Downsides and Challenges of Using Humans in AI Safety Evaluations

Watch on Youtube

Beth is describing the work of the Evaluations team at the Alignment Research Center (ARC) on developing evaluations of safety for cutting edge models, and on developing standards for scaling or deploying models conditional on the results of those evaluations. Beth got into EA via Givewell and Giving What We Can around 2014. After realizing AI might be a big deal she studied computer science, and worked on AI safety at DeepMind and OpenAI. She now leads the Evaluations team at the Alignment Research Center under Paul Christiano.