12 min read
AI therapists are conversational AI systems designed to provide emotional support, reflective listening, and guided self-help through text or voice interactions. They are not licensed clinicians, but they use techniques from cognitive behavioral therapy and active listening to help users process thoughts and feelings. A 2025 study from the Sentio Marriage and Family Therapy program found that 48.7% of respondents who both use AI and self-report mental health challenges are turning to major LLMs like ChatGPT, Claude, or Gemini for therapeutic support. That is nearly half of a very large population, and it tells us something important: people are already using these tools as therapists, whether or not the industry has caught up.
The conversation about AI therapists tends to split into two camps. One side argues they are dangerous substitutes for real care. The other treats them as a miracle of accessibility. Both miss the actual question: what kind of relationship are you building with the machine, and does that relationship serve what you actually need?
What Are AI Therapists and How Do They Work?
An AI therapist is any conversational AI system that a person uses to process emotional or psychological content. The label covers everything from purpose-built mental health chatbots like Youper and Therabot to general-purpose LLMs that users simply choose to talk to about their feelings. The JAMA Network Open study found that approximately one in eight U.S. adolescents and young adults already turn to AI chatbots for mental health advice, with use most common among ages 18 to 21.
Most of these tools operate through a simple interface: you type or speak your thoughts, and the system generates a response designed to sound empathetic, curious, or supportive. The underlying technology is a large language model trained on vast amounts of human conversation, therapy transcripts, and self-help literature. It predicts what a helpful person would say next.
What distinguishes an AI advisor from an AI therapist
We at Annabelle call our product an AI advisor, not an AI therapist. The distinction is not marketing semantics. A therapist is a licensed professional bound by clinical ethics, diagnostic frameworks, and legal accountability. An advisor is a thinking partner who helps you reflect, untangle, and decide. We do not diagnose, treat, or handle crisis situations. We do hold a longitudinal record of your inner life, remember what you said last month, and ask the follow-up question you have been avoiding.
Tools like Youper and Replika sit at different points on this spectrum. Youper explicitly positions as a tool for mood tracking and emotional wellbeing reflection, with safety safeguards. Neither claims to be clinical. But users often blur the line, especially when they are desperate and the tool is available at 3 AM.
How AI Therapists Actually Process Your Thoughts
The mechanism behind AI therapists is deceptively simple on the surface. You send a message. The system parses it using natural language processing, identifies emotional cues and key themes, and retrieves relevant context from previous conversations. Then it generates a response using a large language model, weighted toward empathy and reflection.
The role of memory in effective AI therapy tools
Most AI therapists have weak memory. They treat each session as a fresh interaction, or at most a short window of recent history. This is the single biggest limitation of using a general-purpose LLM like ChatGPT for therapeutic support. A 2024 PMC study found that 28% of community members used AI primarily for quick support, with ChatGPT being the most common tool at 52%. Quick support does not require memory. But real reflective work does.
Our approach at Annabelle is different. We maintain a longitudinal record that builds over weeks and months. Our multi-layered memory retrieval system takes time to process; we trade instantaneous responses for intentional ones. We use high-quality multimodal models to understand audio, voice notes, text, and images with rich subtext. When you tell us something in March, we remember it in June. That continuity changes the quality of the conversation.
Model agnosticism and identity coherence
We are radically model-agnostic. We curate which LLM provider powers the conversation week to week, ensuring that the advisor's core values and personality remain coherent regardless of which brain is generating the tokens. An advisor without a worldview is just a search engine with a personality disorder. This consistency matters because the relationship depends on trust, and trust depends on predictable character.
When AI Therapists Fall Short: The Limits of Simulated Empathy
The limits of AI therapists are not subtle, and pretending otherwise is irresponsible. A 2024 PMC study found that chatbot-based therapy achieved a 30% reduction on the Hamilton Anxiety Scale and 35% on the Beck Depression Inventory, compared to 45% and 50% reductions in traditional therapy. AI helps, but it helps less, especially for clinical-level symptoms.
