Enhancing Decision-Making in LLMs: Exploring Two Modern Approaches

Discover how companies are responsibly integrating AI into production at this invite-only event in San Francisco. Explore how technology and business intersect here – learn how you can attend by clicking here!

Emergent large language models (LLMs), including OpenAI’s ChatGPT (and particularly its latest iteration, GPT-4), Claude AI and Gemini have demonstrated limited decision-making ability. We will discuss current research surrounding LLM decision-making capabilities and what this may entail for their future development.

Traditional decision-making among LLMs involves discovering underlying patterns or rules and applying them appropriately in new scenarios to make decisions. An experiment by the Santa Fe Institute found that LLMs, including ChatGPT, could not “reason about basic core concepts.” To make well-reasoned decisions a deeper understanding of prompt context and output consequences is essential.

Poor legal, financial, and managerial decision-making often results in disastrous outcomes in practice. For instance, in 2023 the National Eating Disorder Association had to temporarily disable their AI chatbot called “Tessa,” after she dispersed insensitive advice such as weekly weigh-ins and eating at 500-1,000 calorie deficit. Following a storm of controversy this decision was quickly reversed and disabled.

LLMs don’t only give wrong information; they may also suggest generic solutions. INSEAD has observed that when ChatGPT was asked to respond to research questions about business strategies, its model tended towards generic advice on participative management such as collaborative working arrangements, creating an innovation culture, or aligning employees with organizational goals – rather than providing tailored strategies tailored specifically for a given organization’s situation. Business strategizing requires complex social and economic considerations which cannot be simplified with generic advice.