Apollo Research proposes a step-by-step framework for auditing AI systems

In short

Apollo Research, an AI safety research organisation, has published a framework for auditing AI systems. It traces the chain of causes that leads from how an AI is built to how it affects the world. It is a policy and research paper, not a test of any AI system.

What happened

Apollo Research published the paper on 8 November 2023. It says the framework focuses on catastrophic risks, meaning the risk of very serious accidents. Apollo argues that auditing should be proportional to an AI system's capabilities and to the affordances available to it.

An affordance, in Apollo's wording, is a resource or opportunity for affecting the world that is available to an AI system. One example it gives is access to the internet. Apollo says affordances determine which of a system's capabilities it can actually use.

The framework works backwards from real-world effects. Apollo lists ten links in the chain:

• What the system actually does, and in what context
• The affordances available to it
• Its capabilities and its tendencies to use them
• Its internal structure, during and after training
• The learning process that produced that structure
• The computing power and training data used
• Security
• Deployment design, including who gets access to what, and when
• The design of the training experiment
• Governance and institutions, including regulations, standards and norms

Apollo names five types of audit, each meant to give assurance about different links in the chain: AI system evaluations, security audits, deployment audits, training design audits, and governance audits.

The paper also lists high-priority research questions. These are interpretability (understanding what happens inside a model), predictive models of capabilities and alignment, structured access, and possible barriers to AI labs being open with regulators.

What was done about it

The source describes no incident, so there was no response to report. Apollo's main policy recommendation is that proposals to change the affordances available to an AI system should themselves be audited. It says affordances can be limited through guardrails, staged deployment, prompt filtering, safety requirements for open-sourcing, and effective security. Apollo says the framework may help in designing auditing and governance regimes.

Why it matters

Apollo says that responsible AI development needs comprehensive auditing and that oversight mechanisms are needed to prevent harm. It says its clearer terms should lead to better governance, and it singles out affordances as a point where regulation could act. The source does not say whether any regulator or company has adopted the framework. Read the original report for the full paper.