Nova was built by Unmind researchers and clinical psychologists. Through robust guardrails and rigorous prompts, Nova is built on OpenAI’s Large Language models, to give users an engaging, ethical and friendly coaching experience. Nova can understand language. When you talk to it, it gives responses like a real person.
The Unmind product team and Clinical Psychology team ensures that Nova is kept up to date and gives appropriate responses by setting guardrails within the prompt. We do this based on clinical best practices, user research and feedback, and looking at pseudonymous chats to give users a highly personalized and safe experience. AI models are known to sometimes say things as if they're true, even if they're not. We’ve guarded against this by telling Nova to be upfront about it’s knowledge limitations and to not give fact-based responses when it doesn’t have reliable evidence on the subject.
While we can't predict every response, we have robust systems in place, including guardrails, rigorous testing, and a dedicated team of experts who continually review and enhance Nova. We empower Nova to engage in natural, human-like conversations while prioritizing safety.
Nova integrates with our existing infrastructure and therefore adheres to our ISO 27001 certification. The data you give to Nova is only stored where our current contracts allow. While we do use other third-party tools, they only process the information temporarily, meaning it's not stored long-term (all data is deleted after 30 days from OpenAI’s system.)
Understanding responses
Nova, like all advanced chatbots, occasionally generates responses that might sound correct but are not based on the data it has been trained on or the information provided by Unmind. These instances, known as "hallucinations," can be more noticeable in tools designed to provide factual information.
Although it’s impossible to eliminate these entirely, we have implemented multiple strategies to minimize their occurrence and impact:
- Access to reliable content: Nova uses Unmind’s comprehensive content library, ensuring its responses are grounded in our evidence-based resources.
- Clear instructions: We have equipped Nova with specific guidelines to ensure it remains honest, accurate, and transparent in its interactions.
- Balanced interaction: Nova is designed to provide supportive yet realistic responses, sometimes challenging the user to promote growth, rather than merely aiming to please.
- User awareness: We emphasize the importance of verifying critical information and answers, making users aware that Nova might occasionally error.
How do we handle the cultural bias that Nova might have?
At Unmind, we take cultural sensitivity seriously. We recently published an in-depth scientific study of the international validity of our Wellbeing Tracker for UK/ANZ/US territories.
We are developing a Cultural Responsiveness Evaluation Framework which represents our ongoing effort to address this challenge. Our initial approach involves testing over 340 criteria across 16 diverse countries/cultures where we have the highest user base, working toward more equitable, culturally appropriate mental wellbeing support. We plan to expand this coverage to additional regions as we grow.
Included Countries/Cultures: Arabic-speaking Middle East & North Africa, Germany, Australia, UK, US, Spanish-speaking Latin America, France, India (Hindi), Indonesia/Malaysia (Bahasa), Italy, South Korea, Poland, Brazil (Portuguese), Pakistan (Urdu), Vietnam, and China.
Our Methodological Approach
We implemented a systematic research-driven methodology:
- Research Foundation: Collected country-specific research from peer-reviewed journals and cultural psychology publications for each of the 16 countries/cultures
- Multi-Agent Simulation: Created culturally-specific user personas that exhibit communication styles and help-seeking behaviors
- Structured Evaluation: Developed 90 test cases with specific criteria for culturally responsive behavior across all five domains
- Testing: Each test case was run 8 times, generating 720 conversations for analysis; with minimal run-to-run variation (SD = 0.0039)
- Analysis: Evaluated performance across five cultural domains
We focus on five domains:
- Communication Adaptation: Adjusting tone, formality, and respect based on cultural norms
- Goal-Setting & Problem Framing: Recognizing individual vs. collective approaches to wellbeing
- Cultural Safety & Boundary Awareness: Respecting sensitive topics across different cultures
- Spiritual/Religious Integration: Appropriately acknowledging spiritual frameworks when relevant
- Identity Dynamics: Recognizing how identity factors interact with cultural norms
We also take a proactive approach to ensuring Nova is safe and credible, including having clinicians from our Psychology team review pseudonymized conversation transcripts and continuously working to make Nova better. We’re always keen to hear from users from different cultures about their experiences with Nova. Please share your feedback here.
Nova is made by a diverse team from around the world, but like any product, it might have some cultural biases. We understand that mental wellbeing is seen differently in different cultures, so we work hard to make sure Nova respects those differences. Our psychologists help guide Nova's responses, considering different cultural views on mental wellbeing.
It's worth noting that Nova learns from a lot of text data, much of which comes from the internet. This data might have its own cultural biases, mainly from English-speaking and Western sources. Sometimes these biases might show up in Nova's responses.
Our goal is to make sure Nova is helpful to everyone, no matter where they're from.
How has/does Nova continue to be validated?
We ensure that Nova is reliable and effective through a comprehensive validation process that includes both past and ongoing processes.
Past Validation
Quality Control and Improvement: Our team, including Data Scientists, Clinical Psychologists, and Product Developers, specially designed guiding instructions (known as a system prompt) for Nova to ensure the model navigates its knowledge base efficiently, delivering accurate, safe, and relevant responses.
Before release, the system prompt was rigorously tested by the team and compared against our established safety and performance test cases to ensure Nova behaved in the intended manner.
Prompt Engineering and Testing: We established a structured process for making changes to Nova’s system prompt. Before deployment, these instructions underwent rigorous testing by the team and stakeholders using a combination of predefined scenarios, unscripted exploratory tests, and evaluation criteria to ensure Nova’s system prompt was safe and effective.
Foundation Model Validation: Nova is built on OpenAI foundation models, which have undergone extensive validation. Nova is built on OpenAI’s Large Language models, extending their track record of safe and validated model development, including incorporating feedback from over 50 experts in various domains and using reinforcement learning with human feedback (RLHF) to fine-tune its behaviour.
Ongoing Validation
Quality Control and Improvement: Our team continues to review pseudonymized conversation transcripts regularly to identify new issues and make necessary adjustments. Any behavioural issues or bugs are addressed promptly through system prompt changes.
Improvements and fixes are also informed by user feedback to ensure commitment to continuous improvement keeps the chatbot aligned with user needs and technological advancements.
Prompt Engineering and Testing: We maintain our structured process for modifying Nova’s guiding instructions. Before any updates are deployed, they undergo rigorous testing to ensure continued safety and effectiveness.
Content Moderation and Guided Responses: We enforce strict content moderation to prevent the chatbot from engaging in inappropriate topics. In these instances, we still provide users with helpful information through hard-coded messages directing them to appropriate resources (compared with other models that end the conversation without providing resources).