After one too many conversations about the ways GenAI + peer review = shitshow for all involved, I dashed off this slightly polemic commentary on how I think we should talk about GenAI as an epistemic carcinogen.

link.springer.com/article/10.1...

The next AusSTS is so soon. Come join us in Melbourne. Cfp is live at: aussts.org/aussts-2025-... #STS

Made it this far? Soon we’ll be starting an interview study with Renee Shelby to explore the development & use of taxonomies of societal impact . If you’d like to be kept in the loop please follow (glenberman.bsky.social).

Read more: arxiv.org/abs/2410.22985
5/5

And, we argue that understanding societal impacts in this way means accepting that predictive evaluations of GenAI impacts are inherently limited. We suggest a ‘governance-first’ approach may be more effective at meaningfully managing harmful consequences of GenAI.
4/5

We put forward 3 premises for rethinking societal impacts & their relationship to the introduction of GenAI in a particular social context. Societal impacts should be thought of as application- and context- specific, incommensurable with each other, and about questions of power.
3/5

To answer, we look at taxonomies of societal impact, developed to support GenAI evaluations. We ask: how do these taxonomies focus attention on which social factors are capable, and deserving, of measurement during GenAI evaluation?
2/5

In a paper we’re workshopping this week at #NeurIPS2024, Ned Cooper, @wesleydeng.bsky.social, and Ben Hutchinson and I ask: what is the model of societal impacts reflected in efforts to evaluate GenAI systems?

Paper: arxiv.org/abs/2410.22985
Workshop: evaleval.github.io
1/5