6.1 AI Features in Keepr
| Feature | Technology | Risk Level |
|---|---|---|
| Speech-to-Text | On-device ML model | Low |
| Mood Detection | Sentiment analysis | Medium |
| Keyword Extraction | NLP | Low |
| Recommendations | Behavioural patterns | Medium |
6.2 Risk Assessment
- Speech-to-Text: risk of inaccuracy and privacy concerns; mitigated by on-device processing, user review and edit capability.
- Mood Detection: risk of misclassification and psychological impact; mitigated by user override, disclaimers and no automated actions.
- Keyword Extraction: risk of bias or misinterpretation; mitigated by on-device processing and user review.
- Recommendations: risk of manipulation or filter bubbles; mitigated by user control, transparency and opt-out.
6.3 Bias Mitigation
- Regular testing for bias in AI models.
- Diverse training data and user feedback mechanisms.
- Regular model updates and human review of edge cases.
6.4 Transparency & Explainability
- Clear labelling of AI-generated content.
- Explanation of how AI features work.
- User control over, and ability to disable, AI features.
- Feedback mechanisms for improvement.
6.5 Human Oversight
There are no automated decisions with legal or significant effects. Users can override AI recommendations, manual review is available for sensitive decisions, and clear escalation procedures exist.
6.6 Monitoring & Evaluation
We carry out regular performance monitoring, user feedback collection, bias testing and mitigation, incident reporting, and continuous improvement.