Generative AI: Opportunities and Effects on Society

Danila Petrelli moderated a panel held shortly after the release of GPT-4 and the Future of Life Institute’s open letter.

Sylvie Saget, University of Gothenburg, described academic work on new paradigms rather than larger models, and argued that factuality cannot be guaranteed, so that these systems suit assistive tasks in which the user keeps the responsibility. Aron Lagerberg, Recorded Future, said that complex annotation tasks still need people, and that without information on GPT-4’s training data, bias in downstream models is hard to trace. Rebecca Oskarsson, Pirr, described the trade-offs between cost, speed, and quality when running generative models in production, and users who now expect to instruct models as they do ChatGPT. Francisca Hoyer, AI Sweden, spoke about who benefits from these systems, the need to explain their limits, and how visible bias in their output can start a debate about the kind of society we want.

The panel agreed that general AI is not here, and that more pressing questions come first.