> Models are biased, but it is a bias we control through the selection of training data. But what if controlling bias is impossible because the very format of

> Models are biased, but it is a bias we control through the selection of training data. But what if controlling bias is impossible because the very format of the data demands its emergence? I always thought both structuralism and poststructuralism were the worst types of in(s/n)ane navel gazing. I think I even wrote on a course feedback section in a final many, many years ago that these topics were massive wastes of my time. (We didn't have "evals" as such, but some faculty would elicit feedback on a separate sheet of paper you could leave with the final exam.) Observation: the more time I spend just letting the algorithms converge under a huge variety of treatments, the more convinced I become that the most basic structures of our language cast a spell on human understanding of the world, and that the interactions between the primitives of our language and human meaning-making is FAR more deterministic than we would like. Especially in terms of how they direct certain emergent phenomena in human groups. Implication: maybe it's reasonable to hypothesize that bias isn't "in the data". Maybe bias emerges, nearly deterministically, from the underlying grammatical structure of the data, at least in certain semantic contexts. Maybe certain utterances, and the persuasive power of those utterances, correspond to a sort of default hard-wired path for a majority subset of the population. A co-occurrence of certain neural pathways and certain grammar/semantics that eventually must arise in a given language and socially situated use of that language. FWIW: I have almost no social science training, the training I did get was received by me with extreme skepticism, and this sounds bunk to me even as I type it. So, for me, this hypothesis would have sounded stupid -- or at least inane/pointless -- before 2019. But I'm becoming increasingly convinced that the languages we've converged on encode certain types of group preferene/bias at an extremely fundamental level. ...either that, or I'm far FAR worse at filtering/generating data than I give myself credit for. But in at least a few cases I have literally reviewed every token within a certain delta of influencing the models behavior on an input sequence and see the same results.