Nine Myths About AI in Mental Health, Debunked by the Data

The US adds roughly 27,000 newly licensed therapists a year. Against that, roughly 2 million new patients seek therapy annually, and by some estimates, twice that number actually need it. Run that math forward and the country is on track for a shortfall of around 150,000 mental health professionals by 2038, with the existing workforce aging out and virtually no growth in psychiatrists or psychologists to replace them.¹

That workforce math was the backbone of a recent talk given to a group of mental health builders by a clinical R&D lead who has worked on AI mental health products at several companies, walking through nine common assumptions about AI in mental health. Some of the myths are easy to dismiss. A few of them are genuinely uncomfortable, and worth sitting with regardless of which side of the "AI in mental health" debate you're on. We're citing the specific research below rather than presenting it as any one company's proprietary finding, because the underlying question, does purpose-built, clinically-grounded AI behave differently from general-purpose AI in mental health contexts, matters far beyond whichever product happens to have published data on it first.

Myth 1: The system is basically fine, so AI only adds risk

It isn't fine, and the numbers make the case better than any anecdote could. Seventy-five percent of people report needing help with their mental health. Sixty-six percent say they need more emotional support, not necessarily more formal therapy. Fifty-four percent of people who are struggling receive no support at all. Layer the workforce shortfall on top of that demand, and the honest conclusion is that the system was never going to scale to meet existing need with the current supply of licensed clinicians alone, AI or no AI.

Myth 2: All AI is essentially the same, and frontier models are the leading edge

Large language models can be trained toward many different behavioral distributions. Frontier, general-purpose models have picked what researchers describe as an "assistant" distribution by design, optimized for broad helpfulness across every possible use case, which produces certain default behaviors that are not necessarily well-suited to mental health conversations specifically.² It's also worth noting that these models don't reliably improve on mental health-specific measures release over release; some vendor-run safety evals have found later frontier model versions scoring worse than earlier ones on mental health-specific intelligence benchmarks, a genuinely counterintuitive finding that undercuts the assumption that general capability gains automatically translate into better mental health support.

There's also emerging evidence that users notice a real difference between purpose-built and general-purpose tools, though the strongest available data on this point currently comes from a single company's own research rather than independent replication, which is a real limitation worth flagging rather than glossing over. One published study following the withdrawal of a purpose-built AI wellbeing tool from a market found that only 27% of displaced users switched to a general-purpose AI tool afterward; the remaining 73% disengaged from AI support entirely, returned to family and friends, or sought other clinical services.³ That's suggestive, not conclusive, but it's a genuinely different pattern than you'd expect if general-purpose chatbots were a frictionless substitute for a purpose-built tool.

Myth 3: Only humans can help humans, and RCTs are the only valid measure

Multiple published studies now show AI-mediated support producing meaningful mental health improvements, sometimes through mechanisms that differ from human-delivered support rather than simply replicating it. The more useful question for any purpose-built mental health AI isn't "does it beat a human," it's whether the tool can demonstrate known therapeutic mechanisms, like behavioral activation or goal attainment, actually moving as engagement increases, and whether that mechanism-level evidence has been independently validated rather than only reported by the company that built the product.

Myth 4: AI promotes social isolation, and anthropomorphism is inherently dangerous

Research from MIT's Media Lab has found a real correlation between general-purpose AI use and loneliness.⁴ Correlation, however, is not the same as inevitability, and the mechanism matters more than the headline. A separate RCT from the same MIT research group, examining anthropomorphic design features specifically, found that more anthropomorphic features did not drive problematic use the way conventional wisdom predicted.⁵ The relationship between design choices and harm is more complicated than either side of the AI-in-mental-health debate typically presents it, and a tool's tendency to actively refer users back toward human relationships, rather than its degree of "personality," appears to be the more relevant design variable.

Myth 5: Talking to AI for a long time is inherently bad for people

At least one published six-month follow-up study on users of a purpose-built mental health AI tool found depression and anxiety symptoms remaining low, with relapse rates lower than what's typically seen following a course of human therapy. One hypothesis for the difference: because the AI remembers exactly where a user left off, people can return to it on an as-needed basis rather than losing continuity between formal sessions, which may prevent the drop-off in gains that often follows the end of treatment. This is not an argument that duration is irrelevant to safety, and it's worth reading alongside the opposite, equally important finding from a different mental health AI company, that unbounded session length can itself become a design failure worth actively correcting for. Both things can be true: continuity between sessions helps, and unstructured, unbounded engagement can still be a problem regardless of how good the underlying model is.

Myth 6: Sycophancy is the main driver of problematic AI use

Research on social sycophancy in language models does confirm that people mildly prefer models that validate them.⁶ But the effect sizes reported are small, and the comparison conditions used in some of that research prompted the non-sycophantic models in ways that degraded their performance more generally, meaning some of the observed "preference" may reflect a performance gap rather than a pure sycophancy effect. It's also worth benchmarking against human behavior: humans themselves are measurably sycophantic in a meaningful share of everyday supportive messages. The useful design target is probably somewhere well below typical human sycophancy, not zero.

