AI Can Simulate Therapy. It Still Can’t Replace the Relationship.

One of our early product conversations at Mentalyc started with a deceptively simple question: how do you actually know if therapy is working?

In most areas of medicine, progress reveals itself through numbers. Blood pressure drops, tumors shrink, lab values move in the right direction. Psychotherapy unfolds differently. A patient might sit in the same chair every week for months, talking about the same relationships, the same memories, the same patterns. From the outside it can be difficult to tell whether anything meaningful is changing.

During that time, I had the opportunity to speak with more than seventy therapists while working on products intended to support their practice. Many of those conversations were part of broader efforts to understand how technology might help clinicians track progress across sessions.

What stood out to me from those discussions, and from the research I was reading alongside them, was that therapists often describe progress in ways that do not easily translate into simple metrics.

Again and again, the same theme appeared in those conversations. The quality of the relationship between therapist and patient seemed to play a central role in whether therapy moved forward.

Psychologists call this the therapeutic alliance, the degree to which a patient feels understood, supported, and collaboratively engaged with their therapist. It is not the only signal of progress in therapy, but decades of research suggest it is one of the most reliable predictors of successful outcomes.

Looking at therapy through that lens gradually changed how I think about technology in mental health.

Many conversations about AI in this space focus on whether machines can simulate therapy. The deeper question may be whether technology can support the relationship that makes therapy effective in the first place.

The Relationship Beneath the Technique

In psychotherapy research, few findings are as consistent as the role of the therapeutic alliance. Across decades of studies, the strength of the relationship between therapist and patient repeatedly predicts treatment outcomes regardless of the therapeutic method being used.

A landmark meta-analysis by Christoph Flückiger, Adam Horvath, and colleagues synthesized results from hundreds of studies. They found that alliance reliably predicts improvement in psychotherapy independently of the techniques applied. Whether a therapist practices cognitive behavioral therapy, psychodynamic therapy, or another modality, the quality of the relationship still explains a meaningful portion of treatment outcomes.

This insight reshapes how we think about therapy itself. Psychological interventions are not delivered in isolation. They are carried through a relationship where the patient feels safe enough to explore difficult emotions and experiences.

Clinicians often describe breakthroughs in relational terms rather than technical ones. Moments when a patient finally feels understood. When trust deepens after weeks of hesitation. When a difficult insight becomes possible because the relationship can hold it.

Progress in therapy is influenced by many factors. Yet the strength of the therapeutic alliance consistently emerges as one of the most important signals.

The Seduction of Autonomous AI Therapy

At the same time, generative AI has introduced a new narrative into mental health technology. If language models can hold natural conversations, perhaps they can deliver therapy directly. AI chatbots could potentially support millions of people who currently lack access to care.

From a systems perspective the appeal is obvious. Mental health services remain scarce in many regions. Waiting lists continue to grow, and clinicians face increasing burnout. AI appears to offer a path toward scale that traditional care systems struggle to achieve.

Recent research suggests this possibility should not be dismissed entirely. A randomized controlled trial led by Heinz and colleagues evaluated a generative AI chatbot designed to support individuals experiencing depression, anxiety, and eating disorders. The study reported promising improvements in symptoms, and participants described experiencing a sense of connection with the system.

Some users even reported what researchers described as a form of therapeutic alliance with the AI. But the researchers issued an important warning. Systems like these should not be deployed autonomously without human oversight.

Once the importance of alliance is taken seriously, the challenge becomes clearer. The question is not simply whether AI can produce empathetic language. The question is whether technology can sustain the relational conditions that make therapy effective.

Why Mental Health Apps Struggle With Complexity

Over the past decade, thousands of digital mental health tools have attempted to replicate pieces of therapy. Some focus on mood tracking. Others offer cognitive behavioral exercises, guided journaling, or conversational support. Many of these tools show encouraging results for mild distress, particularly when users remain engaged.

But the evidence becomes less optimistic as problems become more complex. A meta-analysis of ninety-two randomized controlled trials found that mental health apps tend to produce small to moderate improvements for mild symptoms. They also face persistent challenges with engagement and retention.

Dropout rates remain high, and effectiveness often declines for individuals dealing with more complex mental health conditions.

From a product perspective this pattern is often framed as an engagement problem. Designers experiment with reminders, gamification strategies, and habit-forming mechanics that encourage users to keep returning to the app.

Another interpretation is possible. Many digital mental health tools struggle not because they lack features, but because they lack relationship.

