Operationalism and AI

Alex Mussgnug from Stanford University will join us on Tuesday, August 4, 2026, at 15:00 for a lecture in our AI Ethics Lecture Series.

The event will take place in person at HLRS in Room Berkeley/Shanghai and will also be streamed online via Webex.

Join us online at:
https://unistuttgart.webex.com/unistuttgart/j.php?MTID=m70b6808f0edefeebf3e369cc3164936b

Abstract

AI applications in the social sciences have been criticized for their tendency to reduce complex concepts to the measurements used to operationalize them. This concern also extends to research investigating frontier AI itself, where benchmarks and behavioral tests can stand in for concepts such as intelligence, understanding, emergent capabilities, agency, or deception. Critical scholarship on AI often attributes this practice to naivety or negligence and highlights its ethical and fairness implications.

I argue that the history of science suggests a more constructive interpretation. In early experimental psychology, operational definitions that equated a concept with its measurement helped researchers gain an empirical foothold on poorly understood phenomena and bootstrap the development of new concepts.

Frontier AI is currently in a similar position, with many of its central concepts only beginning to take shape. I therefore propose that operational definitions can serve a similarly productive function in the development of concepts for frontier AI. The lecture will identify criteria for using such definitions responsibly and reliably.