Offert and Dhaliwal flag three method traps in critical AI studies

A short methodological paper posted to arXiv in late 2024 is worth flagging for anyone who writes about AI for a critical audience, because it names three habits that have quietly become the default style of the genre. Fabian Offert and Ranjodh Singh Dhaliwal, in a piece on arXiv titled The Method of Critical AI Studies, A Propaedeutic, argue that the field has three recurring methodological weaknesses they call casuistries.

The first is benchmark casuistry, which they describe as a tendency to overestimate the explanatory power of individual samples. The second is black box casuistry, a dependency on theoretical frameworks derived from earlier conceptualizations of computation. The third is stack casuistry, a preoccupation with a cause-and-effect model of algorithmic harm. Read together, the three names describe a style of critique that picks a single output, reaches for an older theory of how computers work, and then draws a straight line from the system to a harm. The authors argue that this style is reaching its limits.

What they call for is less specific than what they criticize. Rather than proposing a fully formed alternative method, they point toward future methodologies that, in their words, might take into account existing strengths in the humanistic close analysis of cultural objects. The honest caveat is that the abstract does not show what such an analysis looks like applied to, say, a frontier model, and it does not name the scholars whose work it has in mind. Take this as a diagnostic frame rather than a finished toolkit.

The forward-looking value, for practitioners who commission or publish AI criticism, is that the three casuistries make a usable internal checklist. Before sending the next piece on model bias or platform harm, it is now easy to ask whether the argument rests on a single example, on a dated theory of computation, or on a too-tidy causal chain, and whether anything would be lost by reading the system more like a cultural object.