Prose and algospeak: the changing language of AI image generation

Abstract

This paper explores developments in the relationship between prose and prompting in the use of Large Language Models (LLMs) for visual generation. It observes the push towards increasingly descriptive language by two foundation models specifically marketed to creatives: Leonardo’s Phoenix and Adobe’s Firefly. It then considers how and why prose and ‘algospeak’ prompting has become a vital tool for artists in negotiating the affordances of prompt parameters. This paper argues that the language of prompting is changing through fine-tuning and iterative practices at both the technical and experiential level. Tracing these changes offers further insight into the role of text in the aesthetics of visual generation, as well as the tensions between creative work and corporate ethics-washing.

User interactions with LLMs for visual generation are characterised, at the interface level, by prompting. Promptology aims at optimising user input towards optimised model output. Directives for ‘good’ prompts emphasise a formulaic approach that lists subject, style, details and format. As LLMs have evolved, so too has the prominence of ‘details,’ with both Leonardo and Adobe encouraging increasingly descriptive prompts. This paper first evaluates the aesthetic consequences of prose prompting as a technical feature. For instance, when the visual generation AI model Leonardo launched to the public in December of 2022, it included ‘improve my prompt’ as an optional feature. Built on ChatGPT, the feature takes an existing user prompt and extends it into a highly descriptive prose paragraph with an abundance of adjectives. This feature has carried through into its foundation model ‘Phoenix,’ which was rolled out in July of 2024. With relatively minor changes in output between the simplified original prompts and the prompts ‘improved’ by Leonardo, questions arise around the motivations for the feature. Namely, what aesthetic — sensing and sense-making — value is being generated through the addition of AI-generated descriptive prose?

This paper then unpacks the ethical tensions around prompt parameters as a way of contextualising the experiential perspective, wherein prose prompting operates as a negotiated user affordance. Embedded limitations in the prompting process for many models are heralded as holding an ethical imperative, aimed at preventing their misuse. Alongside the fine-tuning of prompt parameters by the companies that own AI models, we have seen the rise of ‘algospeak’ style prompting, whereby users aim to circumnavigate the censorship and moderation of the platform to overcome these limitations. Algospeak prompts are not necessarily mobilised towards nefarious means. Often, they are an attempt at balancing the over-simplified nature of prompt censorship, which conflates things like the post-war art style as a user request for violent imagery, or the depiction of a pregnant stomach with a desire to generate pornographic nudity. The artist-academic Beverley Hood’s process of making her recent work Mother (2024) — an AI-generated photo-film made using Firefly — is provided as a grounded example to elucidate the role of experimental prose and algospeak prompting in visual generation.