Framing what and how to think: Lay people’s metaphors for algorithms

An informed public discussion about explainable artificial intelligence requires that laypeople and experts can negotiate in a language accessible to both. In our paper, we argue that this requires attention to metaphors. Metaphors – roughly speaking linguistic images – are necessary in such discussions to make abstract concepts comprehensible. We also argue that the typical metaphorical “vocabulary” is relatively narrow and sometimes has problematic implications. We conclude with some suggestions on how to systematically generate more appropriate metaphors. 

When explaining or negotiating abstract concepts – such as algorithms, large language models or artificial intelligence – humans almost invariably refer to metaphors. In the understanding to which we refer in our article, metaphors are not a means of embellishing a text linguistically, but are cognitively and even discursively relevant: They render abstract concepts more understandable and negotiable because they structurally relate them to more concrete and thus more familiar experiences or objects. An algorithm, for instance, may be explained as a recipe with several ingredients and steps or as a valet who fulfills certain functions for his mistress. 

Based on conceptual metaphor theory, in our contribution, we argue that such metaphors do more than helping to understand: they create a frame – a local context – for individual thinking and public discussion. Framing artificial intelligence or algorithms as a valet, for instance, highlights that one of their functions is to support human cognition and that algorithms and humans share tasks. But the metaphor hides (among other aspects) that much software is developed by companies with the interest to sale products. 

In order to start an empirical analysis of this problem, we created a corpus of metaphors for algorithms that we collected from voluntary English-speaking participants on the online platform Prolific. The participants responded to the instructions to 1) complete the sentence “An algorithm is like …” with an image or a metaphor and to 2) explain their metaphor in a few sentences. These data were then analyzed with common metaphor analysis procedures: First, all algorithm-related words or phrases in the text were identified according to the MIPVU procedure developed by Steen and coworkers.  Second, conceptual metaphors/ were identified by two coders.  Finally, we classified the metaphors into larger groups derived from philosophy of technology:  metaphors related to architecture (structuralist view) and relevance (functionalist view) and critically analyzed their implications for understanding and negotiating technical artefacts.

Both architecture and relevance metaphors were frequent, and only few metaphors could not be classified as either architecture or relevance. Kroes' distinction indeed appears to be relevant for everyday thinking. In addition, some interesting features emerged. 

The most frequent conceptual metaphors describing an algorithm’s architecture or structure were recipe, map, and pattern. In the relevance domain, the larger conceptual domain machines are people was most frequent. Personifications – an element of this class – liken the algorithm itself to a person directly (best friend, child, spy, crook, teacher) or by implication through its actions (want, know, think, believe). Verb metaphors, which may emphasize agency, were more frequent in in the relevance than in the architecture domain. 

Architecture metaphors rarely occurred on their own. They were often used in combination with relevance metaphors, for example, “an algorithm is like a pattern that computers follow to find the answer to something.” Relevance metaphors, on the other hand, were also used on their own, without any addition of architecture metaphors. Texts that mainly used relevance metaphors often described algorithms as a kind of homunculus and used verb metaphors in which the algorithm was the subject. 

The two classes of metaphors frames open up distinctive argumentative contexts for thinking and talking about algorithms. What we find most remarkable is that texts that begin with relevance metaphors make little reference back to structural relations. The perspective on both elements of the definition of an algorithm seems to succeed more from the direction of architectural metaphors, possibly because of their more obvious limitations. 

To summarize, there seem to be two major source domains from which people draw the metaphors they use to talk about algorithms. Choice of metaphor may support or hinder perspectives in thinking about algorithms.