I have recently written two posts on the issue of multiple meanings, polysemy and ambiguity, in AI discourse. When chatting with various people interested in AI metaphors, which has so far been my research focus, I realised that we all had neglected the topic of polysemy, or the multiple meanings that words have in a language or can have in language use. Metaphor is one way of creating multiple meanings, as in ‘my leg hurts’ and ‘that table has three legs‘ – ‘leg’ has now (at least) two meanings.
In the case of AI, think for example of the metaphor of ‘mirror’ which, in the context of computational technology, can mean a ‘psychological mirror’ that prompts people to reflect on human thought, memory, and identity (Sherry Turkle) or a mirror reflecting back human traits and cultural and political biases (Shannon Vallor). The first metaphor highlights that computers are not just ‘calculators’ and the second that AI systems are not ‘minds’.
In this post I want to explore why awareness of polysemy, over and above awareness of metaphor, might be important in social science studies of AI, especially critical AI literacy studies and critical AI literacy itself as the ability to engage with AI systems by understanding their technical foundations, linguistic framing, ethical implications and embedded power structures.
The ubiquity of polysemy
When thinking about this, it dawned on me that not only metaphor but also polysemy awareness might be important across scientific disciplines, from AI to genomics and beyond.
I still remember Ian Wilmut, the famous ‘creator’ of Dolly the cloned sheep, giving a talk at the University of Nottingham in the middle of debates about replicants, photocopies and so on of humans and animals. During that talk he used the, for him, entirely neutral term, ‘copy’ dozens of times, not realising that this word resonated quite differently with people who were not core scientists in the field. In short, he wasn’t aware of the fact that that the word ‘copy’ had multiple meanings in different contexts of use.
I believe that such blind spots and lack of polysemy awareness should perhaps be discussed more in the context of science and society – across all sciences. We all need more awareness of metaphors and polysemy – lay people as well as scientists, technical experts, and corporations. I shall briefly give examples of polysemy as important to technical, strategic, societal, critical discourses.
The ubiquity of polysemy in AI discourses
In technical AI discourse we find for example metaphors like ‘misalignment honeypots’. This polysemous metaphor has one foot in cybersecurity discourse (where it means a trap, a fake, an enticing target set up to lure attackers and catch them in the act) and now also one foot in AI security discourse (setting up a controlled, artificial scenario designed to tempt a misaligned AI into revealing its bad behaviour before it gets deployed in the real world). Both these meanings are not self-evident to ordinary users of AI, despite the use of the everyday word ‘honeypot’.
Strategic AI discourse exploits multiple meanings of words like ‘intelligence’ for example to hype and deceive (as discussed in the article that sparked my interest in polysemy).
In societal AI discourse numerous metaphors like ‘co-pilot’ or ‘agent’ or ‘slop‘ circulate with one meaning or many meanings and meanings that change and shift over time.
And finally, critical AI discourse investigates all of this and its educational, societal and political impacts, focusing in particular on the use of the emergence and change of metaphors in science and society (as I have done since 2022). I’d argue that in this context polysemy awareness might be more critical than anywhere else.
Polysemy awareness and polysemy literacy
There seems to be something paradoxical about the critical AI studies discourse. Critical AI discourse is largely made of language, it depends on contested terms like ‘intelligence’, ‘understanding’, ‘bias’, ‘alignment’, which are all massively polysemous and doing enormous amounts of ideological work. And yet, so far, this field has mainly focused on metaphors and overlooked polysemy (and I am guilty as well). By not being aware of polysemy (with some exceptions), critical AI studies and critical AI literacy studies may miss out on a conceptual tool that allows us to dig deeper into the societal as well as ideological impacts of AI.
The article I discussed in my earlier post on polysemy and power, which opened my eyes to polysemy, should, I believe, be required reading in critical AI literacy studies. In their preprint, Travis LaCroix et al. introduced the concept of ‘glosslighting’ “the practice of using technically redefined terms to evoke intuitive—often anthropomorphic or misleading—associations”. The glosslighting phenomenon works because the audience doesn’t have the analytical vocabulary to notice it happening. There is a lack of polysemy awareness!
Polysemy awareness is a kind of critical literacy, and its absence leaves people responding to the emotional or rhetorical associations of terms like ‘safe’ or ‘intelligent’ or ‘hallucination’ without being able to pin down what is being claimed. ‘Hallucination’, for example, implies something about the inner life of the system (it’s experiencing something incorrectly), thus suggesting the system is similar to a human, while providing plausible deniability that it is just a technical term.
In this little post, I just wanted to put down a marker highlighting the importance of polysemy awareness. There is much more to do. Metaphor analysis has come a long way as part of critical AI studies and studies of AI literacy, but polysemy awareness, designed to increase AI literacy, might fill a gap that we didn’t even know was there!
Image: Pexels: Landiva Weber: abstract art with flowing shapes in red, green, and blue tones
Making Science Public 
Leave a comment