Mit:

Maria Eriksson
European Centre for Algorithmic Transparency, Sevilla

“AI Benchmarking as Experimental Practice”

AI benchmarks comprise a multitude of quantitative techniques for evaluating the capabilities and potential risks of AI models. They range from multiple-choice questionnaires to frameworks for assessing the free-text outputs of AI models, and evaluations more akin to psychological intelligence tests. Within the AI industry, benchmarks have long been relied upon as valuable sources of insight about the performance of AI models. They are also used in policy contexts to assess and minimize potential societal harms posed by AI. In her talk, Dr Maria Eriksson (European Centre for Algorithmic Transparency, Seville) approaches AI benchmarking as an experimental practice and epistemic template for how to “know” AI models. Drawing on insights from the anthropology and sociology of tests and experimental modes of knowledge production, she examines how benchmarks do not simply describe AI systems, but actively (re)configure research agendas, competitive dynamics, regulatory imaginaries, and visions of technological progress.

Gäste sind herzlich willkommen. Im Anschluss findet ein Apéro statt.