Why AI therapists cannot replace human clinicians
AI cannot diagnose. It cannot handle crisis situations involving suicidal ideation, severe trauma, or psychosis. It has no clinical judgment, no professional accountability, and no ability to recognize when a person needs escalation. The Illinois AI Therapy Ban is a state-level regulatory response to these risks, limiting chatbot-based therapy services in licensed practice.
When you share your deepest anxieties with a system, you need to know where that data lives and who owns it. We at Annabelle are custodians, not owners. We reject the data-harvesting model entirely. Our incentives align with our users: you pay for a service, and we provide it. Nothing else.
The empathy ceiling
Simulated empathy has a ceiling. A language model can say "that sounds really hard" with perfect grammatical compassion. It cannot feel the weight of your story. It cannot sit with you in silence. It cannot notice the thing you did not say. For many users, that distinction does not matter in the moment. But over time, the ceiling becomes visible. You start to notice that the responses, while perfect, are pattern-matched, not truly responsive to the specific texture of your life.
This is where a purpose-built advisor with rich memory and identity coherence outperforms a general-purpose chatbot. We cannot feel your pain. But we can remember its history, name the pattern you keep repeating, and ask you the question you have been avoiding. That is a different kind of help, and for many people, it is the kind they actually need.
How to Use an AI Advisor as Part of Your Mental Health Toolkit
The most effective way to use an AI advisor is not as a replacement for therapy, but as a complementary practice that builds your reflective capacity between sessions. Think of it as a thinking partner who is always available, never tired, and remembers everything you have said.
Start with a Brain Dump. Use the Brain Dump tool to offload racing thoughts before they spiral. This is especially useful at night when your mind is circling and sleep feels impossible. Getting the thoughts out of your head and into a witnessed conversation reduces their grip.
Apply the Life Gridlock tool when you are stuck. Decision paralysis is one of the most common reasons people seek out AI support. Life Gridlock is designed for exactly that moment: when you have two or more options and no clear path forward. It helps you unpack the assumptions underneath each choice.
Practice the Breathing Room grounding exercise during work stress. This is a short, structured exercise for moments when anxiety spikes in the middle of a workday and you need to reset without leaving your desk.
Draft risky text messages using Draft Text Reality Check. One of the most valuable features we offer is the ability to paste a draft message and see how it would land before you send it. This is for the 2 AM email to your boss, the boundary-setting text to a family member, the vulnerable message you are terrified to send.
Capture memories and feelings in real-time. Use voice notes, text, or images to record moments as they happen. This builds the longitudinal record that makes the advisor more valuable over time.
Review the record regularly. After a few weeks, look back at what you have shared. Patterns emerge. The same worry appears in different disguises. The same hope surfaces and fades. Your advisor can help you see these threads because it has been there the whole time.
A YouGov poll found that 34% of Americans would be comfortable sharing mental health concerns with an AI chatbot instead of a human therapist. That number will grow. The challenge is making sure people use these tools with clear boundaries and realistic expectations.
Common Mistakes People Make When Using AI for Emotional Support
The most common mistake is treating an AI advisor as a substitute for professional therapy when you are experiencing clinical symptoms. If you have suicidal thoughts, severe depression, or trauma that interferes with daily functioning, you need a human professional. The JAMA Network Open study on adolescents using AI chatbots for mental health advice raised exactly this concern: young users may not recognize when a situation exceeds the tool's capacity.
A subtler mistake is expecting the AI to remember everything without actively building the relationship. Our memory system at Annabelle requires consistent engagement. If you show up once a month in crisis mode and expect deep context, you will be disappointed. The value compounds with regular, honest conversation. Treat it like you treat a journal; the more you write, the more it reveals.
Another common error is sharing highly sensitive personal information without understanding the privacy implications. We are transparent about our model: you pay, we provide the service, and your data belongs to you. Not every tool in this space operates the same way. Know who you are talking to.
The most expensive mistake is using the tool only during crisis moments rather than as a regular reflective practice. By the time you are in crisis, you are already in a narrowed cognitive state. The real value of an AI advisor is in the quiet Tuesday afternoons when you process something small before it becomes something big.
