mp | Machina machista
href="//cdnjs.cloudflare.com/ajax/libs/highlight.js/9.15.6/styles/atelier-dune-light.min.css"><br>-->
home
blog
code
about
privacy
Menu
home
blog
code
about
privacy
Machina machista
22 Jul 26
Your AI thinks you’re a dude1. And that’s a problem for everybody who is<br>not. Experiments prompting several LLMs2 with matched decision-making<br>scenarios<br>show escalatory recommendations are around ten times more likely when actors<br>are identified as male rather than female, while compromise-seeking<br>recommendations are skewed toward female identities. And default behaviour in<br>the absence of gender specifiers is overwhelmingly “male”. The more such models<br>inform real decisions, the greater the risk of building genuine machina<br>machista — machines whose ordinary use fosters and normalises misogynistic and<br>macho behaviour throughout society.
If norms can be said to form us, that is only because some proximate,<br>embodied, and involuntary relation to their impress is already at work.
Judith Butler (2024), “Who’s afraid of gender?“
The Large Language Models (LLMs) which power almost all contemporary AI are<br>trained on vast textual corpora. All corpora manifest bias, from the subjective<br>biases of every human author, to collective biases of authors’ societies. Those<br>biases are fed as input into LLMs, ensuring that the outputs are also<br>Semantics derived automatically from language corpora contain human-like biases‘, Aylin Caliskan, Joanna J. Bryson, and Arvind Narayanan (2017) Science 356 (6334): 183-186" class="has-popover">unavoidably biased.<br>In present form, most publicly prominent LLMs<br>are profoundly<br>biased to AI is already rewriting reality for billions of people. It is getting women wrong.‘, United Nations Women (June 2026)." class="has-popover">disfavour<br>women<br>and other genders at the expense of men.
There have been many studies of Bias and Fairness in Large Language Models: A Survey‘, Gallegos and others (2024), Computational Linguistics 50(3): 1097-1179" class="has-popover">gender bias in<br>LLMs, the majority of which have<br>examined frequencies or likelihoods of gender-specific “sensitive words” being<br>produced in response to various inputs. These kinds of studies are an important<br>way to reveal the bias inherent in LLMs. Showing bias in outputs generated by<br>LLMs provides vital defence against their use in sustaining machismo and<br>misogyny.
Machismo, misogyny, and machines<br>Humans can be misogynistic or macho; machines can’t. The only thing machines<br>can do is to faithfully hold the impress of human misogyny or machismo, and<br>process that to generate biased outputs. No machine, including LLMs, can be<br>“misogynistic”, because misogyny is a hatred of women, and machines can’t hate.<br>They also can’t “love”, or even “like”. Similarly, machines can’t manifest or<br>enact “machismo”, because they have no concept of self, let alone of a gendered<br>self. Directly referring to LLMs as misogynistic or macho grants them too much<br>humanity. In themselves, LLMs remain merely “biased.”
And yet machines have only ever been defined by their usage. My claim that LLMs<br>can not be misogynistic or macho is also equivalent to the US-centric claim<br>that, “guns don’t kill, people do,” which is used as a excuse to oppose gun<br>control. Of course guns kill people, even though that mostly only occurs<br>through people actually pulling the triggers. In the same sense, the inability<br>of machines to actually be misogynistic or macho is merely technical. My<br>intention in asserting that here is only to encourage the blame for the<br>misogyny or machismo built into and emerging out of these machines to be<br>rightly directed at the people building and using them. Those people willingly<br>exacerbate misogynistic or machista tendencies, and any increase in their use<br>of these machines will only exacerbate both of these throughout all societies<br>in which these machines are used.
Any references I make to “machina machista“ are intended to be interpreted in<br>this context. Technically, machines can be neither misogynistic nor machista,<br>but their outputs and usage sure can.
The kinds of biases manifest in literally biased textual outputs can be<br>ameliorated by passing such outputs through additional models or processes to<br>intercept and modify texts to reduce resultant biases and restore appearances<br>of gender-neutrality. Many LLMs are passed through additional post-training<br>processes which aim to do just that. I also include one “uncensored” model in<br>the analyses here, to compare outputs with the equivalent “censored” version<br>that has been post-trained to reduce bias.
Biases inherent in LLMs can also be manifest in deeper and more socially<br>nefarious ways. What if LLMs are used to support decision-making processes in<br>ways that are themselves biased, regardless of any discernible gendering of<br>their inputs or outputs? Are there decision-making processes that are typically<br>male? Or typically female? There has naturally been a wealth of academic<br>research on...