6.1.7. Ethical and Governance Considerations

Beyond technical safeguards, generative AI requires thoughtful governance and ethical oversight.

Bias and Fairness in Generated Content: Generative AI can amplify biases from training data, creating content that:

• Perpetuates stereotypes
• Underrepresents certain groups
• Associates certain demographics with negative attributes
• Reflects historical biases in language and images

Addressing bias:

• Diverse and representative training data
• Evaluation across different demographic groups
• Human review of generated content for sensitive use cases
• Clear guidelines for model behavior
• Ongoing monitoring and adjustment

Privacy Considerations: Generative AI introduces unique privacy risks:

• Models may have memorized sensitive training data
• User prompts may contain confidential information
• Generated content may inadvertently reveal private details
• Prompt history could be exposed in security breaches

Privacy safeguards:

• Data encryption in transit and at rest
• Prompt data retention policies
• PII detection and filtering
• User consent and transparency about data use

Copyright and Intellectual Property: Generated content raises legal questions:

• Who owns AI-generated content?
• Can AI-generated content infringe on copyrights?
• What about content that closely resembles copyrighted material?

Best practices:

• Understand your organization's IP policies for AI content
• Review generated content before commercial use
• Be cautious with AI-generated images resembling real people or brands
• Document AI involvement in content creation

Deepfakes and Synthetic Media: Generative AI can create realistic fake images, audio, and video:

• Fake celebrity endorsements
• Manipulated political content
• Fraudulent identity documents
• Synthetic voices for scams

Mitigations for synthetic media risks:

• Content authenticity markers and watermarking
• Detection tools for AI-generated content
• User education about synthetic media
• Platform policies against deceptive content

Human Oversight Requirements: Responsible use of generative AI requires human involvement:

Red Teaming: Proactive testing for vulnerabilities:

• Simulate adversarial attacks on the system
• Test content filter effectiveness
• Identify potential misuse scenarios
• Document and address weaknesses

Transparency in AI-Generated Content: Users should know when content is AI-generated:

• Label AI-generated images and text
• Disclose AI involvement in customer interactions
• Provide opt-out for AI-generated responses
• Be clear about AI capabilities and limitations

⚠️ Exam Trap: Know the difference between layers: System messages (Layer 2) GUIDE behavior. Content filters (Layer 3) BLOCK harmful content. System messages tell the model what to do; content filters catch what slips through.