AWS, Anthropic, and DTCC discuss responsible AI and the opportunities and risks of generative AI

A summary of the panel discussion at the AWS Financial Services Symposium

Written with the help of Anthropic Claude 3 Sonnet on Amazon Bedrock

The AWS Financial Services Symposium took place on June 6 in New York City, where several hundred AWS customers and partners in the financial services industry convened to explore how technology is driving innovation. The theme of the event was “rethink intelligence,” and attendees explored how data and AI, and especially generative AI, are enabling organizations to extract game-changing business insights, automate complex processes, and create hyper-personalized customer experiences.

With the promise of generative AI, there also come new challenges. To explore this topic, Michael Kearns, Amazon Scholar, AI/ML, AWS, and Professor, University of Pennsylvania; Johnna Powell, Managing Director, Technology Research and Innovation, The Depository Trust & Clearing Corporation (DTCC); and Michael Gerstenhaber, VP of Product, Anthropic, took to the stage to engage in a discussion about their experiences and learnings from their work with AI over the years. The panel discussion provided an insightful overview of the current state and future prospects of responsible AI (RAI) in foundation models and generative AI. The key topics and insights covered included:

• The evolution of machine learning and AI: Michael Kearns traced the historical development of machine learning, from its early days as an obscure sub-field of AI to the recent advancements driven by the availability of large datasets and computing power, leading to the rise of foundation models and generative AI. Michael contrasted the highly targeted applications of machine learning five years ago, e.g. to predict probability of default on a consumer lending application, to the “extremely general purpose foundation models like large language models and image diffusion models” we have today and the concerns that emerge from that generality.
• RAI concerns: The discussion highlighted the emergence of RAI as a critical consideration, driven by issues such as demographic bias, privacy leaks, and the potential for hallucinations and intellectual property concerns with generative AI. The panelists emphasized the need to address these concerns proactively rather than in a reactive, post-hoc manner.
• DTCC’s AI strategy and use cases: Johnna shared DTCC’s approach to leveraging generative AI, focusing on use cases such as developer productivity and legacy code modernization, with tools like Amazon Q Developer, and client insights. She emphasized the importance of establishing robust governance structures, policies, and risk management processes to ensure the responsible deployment of these technologies.
• Anthropic’s approach to RAI: Michael Gerstenhaber discussed Anthropic’s efforts to build foundation models with RAI considerations in mind from the outset. This includes techniques like constitutional AI, which allows for the generation of synthetic training data aligned with high-level principles and guidelines, as well as providing tools and services to help customers monitor and evaluate the outputs of these models.
• Regulatory landscape and challenges: The panelists explored the evolving regulatory landscape around AI, highlighting the need for a balance between providing clarity and guidance while avoiding overly restrictive measures that could stifle innovation. They noted the importance of ongoing dialogue between industry, policymakers, and regulators to navigate this complex terrain.
• Balancing fairness and performance: The discussion delved into the technical challenges of achieving demographic fairness in machine learning models, particularly in highly general foundation models. The panelists acknowledged the trade-offs between fairness and model performance and the need to explore new approaches that can better capture and enforce fairness constraints during the training process.

Overall, the panel provided a comprehensive overview of the current state of RAI, highlighting the efforts of both industry leaders and financial institutions to harness the power of generative AI while addressing the associated risks and challenges. The discussion underscored the importance of a multifaceted approach encompassing technical innovations, robust governance frameworks, and collaborative engagement with regulators and policymakers to ensure the responsible development and deployment of these transformative technologies.

To watch the panel recording, click here.