At the ongoing re:Invent conference, Amazon Web Services (AWS) unveiled a new service aimed at assisting businesses in minimizing occurrences of artificial intelligence (AI) hallucination. The Automated Reasoning checks tool, launched on Monday, is currently available in preview through the Amazon Bedrock Guardrails. AWS asserts that this tool mathematically validates the accuracy of outputs from large language models (LLMs), effectively addressing factual inaccuracies caused by hallucinations. Notably, its functionality resembles the Grounding with Google Search feature found on both the Gemini API and Google AI Studio.
AWS Automated Reasoning Checks
AI models frequently produce responses that can be incorrect, misleading, or entirely fictitious, a phenomenon referred to as AI hallucination. This challenge undermines the reliability of AI tools, particularly in enterprise applications. While organizations can somewhat alleviate the issue by training AI systems with high-quality internal data, inherent flaws in pre-training datasets and model architecture still lead to hallucinations.
AWS elaborated on its approach to counteracting AI hallucination in a blog post. The Automated Reasoning checks tool has been introduced as a precautionary measure, and is accessible in preview mode under the Amazon Bedrock Guardrails. The company explains that it employs “mathematical, logic-based algorithmic verification and reasoning processes” to validate information generated by LLMs.
The operation of this tool is quite straightforward. Users are required to upload relevant documents outlining their organizational rules to the Amazon Bedrock console. The Bedrock platform will then analyze these documents and formulate an initial Automated Reasoning policy, transforming the content from natural language into a mathematical structure.
After this, users can navigate to the Automated Reasoning menu located in the Safeguards section. Here, they can create a new policy and incorporate existing documents with essential information that the AI needs to learn. Additionally, users can manually define processing parameters and clarify the intent of the policy. It is also possible to add example questions and answers to enhance the AI’s understanding of typical interactions.
Upon completing these steps, the AI will be primed for deployment. The Automated Reasoning checks tool will then automatically verify any inaccurate responses generated by the chatbot. Currently, this tool is available in preview only in the US West (Oregon) AWS region, with plans for expansion to additional regions in the near future.
