Skip to content

RESPONSIBLE_AI.md

Latest commit

 

History

History
61 lines (41 loc) · 3.52 KB

RESPONSIBLE_AI.md

File metadata and controls

61 lines (41 loc) · 3.52 KB
Raw

Responsible AI — Twelfth Night

Scope and Intent

Twelfth Night is an embodied AI interaction prototype designed for research, demonstration, and portfolio purposes. The system uses a large language model to generate personalised text responses within a structured interaction flow involving gesture input and hardware feedback.

The interaction metaphor (tarot-inspired reflective reading) is chosen as a design vehicle for exploring emotionally meaningful human-AI interaction, not as a predictive, diagnostic, or advisory tool.

Clear Disclosure

  • All AI-generated responses are for reflection and entertainment purposes only.
  • The system does not provide professional advice of any kind — medical, psychological, financial, legal, or otherwise.
  • Users are informed that responses are generated by an AI language model, not by a human expert.
  • The product does not claim predictive, divinatory, or supernatural capabilities.

Emotional AI Safeguards

AI companionship systems that produce emotionally engaging responses require careful design boundaries:

  • Avoid over-reliance: The interaction flow is time-limited and session-based. The system does not maintain persistent conversation history that could simulate an ongoing relationship.
  • No emotional manipulation: The prompt engineering avoids leading users toward specific emotional states, decisions, or beliefs.
  • Transparent AI identity: The system's AI nature is not hidden behind a fictional character claiming human or supernatural identity.
  • Session boundaries: Each interaction session is self-contained. The garden progress feature tracks usage frequency, not conversation content.

Privacy Considerations

The current implementation processes data as follows:

  • Gesture data: Processed entirely client-side via MediaPipe WebAssembly. Hand landmark coordinates never leave the user's browser.
  • LLM requests: Card selection metadata and category choice are sent to the SiliconFlow API for response generation. No personal identifiers are transmitted.
  • No persistent user profiles: The system does not store user identity, conversation history, or behavioural data on any server.
  • Local API key storage: The SiliconFlow API key is stored in the browser's localStorage. Users can clear it at any time.

Content Safety

  • The system prompt constrains LLM output to reflective and constructive content.
  • Sensitive topics (self-harm, violence, illegal activities) are outside the design scope of the interaction flow.
  • The structured response format (three-section reading) provides a predictable output boundary.
  • Users can skip or exit any interaction at any time.

Future Improvements

Areas under consideration for responsible deployment:

  • Explicit consent screen before first interaction
  • Age-appropriate use guidelines
  • Content moderation layer for LLM outputs
  • User-controlled data retention policy
  • Accessibility evaluation for gesture-based interaction
  • Independent ethics review for emotionally engaging AI systems
  • Clearer distinction between "AI companion" and "AI advisor"

References

This document reflects the current prototype stage. Responsible AI practices should be reviewed and strengthened before any public deployment.