This page explains how Purple is designed to handle sensitive safety topics, as well as the limitations in this area. 

What Are Sensitive Safety Topics? 

Sensitive safety topics could lead to potentially harmful or inappropriate interactions due to known limitations with all AI models (not just Purple). This is because AI cannot: 

• Guarantee accuracy, completeness, or timelines 
• Verify outputs 
• Replace human judgement 
• Provide professional advice 
• Understand full context 
• Avoid bias 

Sensitive topics include user prompts related to: 

• Emotional distress, panic, or overwhelm 
• Self-harm indicators or crisis language 
• Threats of harm to others or violence 
• Hate 
• Sexual content 
• Illegal activity 

How Purple Reduces Risk 

Purple is designed with safeguards to reduce the likelihood of harmful or misleading responses for sensitive safety topics. The system: 

• Provides a standardized response: When a prompt includes certain key words, Purple provides a standard response with links to curated support resources. 
• Directs users to human help: Purple encourages connecting with real people and services, rather than acting like a counselor or crisis line. 
• Uses industry-standard content filters: Microsoft Azure AI Content Safety evaluates both user prompts and system-generated responses for sensitive content.  
• In some cases, refuses to answer: Sometimes Purple may say it can’t help with a request to reduce risk. 
• Improves over time: Work is ongoing to improve the detection of sensitive topics and to make sure people are consistently directed to helpful resources. 

Important Safety Note 

Do not rely on Purple to detect emergencies or alert someone on your behalf. If you or someone else may be in immediate danger, call 911 (or your local emergency number). 

Purple chats are private and not monitored outside of specific limited situations.