Exposing Assumptions in AI Benchmarks through Cognitive Modelling

Cultural AI benchmarks often rely on implicit assumptions about measured constructs, leading to vague formulations with poor validity and unclear interrelations. We propose exposing these assumptions using explicit cognitive models formulated as Structural Equation Models. Using cross-lingual alignment transfer as an example, we show how this approach can answer key research questions and identify missing datasets. This framework grounds benchmark construction theoretically and guides dataset development to improve construct measurement. By embracing transparency, we move towards more rigorous, cumulative AI evaluation science, challenging researchers to critically examine their assessment foundations.

Apolitical Intelligence? Auditing Delphi's responses on controversial political issues in the US

As generative language models are deployed in ever-wider contexts, concerns about their political values have come to the forefront with critique from all parts of the political spectrum that the models are biased and lack neutrality. However, the question of what neutrality is and whether it is desirable remains underexplored. In this paper, I examine neutrality through an audit of Delphi [arXiv:2110.07574], a large language model designed for crowdsourced ethics. I analyse how Delphi responds to politically controversial questions compared to different US political subgroups. I find that Delphi is poorly calibrated with respect to confidence and exhibits a significant political skew. Based on these results, I examine the question of neutrality from a data-feminist lens, in terms of how notions of neutrality shift power and further marginalise unheard voices. These findings can hopefully contribute to a more reflexive debate about the normative questions of alignment and what role we want generative models to play in society.