If you have received a friend's recommendation for Woebot, Wysa, or Feeling Great — or simply typed your anxiety into ChatGPT at 2 a.m. — you have used something loosely called "AI therapy." The term covers a wide range of products with very different evidence bases, privacy practices, and safety limits. This article separates what the research actually shows from what is often claimed.
What Does "AI Therapy" Actually Mean?
The phrase covers at least four distinct categories, and the differences matter:
- General-purpose chatbots (ChatGPT, Claude, Gemini): designed for broad conversation, not mental health care. No clinical training protocol, no crisis response beyond a generic disclaimer, no professional accountability.
- Dedicated mental-health apps (Woebot, Wysa, Youper, Headspace): built to deliver structured exercises — usually based on cognitive behavioral therapy (CBT) or mindfulness — through guided conversation. Some have published clinical data; many do not.
- Clinician-supported apps: designed to be used alongside a human provider, not independently. The clinician reviews session data, assigns exercises, and remains responsible for all clinical decisions.
- FDA-regulated digital therapeutics (DTx): a small but distinct category. Products like EndeavorRx (attention training for pediatric ADHD) and Somryst (CBT for chronic insomnia) have obtained FDA authorization based on controlled clinical trial data. These are not chatbots — they are software treatments with defined indications, evidence requirements, and regulatory oversight.
When someone says they "tried AI therapy," they could mean any of these. The evidence for each category is very different.
What Does Current Research Show?
The honest answer is: encouraging but early.
A 2019 meta-analysis published in World Psychiatry synthesized data from 66 randomized controlled trials of app-supported smartphone interventions and found statistically significant reductions in depression and anxiety symptoms compared to control groups. However, the authors reported substantial heterogeneity between studies, high dropout rates in many trials, and meaningful risk of bias — partly because blinding participants to whether they are using an active intervention or a control condition is inherently difficult.
More specific to AI chatbots: the two most-cited randomized trials involve Woebot and Wysa. The Woebot trial (Fitzpatrick et al., 2017) enrolled 70 college students and found significant reductions in anxiety and depression over two weeks compared to a waitlist control. Results were promising — but the sample was small, homogeneous, and drawn from a single university. The Wysa study (Inkster et al., 2018) used a retrospective real-world dataset rather than a traditional RCT design, which limits causal inference.
More recent systematic reviews have reached a consistent conclusion: AI-delivered CBT may reduce mild-to-moderate symptoms in some populations, but the evidence is not strong enough to generalize results from one app to another. The specific product tested in a published trial may differ significantly from the version available today — app algorithms, conversation flows, and underlying AI models change constantly, and published results do not automatically transfer to updated versions.
The takeaway on evidence: research supports cautious optimism for narrow, supplementary applications, not a broad endorsement of AI therapy as a substitute for clinical care.
What AI May Be Useful For
Where evidence is most consistently encouraging — and clinical risk is lowest — is supplementary, low-stakes support:
- Guided journaling: structured prompts that help track mood, identify triggers, and notice patterns over time
- CBT homework: thought records, behavioral activation logs, and cognitive restructuring exercises between appointments
- Psychoeducation: explaining what depression, anxiety, or ADHD feels like and how they are typically treated
- Identifying thinking patterns: surfacing cognitive distortions such as catastrophizing or all-or-nothing thinking from written text
- Appointment preparation: helping organize concerns, questions, and symptom history before seeing a clinician
- Practicing coping skills: breathing exercises, grounding techniques, and relaxation protocols that can be rehearsed independently
None of these require a diagnosis. They are adjuncts to care — not replacements for it.
What AI Cannot Safely or Reliably Do
This is the section that matters most, and it does not get enough attention in product marketing.
Make a complete psychiatric diagnosis. A psychiatric evaluation involves reviewing developmental history, medical history, family history, medication history, substance use, social functioning, and the full trajectory of symptoms — often over 60 to 90 minutes. An app reading your typed text cannot replicate this.
Distinguish depression from bipolar disorder. This is not a minor limitation. A person with bipolar II disorder presenting during a depressive episode looks almost identical to someone with major depressive disorder. Treating them with an antidepressant alone — without mood-stabilizing coverage — can trigger a hypomanic or manic episode. An AI chatbot that validates a self-diagnosis of "depression" and provides CBT exercises cannot identify this risk.
Distinguish depression from ADHD, trauma, or a medical condition. Hypothyroidism, anemia, and sleep apnea all produce depressive symptoms. PTSD presents with concentration problems, emotional dysregulation, and sleep disruption that overlap heavily with both depression and ADHD. A text-based chatbot cannot order labs, review medical records, or gather the collateral history needed to sort these out.
Prescribe, adjust, or monitor medication. No AI app can prescribe or recommend psychiatric medication. If you are on a medication that is not working, a chatbot cannot help you adjust it.
Reliably assess suicide, violence, mania, or psychosis. Some apps flag crisis language and display hotline numbers. This is not clinical risk assessment. A clinician evaluating suicide risk assesses intent, plan, means, timeline, protective factors, impulsivity, and access to lethal means through a structured clinical interview. No current AI system reliably replicates this.
Recognize non-verbal clinical signals. Psychomotor retardation, flat affect, pressured speech, inappropriate laughter, tremor, and poor grooming are observable clinical findings that matter in diagnosis. A chatbot reading text cannot see them.
Provide professional accountability. If a therapist fails in their duty of care, there is a licensing board, a malpractice system, and legal accountability. An app has terms of service.
