[Paper Review] A Sign That Spells: DALL-E 2, Invisual Images and The Racial Politics of Feature Space

Fabian Offert, Thao Phan|arXiv (Cornell University)|Oct 26, 2022

Participatory Visual Research Methods20 citations

TL;DR

The paper analyzes how DALL-E 2 and similar models encode and reproduce whiteness through feature extraction and semantic compression, arguing that debiasing is often client-side and that foundation models reshape visual culture in racially charged ways.

ABSTRACT

In this paper, we examine how generative machine learning systems produce a new politics of visual culture. We focus on DALL-E 2 and related models as an emergent approach to image-making that operates through the cultural techniques of feature extraction and semantic compression. These techniques, we argue, are inhuman, invisual, and opaque, yet are still caught in a paradox that is ironically all too human: the consistent reproduction of whiteness as a latent feature of dominant visual culture. We use Open AI's failed efforts to 'debias' their system as a critical opening to interrogate how systems like DALL-E 2 dissolve and reconstitute politically salient human concepts like race. This example vividly illustrates the stakes of this moment of transformation, when so-called foundation models reconfigure the boundaries of visual culture and when 'doing' anti-racism means deploying quick technical fixes to mitigate personal discomfort, or more importantly, potential commercial loss.

Motivation & Objective

• Investigate how large visual models like DALL·E 2 reconfigure visual culture through feature extraction and semantic compression.
• Examine how ostensibly neutral representations reproduce whiteness and racialized concepts.
• Critically assess OpenAI’s debiasing efforts and their political and commercial implications.
• Argue for new humanist critiques of foundation models beyond data or representation deficits.

Proposed method

• Analyze OpenAI’s release notes and public statements about DALL·E 2 and debiasing.
• Examine user reports and experiments (e.g., prompts like 'a sign that spells') to reveal how bias can emerge from user interaction.
• Discuss theoretical concepts of feature space, invisuality, and whiteness in the context of large visual models.
• Draw on critical race theory and posthumanist literature to frame the politics of visual culture in ML systems.

Experimental results

Research questions

• RQ1 How do large visual models reconfigure the boundaries of visual culture through feature extraction and semantic compression?
• RQ2 In what ways does DALL·E 2 reproduce or stabilize whiteness as a latent feature of dominant visual culture?
• RQ3 What are the political and commercial implications of debiasing efforts in foundation models?
• RQ4 How can critical, humanist critiques of machine learning better address the role of language and prompts in shaping outputs?

Key findings

• Debiasing claims by OpenAI are revealed as user-facing keyword additions rather than systemic model changes.
• Feature space in generative models dissolves and resurrects concepts like race in ways that reinforce whiteness.
• Debiasing is tied to commercial interests and liberal multiculturalism rather than confronting racial injustice.
• The critique advocates for methods that examine machinic whiteness as a technical and cultural phenomenon, not just data gaps.
• User prompts can reveal biases and mechanisms of output without requiring full access to proprietary models.