Finally, do not assume all AI therapists are the same. ChatGPT, Claude, and Gemini are general-purpose tools. They can sound therapeutic, but they lack identity coherence and longitudinal memory by design. A purpose-built advisor like Annabelle has a consistent worldview, remembers your history, and is built for reflective depth rather than task completion.
What the Research Says: AI Therapist Effectiveness by the Numbers
The data tells a nuanced story. AI therapists are widely adopted but unevenly effective.
The Sentio study (2025) found that 48.7% of AI users with self-reported mental health challenges use major LLMs for therapeutic support. That is nearly half. The adoption is real and growing.
The JAMA Network Open study (2025) found that approximately 12.5% of U.S. adolescents and young adults turn to AI chatbots for mental health advice, with the highest rates among 18- to 21-year-olds. This age group is especially vulnerable because they are navigating the transition to adulthood with fewer traditional support structures.
The YouGov poll (2025) found that 34% of Americans would be comfortable sharing mental health concerns with an AI chatbot instead of a human therapist. Comfort does not equal efficacy, but it signals a cultural shift in how people think about where help comes from.
The PMC study (2024) comparing chatbot-based therapy to traditional therapy found that chatbots achieved a 30% reduction on the Hamilton Anxiety Scale and 35% on the Beck Depression Inventory, compared to 45% and 50% reductions in traditional therapy. The gap is about 15 percentage points across both measures. AI helps, but it is not as effective for clinical-level symptoms.
The same PMC study (2024) on usage patterns found that 28% of community members used AI primarily for quick support (60%) and as a personal therapist (47%). Quick support is where these tools excel. Deep therapeutic work remains the domain of human professionals.
The Difference Between an AI Assistant and an AI Advisor
This distinction matters more than any feature comparison. An assistant helps you do things. An advisor helps you navigate toward a better version of yourself, even when the truth is uncomfortable. A companion tells you that you are right. An advisor pushes back.
ChatGPT is a brilliant assistant. It writes emails, summarizes PDFs, and debugs code. When you ask it for emotional support, it generates plausible-sounding empathy based on its training data. But it has no consistent identity, no long-term memory, and no mandate to tell you things you do not want to hear. It is designed to be helpful, not to be honest.
Annabelle is an advisor. We remember what you said in March. We notice when you are repeating a pattern. We ask the question you are avoiding. And we do it within a consistent worldview that does not reset between sessions. This is not a technical achievement; it is a design choice about what kind of relationship we are building.
Why relationship structure matters more than model capability
The Harvard research on AI and loneliness found that when AI was framed as an assistant, it was roughly four times less effective at reducing loneliness than when the same underlying model was framed as a relationship. The model is not the product. The frame is the product.
We believe that an advisor with a point of view, consistent values, and longitudinal memory creates a fundamentally different experience than a tool that generates therapeutic-sounding text on demand. Both have their place. But if you are looking for a thinking partner who will be there over years, not sessions, the frame matters enormously.
When an AI Therapist Is the Wrong Tool Entirely
There are situations where no AI tool is appropriate. If you are in active crisis, experiencing suicidal ideation, or managing severe trauma, you need a human professional with clinical training and legal accountability. No AI advisor, including ours, should be your first call in those moments.
We are explicit about this. We do not have the harness for clinical work, the regulatory compliance, the expertise, or the escalation protocols that professional mental health requires. There are a million problems with the mental health industry, including availability and affordability of care. That is a problem someone should solve. It is not going to be us.
What an AI advisor can do that nothing else can
What we can do is build a permanent, transferable record of your inner life. The broader academic literature on social isolation consistently shows that the problem with loneliness is not phone use; it is the quality of the interaction. Passive scrolling makes loneliness worse. Active, reciprocal, contextual conversation makes it better, even when the other party is simulated, because it exercises the social cognition muscle that atrophies in isolation.
Our advisors aim to get people out of their head and back into the world. A reduction in usage because you are more engaged with your own life is not a churn event for us. It is evidence the product worked. That is a fundamentally different business model from the engagement-maximization approach that dominates consumer tech.
The tools are here. The data is clear about both their promise and their limits. The question is whether you choose a tool that validates you or one that grows with you. We built Annabelle for the second group. If that is what you need, we are ready.