Sonia is an AI mental health therapist delivering empathetic care, served through an iOS mobile app. Sonia is a conversational AI, meaning the interactions with her happen through either voice- or text-based chat conversations. The system is powered by the world’s most advanced Large Language Models (LLMs) developed by OpenAI. Team Sonia was co-founded by Chris Aeberli, Dustin Klebe, and Lukas Wolf, three Computer Scientists with strong academic backgrounds in AI research and alumni from the world’s most prestigious universities such as MIT and ETH. Their work was published at top-rated AI conferences such as NeurIPS, ICML, and EMNLP. Their technical skillset is complemented by the therapeutic knowledge of clinical advisor and psychiatrist Nick Nissen Harvard Medical School and David Singer a CBT therapist practising at Northwell with a PhD from Rutgers University.

Sonia is trained to implement Cognitive Behavioral Therapy (CBT), which is a psycho-social intervention that aims to reduce symptoms of various mental health conditions, primarily depression and anxiety disorders. At the core, Sonia is designed to solve two critical and identically important challenges when it comes to conversational AI used in therapy: flexibility and safety. Since an LLM powers Sonia, she can respond to almost every context a client presents, making her superior to all previously available rule-based chatbots. At the same time, we enforce the structure necessary to make the model align well with therapeutic goals and not leave the guardrails as often seen in general-purpose chatbots like ChatGPT.

Sonia’s creators employ an emergency check on every single client message. The emergency checks are weighted based on the semantics of the conversation and once it reaches a set threshold, the app stops and displays emergency hotlines and message services. Further details on the technical implementation of Sonia can be found in this blog.

Sonia does not store audio data and all chat transcripts undergo an anonymization process to remove all personally identifiable data. Nobody, neither Sonia's team members nor collaborating psychologists can read transcripts in clear text. No data is shared with third parties and is exclusively used to make Sonia memorize the client’s history as well as to improve future sessions.

Imagine you're chatting with an advanced digital assistant designed to help you manage stress or improve your thinking patterns, akin to a conversation with a therapist. This assistant is powered by some sophisticated software programs.

A Large Language Model is one and another is called a “State Machine.” - which has nothing to do with any government by the way.

Let's say you're expressing anxiety about an upcoming test. The software within this digital assistant uses a concept called a "State Machine" to navigate through different stages of the conversation based on what you've shared. Initially, it might start with a greeting, recognize your expression of stress, and subsequently offer strategies to alleviate that stress.

Cognitive Behavioral Therapy (CBT) is particularly well-suited to a Large Language Model and a State Machine system due to its structured, goal-oriented approach.

CBT is based on the idea that our thoughts, feelings, and behaviors are interconnected and that changing negative thought patterns can lead to changes in feelings and behaviors. This structured approach can be broken down into distinct stages or “states”, making it an ideal fit for a State Machine system. Here’s why:

Clear Stages: CBT follows a structured format that includes identifying problematic thoughts, challenging these thoughts, and adopting healthier thinking patterns. A State Machine can easily map these stages, guiding the user through each phase systematically.

Rule-Based Progression: The progression from one stage of therapy to the next in CBT is based on specific criteria being met, such as understanding the nature of one's distorted thinking or successfully challenging a number of negative thoughts. A State Machine excels at handling this kind of rule-based progression, moving from one state to another based on defined conditions.

Tailored Responses: CBT is personalized to address the specific concerns of the individual. A State Machine can adapt its path through the therapy stages based on user input, ensuring that the guidance provided is relevant to the user’s current needs and progress.

Feedback Loops: An essential part of CBT involves monitoring thoughts and behaviors and the outcomes of applying new strategies. A State Machine can incorporate feedback loops, allowing for adjustments based on the effectiveness of the strategies employed.

Scalability: CBT's structured nature allows for scalability in delivering therapy to more people. State Machines can manage multiple instances of therapy sessions simultaneously, offering personalized guidance to each user based on where they are in the CBT process.

In summary, CBT’s structured, step-by-step approach to improving mental health fits well with the organized, rule-based system of a State Machine. This compatibility allows for efficient, personalized, and scalable delivery of CBT principles in digital formats, making it accessible to a broader audience.

From Mayo Clinic Website

Overview

Cognitive behavioral therapy (CBT) is a common type of talk therapy (psychotherapy). You work with a mental health counselor (psychotherapist or therapist) in a structured way, attending a limited number of sessions. CBT helps you become aware of inaccurate or negative thinking so you can view challenging situations more clearly and respond to them in a more effective way.

CBT can be a very helpful tool — either alone or in combination with other therapies — in treating mental health disorders, such as depression, post-traumatic stress disorder (PTSD) or an eating disorder. But not everyone who benefits from CBT has a mental health condition. CBT can be an effective tool to help anyone learn how to better manage stressful life situations.

The concept of AI therapy first came up in 1964 when MIT professor Joseph Weizenbaum developed the computer program ELIZA. Modeled after a Rogerian psychotherapist, the chatbot was originally built as an antic to show just how superficial human-to-computer interaction was. To his surprise and confusion, people started spending hours talking to ELIZA, valuing the non-judgmental responses that offered a safe space for expressing their feelings.

Fast forward almost half a century, the rise of Large Language Models (LLMs) and the release of ChatGPT have brought the spotlight back to AI therapy. In the past 15 months since its release, many have reported turning to ChatGPT to talk about their emotions and struggles.

Mental Health Therapy as an LLM State Machine. Read this technical blog for a detailed explanation of how why Sonia is different and works.