From the course: AWS Certified AI Practitioner (AIF-C01) Cert Prep

AWS responsible AI tools - Amazon Web Services (AWS) Tutorial

From the course: AWS Certified AI Practitioner (AIF-C01) Cert Prep

Start my 1-month free trial Buy for my team

AWS responsible AI tools

“

- I'd like to highlight one of the most important AWS tools for ensuring responsible use of AI, and it's called Amazon Bedrock Guardrails. And this provides customizable safeguards, as long as you're using Bedrock, for responsible AI applications. And it helps to enhance the safety protections by blocking harmful content and ensuring truthfulness. Some of the key features are content filters. So you can block harmful topics like hate speech and violence. It gives you custom topic control, so you can define restricted topics that should be avoided by the AI. And for different definitions of sensitive information based on different data privacy frameworks, you can instruct that that information should be redacted. It also supports hallucination detection, where it uses contextual grounding to prevent factually incorrect outputs. Now, some benefits of having done this should be pretty obvious. It enhances trust by filtering harmful content, it prevents privacy violations by redacting PII, and it improves the overall model accuracy by reducing hallucinations. And so let's take a look at a real scenario. A healthcare provider wants to implement a virtual healthcare assistant to provide medical advice and support to patients. The assistant will interact with patients to answer their questions, provide health tips, and schedule appointments. However, to ensure responsible and ethical AI use, and the healthcare provider needs to ensure that the assistant doesn't give incorrect advice, share sensitive information, or promote harmful content. And so one way that we could put this together in the context of a larger AWS infrastructure, we start with Bedrock. And this gives us access to foundation models to power this virtual healthcare assistant. And we can leverage those pre-trained models for natural language processing and response generation. We use Bedrock Guardrails to filter out the harmful content, redact the sense of information, and detect inaccuracies in the results. But we're going to do more than that. We can take advantage of Amazon Comprehend Medical, which can help with the the detection of medical-related sensitive information beyond what we may have missed with the Bedrock guardrails. And it ensures that any medical data is kept private and compliant. There's more we can use SageMaker. Custom training and tuning to allow the assistant to learn from specific healthcare data while maintaining those ethical safeguards. We can use IAM to manage roles and permissions so that only authorized users can modify or access to sensitive data and configuration. We can use Lambda. Lambda functions for triggering actions, such as checking and filtering content in real time as the assistant generates responses. And finally, we use CloudWatch to monitor the various metrics around the assistance performance, and usage patterns and logging, as well as the AI-generated responses for quality control. And we can set up alarms to alert teams and send messages to Slack or wherever to notify when inappropriate or harmful content has slipped through the guardrails.

Contents