Unsettling the Hegemony of Intention: Agonistic Image Generation DOI: XXXXXXX.XXXXXXX Conference: Conference title.; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/18/06 CCS: Human-centered computing Interaction paradigms CCS: Human-centered computing Interaction design theory, concepts and

Andrew Shaw, Andre Ye, Ranjay Krishna, Amy X. Zhang · 2025

Read Unsettling the Hegemony of Intention: Agonistic Image Generation DOI: XXXXXXX.XXXXXXX Conference: Conference title.; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/18/06 CCS: Human-centered computing Interaction paradigms CCS: Human-centered computing Interaction design theory, concepts and by Andrew Shaw, Andre Ye, Ranjay Krishna, Amy X. Zhang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-15242.txt.

Introduction

Shaw, Ye, Krishna, and Zhang present a 2025 image tool built on agonistic pluralism, the idea that political disagreement should be aired, not smoothed over. Today's tools aim to realize the user's intent, but image making is also a struggle over visual meaning, as the 2024 Gemini uproar showed. Gemini's opaque diversity fixes struck users as fake. The new tool instead shows rival visual readings of a prompt and invites debate.

In a lab study with 29 people, the agonistic tool boosted reflection across several measures, compared with standard, Gemini-style, and intent-refining designs. Reflection rose only when users found the tool both fitting and empowering. The authors conclude diversity and intent are not rivals to balance, but values to negotiate openly.

Catalog record

CREATORAndrew Shaw, Andre Ye, Ranjay Krishna, Amy X. Zhang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE79 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-15242