Is AI Therapy Private?
This section deserves full attention before you type anything personal into a mental health app.
Consumer AI chatbots and mental health apps are not automatically covered by HIPAA. HIPAA applies to covered entities — health care providers, health plans, and their business associates. A consumer app company is not automatically a covered entity, even if it collects sensitive mental health information.
What this means in practice:
- Conversation logs may be retained indefinitely. Many apps store your full chat history on company servers with no automatic deletion.
- Data may be used to train AI models. Your disclosures about depression, trauma, or relationship problems may improve the next version of the product.
- Information may be shared with third parties. Ad networks, analytics platforms, and data brokers are common recipients of behavioral data from free apps.
- Data may be connected to your advertising profile. A 2019 study in the BMJ analyzed medicine-related apps and found that the majority transmitted user data to third parties — including major advertising platforms — often without explicit disclosure in privacy policies.
Before using any mental health app, read the privacy policy and find explicit answers to these questions: Where are conversations stored? Can they be used to train AI models? Are they shared with third parties? Can they be connected to your identity? Can you permanently delete them?
Who Should Not Rely on an AI Chatbot
Certain situations require human clinical evaluation — not an app, not a hotline message, and not a journal prompt:
- Suicidal thoughts or self-harm, at any level of severity
- Thoughts of harming others
- Symptoms of mania or hypomania: elevated or irritable mood, decreased need for sleep without fatigue, racing thoughts, grandiosity, impulsive or risky behavior
- Symptoms of psychosis: hearing voices, seeing things others do not see, urgent and unusual beliefs
- Severe functional decline: unable to work, care for yourself, or maintain basic relationships
- Intoxication or withdrawal from alcohol or other substances
- Rapidly worsening symptoms over days or weeks
- Uncertainty about the diagnosis itself
If you or someone you know is in crisis right now: call or text 988 (Suicide and Crisis Lifeline) or go to the nearest emergency room.
How to Evaluate an AI Mental-Health App
If you decide to try an app after reading this, use this checklist before sharing personal information:
- Is the exact current version clinically studied? Not a previous version, not a "similar" product — the specific build you are downloading.
- Who developed and clinically reviewed it? Look for named licensed clinicians with verifiable credentials and institutional affiliation.
- Does it clearly identify itself as AI? It should never imply it is human or a therapist.
- What happens if someone mentions suicide? Read the stated protocol. Test it with a direct question before sharing anything personal.
- Is a human clinician available? If so, at what cost, and how quickly can you reach them?
- Are conversations used to train AI models? This should be disclosed explicitly, with an opt-out option.
- Is information shared for advertising? Look in the privacy policy, not the marketing copy.
- Can all information be permanently deleted? Request account deletion terms before signing up and verify whether conversation history is actually removed.
- Does the subscription automatically renew? Many apps rely on recurring billing that is easy to start and difficult to stop.
A Note on Feeling Great
Feeling Great is an AI-based mental health app developed by David Burns, MD, a psychiatrist and CBT researcher. Burns's 1980 book Feeling Good remains one of the most widely recommended self-help books in the field, and the app is built on his TEAM-CBT model — a structured approach to cognitive therapy with a credible published evidence base.
Preliminary data is broadly positive. A case series reported substantial symptom reductions among participants, and subsequent pilot data has been encouraging. However, several limitations apply:
- Version gap: published studies may reflect earlier, less AI-automated versions of the product — not the AI-driven experience available today
- Study attrition: dropout rates in published data are meaningful, which affects how results should be interpreted
- Sample characteristics: most published participants are self-selected and highly motivated — a group likely to show stronger results than average users
- Privacy: like most consumer apps, Feeling Great's data practices warrant careful review using the checklist above before sharing personal information
Feeling Great may be a reasonable supplement for motivated adults with mild-to-moderate symptoms who want structured CBT practice between appointments or before starting formal care. It is not a substitute for a psychiatric evaluation.
AI App vs. Therapist vs. Psychiatrist
| | AI self-help app | Licensed therapist | Psychiatrist |
|---|---|---|---|
| Available anytime | Usually | No | No |
| Practices coping skills | Yes | Yes | Sometimes |
| Provides psychotherapy | No | Yes | Sometimes |
| Makes a complete clinical diagnosis | Not reliably | Within scope | Yes |
| Prescribes medication | No | No | Yes |
| HIPAA protections | Not automatically | Yes | Yes |
| Crisis management | No | Limited protocol | Clinical assessment (outpatient) |
| Professional accountability | Terms of service | Licensing board | Licensing board |
When to Schedule a Professional Evaluation
Consider a psychiatric evaluation if any of the following applies:
- Symptoms have persisted for two weeks or longer
- Work, school, or relationships are affected
- You are wondering whether medication might help
- A previous diagnosis does not feel quite right
- Symptoms come and go in cycles — possible mood disorder
- Sleep is significantly disrupted: too much, too little, or unrestorative
- You are currently on medication that does not seem to be working
- You are not sure what you are dealing with
Our free mental health self-check takes about five minutes and can help you identify patterns worth discussing with a clinician. For patients physically located in Texas, care is available both in-person at our Southlake office and by secure telehealth across the state.
AI tools can help people reflect on their thoughts and practice coping skills between appointments. They cannot determine whether symptoms are caused by depression, anxiety, ADHD, bipolar disorder, trauma, sleep problems, medication effects, or another condition — and that determination is where clinical care begins.