Auditing large language models: a three-layered approach (2302.08500v2)

Abstract: LLMs represent a major advance in AI research. However, the widespread use of LLMs is also coupled with significant ethical and social challenges. Previous research has pointed towards auditing as a promising governance mechanism to help ensure that AI systems are designed and deployed in ways that are ethical, legal, and technically robust. However, existing auditing procedures fail to address the governance challenges posed by LLMs, which display emergent capabilities and are adaptable to a wide range of downstream tasks. In this article, we address that gap by outlining a novel blueprint for how to audit LLMs. Specifically, we propose a three-layered approach, whereby governance audits (of technology providers that design and disseminate LLMs), model audits (of LLMs after pre-training but prior to their release), and application audits (of applications based on LLMs) complement and inform each other. We show how audits, when conducted in a structured and coordinated manner on all three levels, can be a feasible and effective mechanism for identifying and managing some of the ethical and social risks posed by LLMs. However, it is important to remain realistic about what auditing can reasonably be expected to achieve. Therefore, we discuss the limitations not only of our three-layered approach but also of the prospect of auditing LLMs at all. Ultimately, this article seeks to expand the methodological toolkit available to technology providers and policymakers who wish to analyse and evaluate LLMs from technical, ethical, and legal perspectives.

Audit Methodologies for LLMs

The proposed audit methodologies for LLMs are an essential step in addressing the ethical and social challenges associated with these advanced AI systems. The three-layered approach to audits—encompassing governance, model, and application levels—provides a comprehensive framework that is effectively designed to capture the complex dynamics and potential risks that LLMs pose.

Introduction to Auditing LLMs

LLMs, such as those based on transformer architectures, have seen widespread adoption due to their impressive ability to generate human-like text across various tasks. While presenting numerous opportunities, including unprecedented adaptability and scaling capabilities, LLMs also come with ethical concerns like discrimination, information hazards, misinformation, misuse, interaction harms, and environmental footprint. Effective auditing can thus serve as a robust governance mechanism, ensuring that the deployment of LLMs aligns with ethical, legal, and technical benchmarks.

Layers of Auditing

1. Governance Audits
   • Objective: Examine organizational procedures and governance structures to ensure effective risk management systems are in place.
   • Key Activities:
     • Review adequacy of governance frameworks to adhere to best practices.
     • Develop audit trails documenting LLM development processes thoroughly.
     • Map roles and responsibilities for accountability concerning adverse system outcomes.
2. Model Audits
   • Objective: Evaluate models post-training but pre-deployment, focusing on their intrinsic properties to inform potential applications and limitations.
   • Key Characteristics Evaluated:
     • Performance: Using structured benchmarks like GLUE and BIG-bench to measure real-world task execution.
     • Robustness: Assessing sensitivity to unexpected inputs using tools like Robustness Gym.
     • Information Security: Determining potential for sensitive data leakage, employing techniques like exposure metrics.
     • Truthfulness: Evaluating capability to produce accurate information, employing metrics from benchmarks such as TruthfulQA.
3. Application Audits
   • Objective: Monitor applications built on LLMs post-deployment to continuously assess their impact and compliance.
   • Approach: Implement continuous monitoring technique to evaluate performance over time, keeping track of dynamic risk factors and emergent behavior in LLMs.

Harmonization of Audits

Cross-leveled harmonization of audits is crucial, where outputs and insights from one level inform approaches and focus at another. For instance, data from ongoing application audits can refine governance protocols and adjust model training to mitigate identified shortcomings. Moreover, audit methodologies ought to adapt to anticipated risks that vary significantly across contexts.

Conclusion

The proposed audit structure is a robust response to the inherent challenges posed by LLMs. By employing external yet collaborative audits across governance, model, and application levels, technology providers can effectively manage risks associated with LLMs and provide a foundational level of assurance for their widespread and responsible use. While the methodology presents a sophisticated framework, it is essential to acknowledge its limitations, requiring continual refinement and adaptation to emerging risks and ethical questions as the societal impact of AI evolves.