A user may complete exercises or track moods. Yet they still feel fundamentally alone in the process.

The Fragile Nature of Digital Alliance

Humans are remarkably capable of forming emotional connections with conversational systems. When an interface responds with language that appears attentive and supportive, people often project understanding and empathy onto the interaction.

A qualitative study by Brotherdale and colleagues explored how users experience therapeutic alliance with fully automated mental health apps. Participants reported moments where the system felt supportive and responsive, creating a temporary sense of relational connection.

But these connections proved fragile. When responses felt repetitive, generic, or slightly misaligned with the user's experience, trust eroded quickly. Participants described the lingering feeling that the system did not truly grasp the depth or nuance of what they were going through.

Humans are highly sensitive to signals of genuine understanding. A subtle misinterpretation or a formulaic response can quickly disrupt the sense that someone, or something, is truly listening.

What Human Support Changes

Another body of research highlights the importance of relational presence in digital mental health interventions. Programs that combine technology with human support consistently outperform those that rely entirely on automation.

A large meta-review analyzing more than five hundred studies found that digital interventions supported by humans produce stronger clinical outcomes than fully automated programs. Human guidance significantly increases both engagement and effectiveness.

Technology can structure the intervention through exercises, psychoeducation, and behavioral tools. Human support provides relational continuity. Someone remembers previous conversations, notices subtle shifts, and helps anchor the process over time.

The intervention may be digital. The transformation remains relational.

Designing AI That Strengthens the Alliance

Reflecting on my time working on AI systems for therapists at Mentalyc, my perspective on the role of technology in mental health gradually shifted.

Much of the public conversation focuses on whether AI might eventually replace therapists or conduct therapy sessions autonomously. But when you listen to clinicians describe their daily work, a different set of challenges becomes visible.

What drains therapists is rarely the therapeutic conversation itself. The emotional presence required in therapy is demanding, but it is also the core of why many clinicians entered the profession. What exhausts them is everything surrounding the session. Documentation requirements, session notes, administrative tasks, and the cognitive effort required to hold dozens of patient narratives while moving rapidly between emotionally intense conversations.

These invisible layers of work slowly erode the attention therapists can bring to the patient sitting in front of them.

This is where AI may offer its most meaningful contribution.

Systems that summarize sessions, surface patterns across conversations, and help clinicians track long term progress can reduce the friction around therapy rather than attempting to replace the therapist. When that surrounding workload decreases, clinicians regain the cognitive bandwidth required for presence, curiosity, and careful listening.

In that sense, the most promising role for AI in mental health may not be as a therapist. It may be infrastructure that helps protect the relationship at the center of therapy.

The Future of AI in Therapy

The vision of autonomous AI therapists will likely continue to attract attention. The shortage of mental health professionals is real, and technology will inevitably play a role in expanding access to care.

But decades of psychotherapy research point toward a consistent insight. The therapeutic alliance is not simply a supportive feature of therapy. It is one of the mechanisms through which therapy works.

That insight suggests a different direction for mental health technology. The most successful AI systems will likely be those that strengthen the relationship between therapist and patient rather than attempting to replace it.

They will help clinicians listen more carefully across time, remember patterns that unfold over months of conversation, and reduce the administrative friction that pulls attention away from the human encounter.

If the therapeutic alliance remains one of the strongest predictors of success in therapy, the design principle for AI in mental health becomes surprisingly clear.

AI should not replace the relationship.

It should help it deepen.

Sources and Further Reading

Brotherdale, E., et al. (2024). User experiences of therapeutic alliance with fully automated mental health apps: A qualitative study. Digital Health, 10. https://doi.org/10.1177/20552076241277712

Flückiger, C., Del Re, A. C., Wampold, B. E., & Horvath, A. O. (2018). The alliance in adult psychotherapy: A meta-analytic synthesis. Psychotherapy, 55(4), 316–340. https://doi.org/10.1037/pst0000172

Heinz, M. V., et al. (2025). Randomized trial of a generative AI chatbot for mental health treatment. NEJM AI. https://ai.nejm.org/doi/full/10.1056/AIoa2400802

Valentine, L., et al. (2025). A meta-analysis of persuasive design, engagement, and efficacy in 92 RCTs of mental health apps. npj Digital Medicine, 8(229). https://doi.org/10.1038/s41746-025-01567-5

Werntz, A., et al. (2023). Providing human support for the use of digital mental health interventions: Systematic meta-review. Journal of Medical Internet Research, 25, e42864. https://doi.org/10.2196/42864