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Unsettling the Hegemony of Intention: Agonistic Image Generation DOI: XXXXXXX.XXXXXXX Conference: Conference title.; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/18/06 CCS: Human-centered computing Interaction paradigms CCS: Human-centered computing Interaction design theory, concepts and paradigms Andrew Shaw, Andre Ye, Ranjay Krishna, Amy X. Zhang Abstract . Current image generation paradigms prioritize actualizing user intention — “see what you intend” — but often neglect the sociopolitical dimensions of this process. As these tools become cornerstones of the media landscape, it is increasingly evident that image generation is political , contributing to broader social struggles over visual meaning. This sociopolitical aspect was highlighted by the March 2024 Gemini controversy, where Gemini faced criticism for inappropriately injecting demographic diversity into user prompts. Although the developers sought to redress image generation’s sociopolitical dimension by introducing diversity “corrections,” their opaque imposition of a standard for “diversity” ultimately proved counterproductive. In this paper, we present an alternative approach: an image generation interface designed to embrace open negotiation along the sociopolitical dimensions of image creation. Grounded in the principles of agonistic pluralism (from the Greek agon , meaning struggle), our interface actively engages users with competing visual interpretations of their prompts. Through a lab study with 29 participants, we evaluate our agonistic interface on its ability to facilitate reflection — engagement with other perspectives and challenging dominant assumptions — a core principle that underpins agonistic contestation. We compare it to three existing paradigms: a standard interface emulating current image generation tools, a Gemini-style interface that produces “diverse” images, and an intention-centric interface that suggests aestheticized prompt refinements before generation. Our findings demonstrate that the agonistic interface enhances reflection across multiple measures, but also that reflection depends on users perceiving the interface as both appropriate and empowering; introducing diversity without grounding it in relevant political contexts was perceived as inauthentic. Our results suggest that diversity and user intention should not be treated as opposing values to be balanced. Instead, interfaces can productively surface and navigate tensions between competing perspectives, enabling users to engage with and evolve their intentions in meaningful ways. 1. Introduction: Unsettling the Hegemony of Intention What does “a Founding Father” look like? What does “Jesus” look like? What about “a migrant,” “a doctor,” “World War II”? The production and consumption of visual meanings is a sociopolitical endeavor. When people consider their mental image of “a Founding Father,” their preconceptions are in conversation with others’ interpretations about who a Founding Father is and — just as importantly — is not. Today, our collective “answers” to these questions increasingly draw from images that are generated by biased AI models and then proliferate in our digital media ecosystem ( Jingnan 2024 ; DiResta and Goldstein 2024 ; Herrie et al. 2024 ) . The political context of image generation was further underscored when in March 2024 Google’s Gemini model came under criticism for generating images with inaccurate racial and gender diversity — for example, depicting Nazi soldiers as people of color. Such examples show how the values embedded in image generation tools shape how we view and understand our social and political world. Despite the increasing politicization of image generation tools, the dominant paradigm behind most current AI image generation tools is the actualization of intention : they should enable anyone to see exactly what they were intending . This “intention is all you need” paradigm ( Sarkar 2024 ) neglects the political dimension of image generation. By inhibiting engagement with competing interpretations of user prompts, intention-centric paradigms run the risk of entrenching problematic user intentions ( Cai et al. 2024 ) . Certainly, the Gemini controversy was as a meaningful attempt to grapple with the political dimensions of image generation by implementing corrections to address a lack of diversity in image generation models. Yet, while Gemini’s diversity intervention may have challenged the “‘hegemony of intention”, it did so in a way that caused significant public backlash, inciting “ an explosion of negative commentary from figures such as Elon Musk who saw it as another front in the culture wars ” ( Milmo and Kern 2024 ) . The swift pushback indicates that top-down interventions are ill-suited to the dynamic political landscape that image generation tools find themselves situated in. In short, the problem is that shallow efforts that do not involve users in discourse but simply alter their intent can come across to users as patronizing. In this work, we aim to build and evaluate an image generation interface that acknowledges the sensitive political contexts of image generation, including questions about diversity and representation, but unlike the Gemini case, encourages users to engage and reflect on their own intentions in relation to others on their own terms. Our design draws from agonistic pluralism (from the Greek agon meaning “struggle”), a political philosophy emphasizing the importance of conflict in the political realm and the productive ways it can be channeled instead of avoided. Given a user prompt, our image generation interface, Agonistic , collects and presents a range of interpretations of the prompt, prioritizing more controversial or challenging interpretations (Figure 1 ). Interpretations come with a justification grounded in concepts extracted from Wikipedia. Only after examining different interpretations and selecting one does a user then receive image generations. A critical part of agonistic pluralism is reflection — the process of considering other viewpoints and challenging dominant assumptions. For this reason, we conduct a comparative lab evaluation of our interface for its ability to foster productive critical reflection. We build three other paradigmatic interfaces for comparison — Baseline : produces images generated from the user’s prompt with no further interaction; Diverse : rewrites the user’s prompt to be more diverse before image generation; and Reformulative : provides prompt reformulations to produce more aesthetic and preferable images, inspired by existing work in prompt reformulation. Twenty-nine participants used these interfaces to iteratively build and refine a collage of 10 images representing a human subject of their choice while thinking aloud. Across several measures of reflection, we find that Agonistic induces the highest reflection out of any interface (§ 5 ), followed by Reformulative , then Diverse . For instance, when participants went to add images to a collage, they described the image as an expansion of their original intention 21% of the time while using Agonistic , but only 8%, 7%, and 3% of the time for Reformulative , Diverse , and Baseline , respectively. In addition, we find that participants express diversity as a justification for adding images to their collage 48\% of the time when using Agonistic , but only 28\% under Diverse . Our findings confirm that participants are likely to reject Gemini-style interventions due to a lack of factuality and authenticity, while Agonistic more authentically expresses diversity by situating user prompts and the generated images in relevant sociopolitical discourse. We find that getting users to acknowledge and prioritize diversity requires situating diversity in authentic political controversies, a different finding from current interventions that either seek to impose diversity for its own sake or constrain diversity with factuality. Agonistic aids critical reflection by giving users a sense of control and provide contextualizing elements to help them interpret images. These results point to richer ways that researchers and developers of

Read the complete text