MOM's Bridge uses artificial intelligence responsibly. This page explains how AI is used, what safeguards are in place, and what limitations exist.

How We Use AI Prompts

Every AI interaction uses carefully designed prompts that include the user's pregnancy week for medical context. Prompts are translated into all 11 supported languages.

• Symptom Chat: The AI receives the pregnancy week and user's message, and responds with a brief supportive message and comfort measures in structured JSON format.
• Photo Analysis: The AI receives the pregnancy week, user's question, and uploaded photo. It analyzes visual content in the context of pregnancy health.
• Voice / Translation: The AI translates the user's spoken or typed message into Korean with pregnancy-specific medical vocabulary awareness.

AI models can reflect biases present in their training data. We take the following steps to mitigate bias in MOM's Bridge:

• Multilingual Equity: All 11 supported languages receive the same AI prompt structure and quality. Prompts are individually translated (not machine-translated from English) to preserve cultural nuance.
• Neutral Prompting: AI prompts do not include demographic data (age, ethnicity, nationality). The AI receives only the pregnancy week and symptom description to avoid discriminatory responses.
• Fallback Consistency: When AI is unavailable, keyword-based mock responses provide the same quality of care regardless of language. English fallback ensures no user sees broken content.
• Health Data Sources: Community statistics use CDC PRAMS national data, which represents diverse populations. We display population averages without making individual predictions based on demographics.

Known Limitations

• AI responses are informational only and should never replace professional medical advice
• AI may generate incorrect or misleading information (hallucination). Responses are constrained to brief formats to minimize this risk.
• Translations may not capture all medical nuances. The Korean doctor summary should be verified by a healthcare provider.
• Photo analysis cannot replace clinical examination. The AI may miss or misidentify visual symptoms.

Built-in Safety Mechanisms

• All AI responses are validated as structured JSON. Malformed responses are rejected rather than shown to users.
• Empathetic responses are truncated at 300 characters and comfort measures at 120 characters to prevent overly long or rambling outputs.
• If the AI API fails, the system falls back to a keyword-based response rather than showing an error, ensuring users always receive some guidance.
• Emergency keyword detection operates independently of AI. It uses a local keyword database in 6 languages and does not depend on API availability.
• All AI interactions are logged server-side for debugging and quality monitoring. Logs include response length, parsing success/failure, and error details.
• Daily tracking risk advisories use deterministic rules (kick count thresholds, ACOG 5-1-1 contraction pattern), never AI predictions, ensuring consistent and reliable safety guidance.

MOM's Bridge is designed to connect users to real healthcare resources, not replace them:

• Emergency Alert System: Detects emergency keywords in 6 languages and triggers GPS-located alerts with a 10-second countdown, connecting to real emergency contacts.
• Doctor-Ready Summaries: Every AI interaction produces a Korean-language clinical summary designed to be shown directly to a healthcare provider at the point of care.
• Evidence-Based Statistics: Community health comparisons use real CDC PRAMS data (nationally representative pregnancy statistics), not AI-generated numbers.
• Medical Disclaimers: Every chat page displays a prominent warning that AI is not a substitute for medical care. Users are consistently directed to consult healthcare providers.
• Daily Tracking (Kick Counter & Contraction Timer): Fetal movement counting and contraction timing use deterministic rule-based risk assessment (not AI), following ACOG clinical guidelines. Three safety disclaimers are always visible, and doctor summaries can be generated for provider visits.

Our AI system is evaluated and improved through the following process:

1. Prompts are tested across all 11 languages to ensure consistent JSON output structure and appropriate response length
2. Response truncation limits (50 words empathetic, 20 words comfort) were determined through iterative testing to balance helpfulness with safety
3. Mock responses serve as baseline comparison to ensure AI responses meet minimum quality standards
4. Server-side logging tracks AI response parsing success rates, token usage, and error patterns for ongoing monitoring
5. User data persistence (SymptomLog) enables retrospective review of AI interaction quality and identification of systematic issues