Using Knowledge Elicitation Techniques To Infuse Deep Expertise And Best Practices Into Gen AI

Quick Overview

The main challenge in developing expert AI systems is moving beyond the general knowledge of foundation models by infusing deep, often unwritten, human expertise through knowledge elicitation techniques like interviewing experts and analyzing their historical decision-making processes.

Key Points: The primary challenge in AI development is moving beyond general LLMs to create specialized experts, often requiring deep, unwritten knowledge. The methodology involves layering three techniques: Human-to-Human (H2H) interviews, Human-to-AI (H2AI) elicitation, and Data Analysis for rule validation. The H2H step involves interviewing experts (like a stock trader) to extract tacit knowledge, including gut feelings and unwritten rules, which are then structured or coded. The 'Momentum Rule' example from the study requires a stock to drop 8% below purchase price and the expert trader to avoid forcing a stop-loss if the market sentiment is positive. The final goal is to create a 'Synthetic Expert' AI whose rules are validated against both explicit data and tacit expert knowledge, avoiding the pitfalls of pure data reliance. The combination of structured rules (like the 8% stop-loss) and nuanced, elicited knowledge prevents the AI from generating 'fake rationalizations' that sound logical but are contextually wrong.

Context: This AI Papers Daily podcast episode discusses the major challenge facing current AI development: how to transition powerful general-purpose Large Language Models (LLMs) into reliable, specialized experts in fields like finance or medicine, which requires capturing the nuanced, often uncodified, 'gut feeling' knowledge held by human professionals.

Detailed Analysis

The discussion focuses on the critical challenge of infusing deep, specialized expertise into AI systems, moving beyond the capabilities of general LLMs like GPT-4. The solution proposed involves a layered methodology consisting of three steps: Human-to-Human (H2H) interviews, Human-to-AI (H2AI) elicitation, and Data Analysis for validation. The H2H phase involves interviewing domain experts (like a stock trader) to extract tacit knowledge—rules, intuition, and gut feelings—which are then either structured into formal rules or encoded in documents like JSON or YAML. The H2AI phase involves having the AI interview the expert in real-time to articulate their thought process, ensuring the AI captures the nuances missed in static documentation. The validation phase checks if the AI's newly formed rules conflict with existing general patterns learned from broad data. An example provided is the 'Momentum Rule' from a 2021 paper by Carrigan, Hallman, and Bertini, which dictates selling a stock if it drops 8% below purchase price unless the expert intuition suggests otherwise (e.g., positive market sentiment). The success of this layered approach—combining explicit rules with deep, elicited knowledge—is crucial for creating trustworthy, specialized AI agents that avoid the pitfalls of relying solely on generalized, potentially biased, training data.