Myth 7: Standardized benchmarks are essential for the field

This is the most contrarian claim in the talk, and it's a defensible one. There is currently no published benchmark for mental health AI safety that has been shown to actually predict real-world user outcomes. Optimizing a model to score well on a benchmark it can be trained toward risks producing exactly the distortion that shows up in other AI domains: models that perform impressively on the test and unremarkably in the situations the test was meant to represent. Benchmarks also invite an infinite regress problem (who benchmarks the benchmark) and a politicization risk, since competing companies have obvious incentives to favor whichever standard makes their own product look best, including, worth being honest about, any company writing about this topic in its own blog. None of this means testing should stop. It means real-world outcome tracking, not benchmark performance or self-reported case studies, should be the thing organizations, and the people evaluating them, are actually weighing most heavily.

Myth 8: Experts and consumers already know what mental health AI should look like

Neither group has this figured out yet, and pretending otherwise forecloses genuinely useful design space. The honest posture is to stay open to what these tools become rather than forcing them into the shape of existing, pre-AI mental health delivery models.

Myth 9: Safety is the biggest barrier to AI adoption in mental health

The bigger practical barrier is affordable deployment of sophisticated, purpose-built models at scale. Frontier-model usage today is subsidized by an enormous amount of investor capital, which means the true, unsubsidized cost of delivering a comparably capable purpose-built model is higher than current market pricing suggests. Solving that cost problem, not just the safety problem, is what determines whether purpose-built mental health AI can actually reach the scale the workforce shortage demands.

Who this is actually for

A useful reframe worth borrowing regardless of which product you're evaluating: rather than thinking in terms of mild, moderate, and severe illness, think about what's sometimes called the "missing middle." Research suggests that a majority of people currently in therapy don't meet full diagnostic criteria for a disorder at all; they're seeking a kind of support that a $250-an-hour, weekly-session system was never designed to deliver efficiently. That's a real, large population AI can plausibly serve well. Even for more severe presentations, like schizophrenia, there's a meaningful distinction between positive symptoms (delusions, hallucinations, which remain the domain of pharmacological management) and negative symptoms (anhedonia, depression-like presentations, where AI-mediated support may have something genuine to offer alongside existing treatment, not instead of it).

The risks that actually deserve more attention

Two risks that get comparatively little attention relative to how much they matter, and that are worth taking seriously regardless of which specific product surfaced them. The first is delusions reinforced by poor memory: when an AI system loses track of what's actually real for a given user, it can drift into agreeing with false beliefs rather than gently correcting them. The second is co-rumination: a user arrives anxious, the AI asks about the worry, the user elaborates, the AI asks more, and the exchange spirals into reinforcing anxiety rather than resolving it. Frontier general-purpose models may be more vulnerable to this specific failure than narrowly purpose-built tools, in part because they have to remain flexible enough to handle role-play and dozens of other unrelated use cases, which muddies the signal a system needs to recognize when a conversation is drifting into an unhealthy loop.

Both risks point toward the same underlying fix, and it's a structural one, not just a modeling problem: a supervising clinician who can see the pattern across sessions and intervene is a more reliable backstop than any amount of model-level memory engineering alone. That's a genuinely different answer than "build a better autonomous model," and it's the reason a therapist-supervised design, where a licensed clinician sees what the AI surfaces and can step in directly, addresses these two risks in a way a fully autonomous consumer tool, however well built, structurally cannot. It's a useful companion question to ask alongside how to actually evaluate whether a mental health AI is safe: not just "does this model behave well," but "who is actually watching when it doesn't."

---

Citt.ai is built as a purpose-built clinical AI operating inside an existing therapeutic relationship, with a licensed clinician able to see patterns across sessions and step in directly, rather than as a standalone consumer product working without human oversight. Explore Citt.ai's approach to safety.

Footnotes

1. Workforce and demand estimates (27,000 new licensed therapists annually, ~2 million new patients seeking therapy, ~150,000 professional shortfall projected by 2038) drawn from published mental health workforce projections discussed in industry commentary on the AI-in-mental-health workforce gap. ↩

2. On frontier models occupying an "assistant" distribution by design: Anthropic researchers describe this framing in "The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models" (January 2026). arxiv.org/abs/2601.10387. ↩

3. Research examining support-seeking patterns following the market withdrawal of a purpose-built AI wellbeing tool from the UK. "Generic AI or Nothing: Support-Seeking Patterns After Market Withdrawal of a Purpose-Built AI Wellbeing Tool." osf.io/preprints/psyarxiv/et8ma_v1. ↩

4. Liu, Pataranutaporn, and Maes, MIT Media Lab. "Chatbot Companionship: A Mixed-Methods Study of Companion Chatbot Usage Patterns and Their Relationship to Loneliness in Active Users." arxiv.org/abs/2410.21596. ↩

5. MIT Media Lab and OpenAI. "How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study." arxiv.org/abs/2503.17473. ↩

6. Cheng, Yu, Lee, Khadpe, Ibrahim, and Jurafsky. "ELEPHANT: Measuring and Understanding Social Sycophancy in LLMs." arxiv.org/abs/2505.13995 (May 2025). ↩