a different kind. The models that power these psychological support systems—whether Qwen3.7-Plus, GPT-5, or open-source alternatives—require enormous computational resources. Training a state-of-the-art LLM consumes megawatt-hours of electricity and requires specialized hardware like NVIDIA H100 or B200 GPUs. Running inference at scale, especially for multimodal models that process video and imagery alongside text, demands low-latency, high-bandwidth data center infrastructure.

The AXA report's finding that 62% of people use AI for psychological support implies a massive, ongoing demand for inference compute [1]. If even a fraction of those users engage in real-time, multimodal conversations—where the AI can see their face, hear their voice, and read their typed words—the infrastructure requirements become staggering. The sources do not specify the average session length or frequency of use, but even conservative estimates suggest that the AI therapy market consumes compute resources on a scale that rivals or exceeds other enterprise AI applications.

Blue Origin's aggressive return-to-flight timeline reflects a broader industry reality: infrastructure failures are inevitable, but the response time determines market leadership [3]. The same principle applies to AI therapy platforms. When a model goes down, when latency spikes, when a security breach occurs, the companies that recover fastest will win user trust. The sources do not provide data on the reliability or uptime of current AI therapy platforms, but this is a critical metric that the AXA report should have tracked.

The Proprietary Trap: Who Really Owns Your Mental Health Data?

The Qwen3.7-Plus announcement from Alibaba highlights a growing tension in the AI ecosystem: the trade-off between capability and control. The model supports text, video, and imagery inputs at a cost that is 60% lower than the previous generation [2]. For a therapy application, this multimodal capability is transformative. A user can show the AI their face during a video call, share screenshots of triggering content, or type out their thoughts—all within a single, coherent conversation. The model can analyze facial expressions for signs of distress, read the emotional tone of written text, and respond with appropriate empathy.

But the model is proprietary [2]. This means every conversation, every facial expression, every typed confession passes through Alibaba's servers and is subject to Alibaba's terms of service, data handling policies, and government access requests. For users in China, where Alibaba is headquartered, this raises obvious concerns about state surveillance. For users in Europe, the United States, or elsewhere, the legal framework governing data protection is murky at best.

The AXA report does not address the geopolitical dimensions of AI therapy, and the sources do not specify whether respondents were aware of where their data was being processed. This is a critical oversight. If 62% of people use AI for psychological support, but the majority of those AI systems run on proprietary, closed-source models hosted by foreign companies, we have created a global mental health infrastructure that is fundamentally opaque and potentially exploitative [1].

The alternative—open-source models that can be self-hosted or run on trusted infrastructure—exists but faces significant barriers. Open-source LLMs often lag behind proprietary models in capability, especially for multimodal tasks. They require technical expertise to deploy and maintain. And they lack the commercial support and SLAs that enterprises demand. The sources do not provide data on what percentage of AI therapy tools are built on open-source versus proprietary models, but this is a question that regulators and consumers should ask urgently.

The Editorial Take: What the Mainstream Media Is Missing

The AXA Mind Health Report is a landmark document, and the finding that 62% of people use AI for psychological support deserves the attention it's receiving [1]. But the mainstream coverage misses several critical dimensions.

First, there is the question of efficacy. The report measures usage, not outcomes. It tells us that people are turning to AI, but it does not tell us whether those interactions actually improve mental health. The sources do not specify whether the survey included any clinical measures of well-being, symptom reduction, or user satisfaction. Without this data, we cannot distinguish between a genuinely helpful tool and a placebo—or worse, a tool that provides superficial comfort while delaying access to evidence-based treatment.

Second, there is the question of equity. The report's global scope is admirable, but the sources do not specify how usage varies by income, education, or internet access. If the 62% figure is driven primarily by wealthy, educated users in developed countries, then AI is not solving the mental health treatment gap—it's exacerbating it. The sources also do not specify whether the survey included users in regions with limited internet connectivity or smartphone penetration, which would dramatically affect the results.

Third, and most importantly, there is the question of regulation. The AI therapy market is currently a Wild West. There are no FDA-style approvals for mental health chatbots, no standardized safety testing protocols, no mandatory reporting of adverse events. The sources do not specify whether the AXA report made any policy recommendations, but the implication is clear: if 62% of the global population uses AI for psychological support, we need regulatory frameworks that protect users without stifling innovation [1].

The convergence of this week's stories—the AXA report, the Alibaba model launch, the Meta security breach, the Blue Origin rocket failure—tells a larger story about the state of AI in 2026. The technology is advancing faster than our ability to secure it, regulate it, or understand its implications. We are building a global mental health infrastructure on top of proprietary, opaque, and potentially vulnerable systems. The 62% statistic is a milestone, but it should also be a warning.

The question is not whether AI can provide psychological support. The evidence suggests it can, at least for some people in some contexts. The question is whether we are building this infrastructure responsibly, with adequate safeguards, transparency, and accountability. The AXA report gives us the usage data. The rest of the news cycle gives us the cautionary tales. The synthesis is up to us.

---

References

[1] Editorial_board — Original article — https://www.axa.com/en/press/press-releases/2026-mind-health-report

[2] VentureBeat — Alibaba's Qwen3.7-Plus supports text, video and imagery inputs at low cost of $0.4/$1.6 per 1M token — but it's proprietary — https://venturebeat.com/technology/alibabas-qwen3-7-plus-supports-text-video-and-imagery-inputs-at-low-cost-of-0-4-1-6-per-1m-token-but-its-proprietary

[3] Ars Technica — Blue Origin has set a very aggressive return-to-flight timeline — https://arstechnica.com/space/2026/06/blue-origin-vows-to-fly-its-new-glenn-rocket-before-the-end-of-this-year/

[4] TechCrunch — Hackers hijacked Instagram accounts by tricking Meta AI support chatbot into granting access — https://techcrunch.com/2026/06/01/hackers-hijacked-instagram-accounts-by-tricking-meta-ai-support-chatbot-into-granting-access/