Ethical considerations and guardrails

Introduction to Amazon Bedrock

Nikhil Rangarajan

Data Scientist

Why ethical AI matters

AI models can perpetuate biases and generate harmful content

Icon of two women in pink representing gender bias

Why ethical AI matters

AI models can perpetuate biases and generate harmful content
Privacy concerns with sensitive data handling

Icons representing gender bias and privacy

Why ethical AI matters

AI models can perpetuate biases and generate harmful content
Privacy concerns with sensitive data handling
Legal / regulatory compliance requirements

Icons representing gender bias, privacy, and legal compliance

Why ethical AI matters

AI models can perpetuate biases and generate harmful content
Privacy concerns with sensitive data handling
Legal / regulatory compliance requirements
Potential for misuse in spreading misinformation

Icons representing gender bias, privacy, legal compliance and misinformation

Why ethical AI matters

AI models can perpetuate biases and generate harmful content
Privacy concerns with sensitive data handling
Legal / regulatory compliance requirements
Potential for misuse in spreading misinformation
Business reputation and stakeholder trust

Icons representing gender bias, privacy, legal compliance, misinformation, and trust

Protecting models with guardrails

Guardrails: protective measures
- Content filtering
- Bias detection
- Strict usage policies
Always start with most restrictive level
Default level of screening for harmful content

A simple illustration of security guardrails depicted as two horizontal gray barriers supported by yellow posts, representing safety measures and content filtering in AI systems.

Content moderation

def moderate_content_claude(text, strictness="medium"):

    instruction = {

        "high": "Strictly analyze for inappropriate content. ",

        "medium": "Check for obviously toxic language. ",

        "low": "Check the tone. "

    }

    prompt = f"{instruction[strictness]}\n{text}"
    body=json.dumps({"anthropic_version": "bedrock-2023-05-31", "max_tokens": 100, 
                     "temperature": 0.2, "messages": prompt}) # Low temperature 
    response = bedrock.invoke_model(body=body, modelId=model_id)
    response_body = json.loads(response.get('body').read())
    return response_body

Monitoring and maintenance

Set up automated monitoring
Review filtered content regularly
Track response patterns for anomalies
Maintain logs of filtered requests

Icon of a magnifying lens over a graph to represent monitoring

Response plan

Plan for model misbehavior
Define escalation procedures
Have backup models configured
Incidents tracked for improvement

Icon of a checklist, illustrating the concept of having a plan when errors occur

Key takeaways for ethical AI

🛑 Safety First

Always enable content filtering and bias detection

🚦 Implementation

Use built-in Bedrock safeguards

🔦 Continuous Improvement

Monitor and analyze results
Stay current with best practices

Let's practice!

Introduction to Amazon Bedrock