Tailoring Responsible AI with New Safeguard Tiers in Amazon Bedrock Guardrails
The introduction of safeguard tiers in Amazon Bedrock Guardrails marks a significant advancement in the way organizations can approach responsible AI. These tiers provide a framework for integrating safety and privacy measures across various foundation models (FMs), thereby empowering businesses to build trusted generative AI applications at scale.
Overview of Amazon Bedrock Guardrails
Amazon Bedrock Guardrails offers configurable safeguards that help prevent unwanted content while aligning AI interactions with an organization’s responsible AI policies. The system provides a model-agnostic approach through the standalone ApplyGuardrail API, which supports models hosted outside of Amazon Bedrock.
Key Safeguards
Guardrails currently offer six key safeguards:
- Content filters
- Denied topics
- Word filters
- Sensitive information filters
- Contextual grounding checks
- Automated Reasoning checks (preview)
Challenges in Implementing Responsible AI
As organizations strive to implement responsible AI practices, they face the challenge of balancing safety controls with varying performance requirements across different applications. A one-size-fits-all approach is often ineffective. To address this issue, Amazon has introduced safeguard tiers that allow organizations to choose appropriate safeguards based on specific needs.
Benefits of Safeguard Tiers
The introduction of safeguard tiers provides three key advantages:
- Control Over Guardrail Implementations: Organizations can select the appropriate protection level for each use case, allowing for tailored safety controls.
- Cross-Region Inference Support (CRIS): This feature enables the use of compute capacity across multiple regions, enhancing scalability and availability for guardrails.
- Advanced Capabilities: The tiers offer configurable options for use cases where robust protection or broader language support is critical, albeit with a modest increase in latency.
Understanding the Tiers
Safeguard tiers are applied at the guardrail policy level specifically for content filters and denied topics:
- Classic Tier (Default): Maintains existing behavior with limited language support (English, French, Spanish) and is optimized for lower-latency applications.
- Standard Tier: Offers multilingual support for over 60 languages, enhanced robustness against prompt attacks, and requires CRIS, with a potential increase in latency.
Organizations can select tiers independently for different policies, providing flexibility to implement the right level of protection for each application.
Quality Enhancements with the Standard Tier
Tests indicate that the new Standard tier improves harmful content filtering recall by over 15% and balanced accuracy by more than 7% when compared to the Classic tier. The multi-language support is particularly noteworthy, providing strong performance across 14 common languages.
Benefits for Different Use Cases
Different AI applications have distinct safety requirements. For instance:
- Customer-facing applications often require stronger protection against misuse.
- Global applications need guardrails that work effectively across many languages.
- Internal enterprise tools might prioritize specific topics in a few primary languages.
Configuring Safeguard Tiers
On the Amazon Bedrock console, organizations can configure the tiers for their guardrails in the Content filters tier or Denied topics tier sections. The use of the Standard tier necessitates setting up CRIS, allowing for optimal performance and availability.
Evaluating Guardrails
To thoroughly assess the performance of guardrails, organizations should consider creating a test dataset that includes:
- Safe examples: Content that should pass through guardrails.
- Harmful examples: Content that should be blocked.
- Edge cases: Content that tests the boundaries of policies.
- Multi-language examples: Especially important for the Standard tier.
Using a labeled dataset allows for accurate assessment of guardrails’ performance, helping organizations refine their AI applications.
Best Practices for Implementation
Organizations are encouraged to consider the following best practices when implementing the tiers:
- Start with staged testing: Test both tiers with representative samples.
- Consider language requirements: Evaluate the necessity of expanded language support.
- Balance safety and performance: Weigh accuracy improvements against potential latency increases.
- Use policy-level tier selection: Optimize your guardrails by choosing different tiers for different policies.
- Account for cross-region requirements: Ensure your architecture can accommodate CRIS.
Conclusion
The introduction of safeguard tiers in Amazon Bedrock Guardrails significantly enhances the ability of organizations to implement responsible AI. By providing flexible and evolving safety tools, businesses can develop AI solutions that are both innovative and ethical. The Standard tier, in particular, offers substantial improvements in multilingual support and detection accuracy, making it ideal for applications serving diverse global audiences.
With the customizable protection levels offered by these tiers, organizations are better equipped to balance performance and safety, ensuring that their AI applications align with both organizational values and regulatory compliance.
