OpenMined’s Policy Lead, Lacey Strahm, joins researchers from leading institutions to propose guidelines for evaluating powerful AI systems.
As artificial intelligence systems become increasingly powerful and integrated into our society, how do we ensure they’re developed and deployed responsibly? A new paper published at NeurIPS 2023’s Workshop on Socially Responsible Language Modelling Research (SoLaR) tackles this crucial question. The authors, including OpenMined’s Policy Lead Lacey Strahm, propose a framework for external scrutiny of frontier AI systems.
Why External Scrutiny Matters
Frontier AI models – the most capable large language models available – come with significant risks, from potentially amplifying discrimination and disinformation to enabling cyberattacks. While AI developers are primarily responsible for ensuring their systems are safe, the paper argues that they shouldn’t be the only ones evaluating these powerful tools.
Just as we wouldn’t rely solely on pharmaceutical companies to verify drug safety or airplane manufacturers to certify their aircraft, we need independent external scrutiny of AI systems. This scrutiny can help:
- Verify claims made by AI companies about their systems
- Uncover potential issues that developers might have missed
- Provide trustworthy information to guide policymakers and the public
The ASPIRE Framework
The paper introduces ASPIRE – a framework outlining six key requirements for effective external scrutiny of AI systems:
Access: Scrutinizers need appropriate access to AI models and relevant information about their development and deployment. This includes access to model outputs, fine-tuning capabilities, and training data, while carefully balancing transparency with security concerns.
Searching attitude: Rather than just following checklists, evaluators need to actively probe for potential issues and unknown capabilities. This requires legal protections for good-faith scrutiny and incentives to discover problems.
Proportionality: The level of scrutiny should match the level of risk. More capable systems that could affect more people deserve more thorough evaluation.
Independence: External evaluators must be truly independent from AI developers in terms of selection, compensation, scope of work, and reporting of findings.
Resources: Proper scrutiny requires adequate time, funding, and computational resources. The paper suggests that current frontier AI models should receive at least six months of external evaluation before deployment.
Expertise: Given the complexity of AI systems, scrutiny requires diverse expertise across multiple fields and perspectives.
Looking Forward
The paper comes at a critical time, as governments and organizations worldwide are calling for increased external evaluation of AI systems. The UK has established a £100 million Frontier AI Task Force, the EU is considering requiring independent expert involvement in AI development, and major AI companies have voluntarily committed to external red-teaming of their models.
However, the authors emphasize that building an effective external scrutiny ecosystem will take time and careful consideration. While external scrutiny alone isn’t sufficient for ensuring AI safety and accountability, it’s a crucial piece of the puzzle for responsible AI development.
This blog post summarizes key findings from “Towards Publicly Accountable Frontier LLMs: Building an External Scrutiny Ecosystem under the ASPIRE Framework” presented at NeurIPS 2023. Read the full paper to learn more about implementing external scrutiny throughout the AI lifecycle and specific policy recommendations.