AI Therapy Apps: What They Are, What They're Not, and When to Use Which

The app stores now contain dozens of products that use the word "therapy" in their descriptions, alongside AI, mental health, and emotional support. This language is doing a lot of work, not all of it honest, and the people most likely to download these apps are often the people most likely to be confused by the ambiguity. A clear-eyed account of what these tools actually offer — and what they do not — is more useful than either uncritical enthusiasm or blanket dismissal.

What AI Therapy Apps Actually Are

Most AI therapy apps are not therapy. The word is used because it is familiar and comforting, but genuine therapy involves a licensed clinician, a therapeutic relationship developed over time, legally protected confidentiality, clinical assessment and diagnosis, and professional accountability. Almost none of the AI apps on the market provide any of these. What most provide is some combination of: Conversational support using large language models — the user types or speaks, the AI responds with reflective, empathic, or gently directive language. Products like Woebot and Wysa fall broadly into this category. Structured psychoeducation — teaching CBT skills, mindfulness techniques, or DBT concepts through interactive modules. These are closer to digital workbooks than to therapy, but well-designed ones can be effective. Mood tracking and journaling with AI-generated feedback — the user logs emotions, thoughts, or behaviors, and the AI provides patterns, observations, or prompts. Some products combine these. Some offer the ability to escalate to a human counselor (licensed or not, depending on the platform). The range in quality and honesty about what is being offered is extremely wide.

What Research Actually Shows

The honest picture from research is mixed, which is different from negative. Woebot, one of the most studied AI mental health apps, has produced trial data suggesting meaningful reductions in anxiety and depression symptoms over two-week periods, using CBT-based protocols. Research published by teams at Stanford University found effect sizes comparable to some brief human-delivered interventions for subclinical anxiety. The researchers were careful to note that these were short-term outcomes in populations without severe symptoms. What is less established is whether these effects persist, whether they translate to people with moderate-to-severe conditions, and whether regular use produces meaningful change in functioning rather than just symptom ratings. A meta-analysis from University College London reviewing digital mental health interventions found the largest evidence base for depression and anxiety management in subclinical and mild populations, with weaker evidence for moderate presentations and very limited evidence for severe mental illness. This is not a criticism of the tools — it reflects where research has been done and where caution is warranted.

A Tangent: The Therapist Shortage Is Not Abstract

There are not enough licensed therapists to meet population-level mental health need. In the United States, roughly half of the counties have no practicing psychiatrist. Waitlists for mental health services in the UK run to months. The shortage is structural, persistent, and not resolving quickly. AI tools exist in this context, and their use needs to be understood partly as a response to a gap that is not being filled by traditional services. Dismissing AI mental health tools without reckoning with this gap is intellectually incomplete. The relevant question is not whether these tools are as good as therapy — they are not — but whether they are better than the alternative available to the specific person in front of the screen, which in many cases is nothing.

Who Benefits, Who Should Be Careful

AI therapy apps appear most appropriate and most evidenced for: People managing mild-to-moderate anxiety, stress, or low mood who are functioning well enough to engage with structured interventions. People who want skill-building — particularly CBT or mindfulness skills — and cannot access or afford a clinician. People for whom the low barrier to entry (available instantly, no judgment, anonymous) makes it more likely they will engage than with formal therapy. More caution is warranted for people with severe depression, active suicidal ideation, eating disorders, psychosis, trauma requiring careful titration, or complex presentations that require clinical assessment. For these populations, AI apps are at best supplementary and at worst a delay to necessary care.

The Honesty Test

The most important criterion for evaluating any AI mental health product is whether it is honest about what it is. Products that clearly describe themselves, acknowledge their limitations, have transparent data practices, and facilitate connection to human services when appropriate are operating in good faith. Products that use "therapy" as a marketing term without clinical substance, that make outcome claims not supported by evidence, or that fail to address safety protocols for users in crisis are not. The category is real. Its quality is variable. Reading the label matters.