How to audit your LLMs to build trusted AI

Uncover hidden risks and drive business value with LLM auditing

Donovan @trybricks.ai
2 min readFeb 12, 2024

🔍 The Three-Layered LLM Auditing Framework 🔍

In the ever-evolving landscape of AI, transparency and accountability have never been more critical. AI researchers from Oxford proposed a three-layered approach to LLM auditing, encompassing Governance, Model, and Application Audits. Each layer is designed to ensure that LLMs are developed, deployed, and utilized with the highest ethical standards and compliance in mind.

Source: Mökander, Jakob, et al. “Auditing large language models: a three-layered approach.” AI and Ethics (2023): 1–31.
  • Governance Audits: Assess the organizational frameworks and procedures of companies developing LLMs to ensure ethical and responsible AI governance, needed during the foundational stages of LLM development and before deployment.
  • Model Audits: Evaluate the technical capabilities, limitations, and ethical implications of LLMs themselves, necessary after initial model training but before their application in real-world scenarios.
  • Application Audits: Examine the legality, ethical alignment, and societal impact of LLM-powered applications, crucial for continuous oversight post-deployment to ensure ongoing compliance and positive societal contribution

🌟 The Spotlight on Application Auditing 🌟

Among these, Application Auditing stands out for its crucial role for most companies in assessing the legality, ethical alignment, and impact of LLM applications. Why focus on Application Auditing, you ask? Because it’s where the rubber meets the road: ensuring that LLM applications are not only legally compliant but also positively impact users. Specifically, application Auditing examines:

  • Functionality Audit: Is the application’s purpose and operation within legal and ethical boundaries? For example, verifying that a customer service bot provides accurate, non-discriminatory advice.
  • Impact Auditing: What actual effects does the application have on its users and the broader community? For example, assessing whether an LLM-based content recommendation system fosters positive engagement without amplifying harmful content.

More than Ethical Concern: Discover Insights Through LLM Auditing.

While the recent EU AI Act has raised companies’ concerns about auditing to fulfill legal compliance, an often overlooked dimension is auditing how the input and output of LLM impact the end user from a business perspective.

Auditing the relationship between LLM application’s input and output can provide valuable insights for businesses and mitigate risks. For example:

  • What type of user query leads to conversion, and what type leads to customer churn?
  • Are there any dynamics in users’ inputs that indicate an emerging customer need?
  • How does the characteristic of LLM output lead to your key metrics like retention rate?

At BricksAI, we’re building LLM auditing features into our one-stop LLM infrastructure platform to make this process as easy as possible, so companies can focus on building features that drive business value.

If you’re an enterprise looking to start implementing LLM audits, we’d love to chat to see how we can help.

--

--

No responses yet