For regulated industries, AWS’s neurosymbolic AI promises safe, explainable agent automation

Amazon Web Services (AWS) is making strides in the field of artificial intelligence with the introduction of its Automated Reasoning Checks feature on Bedrock, now available to a wider audience. This initiative aims to instill greater confidence in enterprises and regulated sectors as they explore the deployment of AI applications and agents. The company is optimistic that advancements such as automated reasoning—an approach grounded in mathematical validation—will facilitate a smoother transition into the realm of neurosymbolic AI. AWS believes this innovative method will set it apart in the competitive landscape of AI technology. Automated Reasoning Checks empower enterprise users to verify the accuracy of AI-generated responses and effectively identify instances of model hallucination. The feature was first showcased during AWS's annual re:Invent conference last December, where it was claimed to detect nearly all hallucinations in AI outputs. A select group of users had access through Amazon Bedrock Guardrails, enabling organizations to establish responsible AI governance. Byron Cook, a distinguished scientist and vice president at AWS’s Automated Reasoning Group, discussed the early success of this rollout in an interview with Venture Beat, highlighting that organizations have begun to appreciate the potential of combining symbolic reasoning with the generative capabilities of AI. "Neurosymbolic AI is an emerging concept that embodies automated reasoning," Cook explained. As organizations permitted AWS to analyze their data and the documents utilized to validate AI responses, it became clear that the tool's performance was comparable to that of humans equipped with a rulebook. Cook noted that the subjective nature of what constitutes 'truth' can often lead to debate, whereas automated reasoning provides a more definitive solution. The newly released Automated Reasoning Checks validate the correctness of AI outputs, offering reassurance to regulators and industries concerned about the unpredictable nature of generative AI. Cook emphasized that this approach substantiates many principles underpinning neurosymbolic AI, which merges the pattern recognition of neural networks with the logical structure of symbolic reasoning. While foundation models typically rely on neural networks, they can be susceptible to inaccuracies, raising concerns for enterprises. Conversely, symbolic AI tends to be less adaptable without explicit instructions. Influential figures in AI, like Gary Marcus, have highlighted the importance of neurosymbolic AI in achieving artificial general intelligence, and both Cook and AWS are enthusiastic about integrating these concepts into enterprise solutions. The potential applications of automated reasoning are vast, particularly in areas requiring precise validation, such as financial audits. For instance, if a model indicates that a report includes unapproved payments, automated reasoning can deconstruct the response into logical statements, confirming the model's accuracy based on predefined criteria. Cook envisions that as more enterprises gain access to these features through Bedrock, the practical benefits of automated reasoning will become evident. However, he also cautioned that these neurosymbolic AI methodologies are still in their infancy. "The impact on agentic AI is promising, though much of that work remains speculative at this stage," he remarked. Overall, AWS's commitment to enhancing the reliability of AI in regulated sectors marks a significant milestone in the ongoing evolution of artificial intelligence technology.