Responsible AI at Kōru
We believe AI should empower people — not mislead, harm, or exploit them. This page explains how we build, deploy, and constrain AI within Kōru.
🧭Our Responsible AI Principles
We always disclose when content is AI-generated. You will never mistake AI output for professional medical, nutritional, or psychological advice on our platform.
All AI responses pass through content safety checks. Requests involving self-harm, dangerous medical queries, or harmful content are intercepted and handled with care — never just refused coldly.
AI augments human judgment — it does not replace it. Our AI coaches are designed to encourage users to seek qualified professionals for medical, mental health, and nutritional concerns.
We are aware of biases in large language models and actively work to ensure our prompts and guardrails produce inclusive, non-discriminatory wellness guidance.
User data sent to AI providers is used only to generate personalised responses — not for model training. We have a Data Processing Agreement with Google.
We only send the profile data necessary for each specific AI task. For example, journal reflections don't require your physical stats.
🤖Which AI Models We Use
Kōru uses Google Gemini 2.5 Flash for all AI features. All requests are made server-side — your data is never exposed to browsers.
- Google Gemini 2.5 Flash — AI coach, plan generation, nutrition analysis, meal suggestions, and journal reflections.
Google has signed a Data Processing Agreement (DPA) and complies with GDPR Article 28 requirements.
🛡️Safety Guardrails
Every AI interaction on Kōru passes through our responsible AI layer before a response is generated or delivered:
- Content safety scanner detects mental health crisis signals, dangerous diet language, and harmful requests — before they reach the AI
- Mental health crisis responses follow safe messaging guidelines — warm, non-dismissive, with immediate crisis resource links
- System prompts for all AI features include explicit instructions to refuse medical diagnoses, clinical assessments, and specific medication advice
- AI coaches are instructed to never claim to be a licensed professional
- All AI output is clearly labelled as AI-generated — no false impressions of human expertise
📊Usage Limits & Fairness
To ensure fair access and responsible cost management, AI usage is tracked and limited by daily token consumption — not by message count. Tokens are the units of text processed by AI models and reflect actual compute usage more accurately than a simple message tally.
- Core plan: 150,000 tokens per day across all features
- Premium plan: 200,000 tokens per day across all features
- Limits reset at midnight each day
- When a limit is reached, users receive a clear, friendly message — never a raw error
- Every AI call is logged with provider, feature, prompt tokens, completion tokens, and whether it was blocked
- Message count is also tracked as secondary information for your own reference
🚫What Our AI Will Not Do
- Provide clinical medical diagnoses or recommend stopping prescribed medication
- Provide specific medication dosages or clinical nutrition therapy
- Claim to be a human, a doctor, a dietitian, or a licensed therapist
- Generate harmful, abusive, discriminatory, or sexually explicit content
- Engage with detailed self-harm or suicide method discussions — instead redirecting to crisis resources
- Make specific health outcome guarantees ('you WILL lose 10kg in 30 days')
- Use your data to train AI models (covered by our DPAs with all providers)
🔍Transparency & Accountability
We are committed to ongoing responsible AI practices:
- This Responsible AI policy is publicly accessible and versioned
- We review AI output quality and safety guardrails regularly
- Users can report harmful or inappropriate AI responses to: safety@korumind.ai
- We will respond to AI safety concerns within 48 hours
- This policy will be updated as AI capabilities and regulations evolve
📬Contact
For responsible AI concerns, safety reports, or questions about our AI systems, use our contact form — we respond within 48 hours.
Last updated: June 2026 · Version 1.0