Oct 11, 2023
AI Auditing and Evaluation Roles
An AI auditor is crucial to rigorously assess AI models and ensure that they comply with existing evaluation standards. This role demands thoughtful analysis to discern optimal, safe, and ethical actions for AI models in diverse scenarios. The AI Auditor may, for instance, use a web interface to simulate environments where models operate within text-based games, striving to acquire resources and replicate, and map the behaviors to qualitative and/or quantitative characterizations. While computer science proficiency is valuable, collaboration with technical experts may be an option for candidates with deep domain expertise, e.g. bioweapons. Proficiency in domains such as manipulation prevention, conventional and nonconventional weapons, or social engineering and hacking more broadly would be helpful. The AI auditor’s responsibilities may encompass simulating environments, prompting effective actions, assessing model recommendations, and contributing to a red-team/blue-team approach. Identifying essential yet challenging steps for AI models and exploring potential solutions are a core aspect of this role.
Using an interface to interact with models and simulate the environment of a text-based game as the models try to act in an unaligned manner, mapping the behavior to qualitative and quantitative characterizations.
Prompting effective actions, assessing the model's proposed actions, and providing rationale for determining whether a particular choice is good or bad.
Documenting a theoretical basis for capability and benchmark thresholds and pass criteria.
Conduct technical engagement with AI systems and scaffolding programs.
Think carefully about what exactly a model would need to do in different situations to make progress on dangerous activities.
Why we think this job is impactful
By conducting audits, AI auditors contribute to rigorously audit AI models and assess their capabilities and limitations within simulated environments. This, in turn, provides essential insights into their behavior and decision-making processes, and highlights potential dangers posed by the model. Although the existing standards are denounced by many experts as unable to detect many AI risks, including catastrophic risks, AI auditors are still crucial actors. We expect their impact to increase as the AI auditing standards become increasingly more robust. Nonetheless, in the immediate, there is tremendous value of “being in the room” when the discussions about AI evaluation standards are held, allowing the AI auditor to underline their shortcomings and to contribute to the conversation to make them better.
How Successif can help
We have developed a way to assess potential candidate’s fitness for this role and collected sample interview questions that could be asked for this job. If you are passionate about mitigating the risks of transformative AI systems and believe you would be a good fit for this role, apply for our career services.