AWS has made its Automated Reasoning Checks feature available through its Bedrock platform, aiming to provide enterprises, particularly those in regulated industries, with greater confidence in using AI applications. This feature allows users to verify response accuracy and identify model hallucinations, which are the incorrect outputs often produced by AI systems.
Introduced during AWS’s re:Invent conference in December, the Automated Reasoning Checks can potentially detect nearly all hallucinations, as indicated by usability tests conducted with a select group of users. These users reported that the tool’s performance was comparable to that of a human expert equipped with relevant guidelines. Byron Cook, vice president at AWS’s Automated Reasoning Group, noted that the initial rollout demonstrated the system’s effectiveness in practical settings, particularly in merging symbolic logic with the capabilities of neural networks.
The recent updates to Automated Reasoning Checks include the ability to handle large documents, simplified policy validation, automated scenario generation, and natural language suggestions for feedback on policies. According to Cook, these enhancements enable users to confirm the model’s logic and correctness without repeated validations.
The underlying concept of neurosymbolic AI, which combines neural networks with symbolic reasoning, is highlighted as a significant advancement in AI technology. By utilizing mathematical proofs, Automated Reasoning Checks can address the concerns associated with the non-deterministic nature of generative AI, making it more appealing to enterprises that require reliable and accurate AI responses.
Cook mentioned that while the technology is still in its early stages, it has the potential to impact various applications in agentic AI, indicating a growing interest and investment in these innovative methodologies. However, he cautioned that the development of such systems remains speculative and still requires refinement and further investigation.
Source: https://venturebeat.com/ai/for-regulated-industries-awss-neurosymbolic-ai-promises-safe-explainable-agent-automation/