TLDR: A new research paper by Amanda Wasielewski explores what AI-generated images ‘want,’ reframing W.J.T. Mitchell’s art theory question. It argues that these images, fundamentally abstract, desire specificity and concreteness, aspiring to ‘objecthood’ but often producing ‘phantom objects’ due to statistical blending. Moving past speculative AI consciousness, the paper concludes that the most critical ‘desire’ of AI-generated images is for vast resources, particularly water, highlighting the significant environmental impact of data centers and the abstract nature of this material cost.
In a thought-provoking new paper, Amanda Wasielewski reframes a classic question from art theory to explore the hidden ‘desires’ of AI-generated images. Moving beyond the speculative fears of AI consciousness, the research delves into what these digital creations truly ‘want’ from an art historical and philosophical perspective, ultimately revealing a surprising and critical environmental cost.
The paper, titled “What Do AI-Generated Images Want?”, draws inspiration from W.J.T. Mitchell’s influential essay, “What do pictures want?” Mitchell shifted focus from human interpretation to the idea that pictures themselves possess agency and desires. Wasielewski adapts this concept to the realm of contemporary AI image generation, arguing that these images fundamentally crave specificity and concreteness because they are, at their core, abstract.
The Abstract Nature of AI Images
Multimodal text-to-image models, which are the primary focus, operate on the premise that text and image are interchangeable mathematical tokens. While users see a seamless transformation from textual input to visual output, this process obscures the underlying representational regress. The article explains that AI-generated images are statistical constructs, representations of data that are themselves representations of signs, ultimately stemming from collective thought. They are, in essence, mathematical abstractions.
Seeking Reality and Objecthood
According to the paper, AI companies often market their tools by emphasizing the creation of “realistic” or “photo-realistic” images. This desire for reality, or at least its semblance, is central to what AI-generated images ‘want’. They aspire to Mitchell’s concept of “objecthood” – the idea that a picture is not just an image but a material unit combining image, object, and discourse. However, AI-generated images produce “phantom objects” because they lack an individual reference object and are statistical constructs seeking bodies.
The paper illustrates this tension between abstraction and concreteness with compelling examples. When prompted to create “a school classroom with a hotel bed in the middle,” the AI often generates a child’s bed, complete with colorful sheets and curtains, rather than a neutral hotel bed. Similarly, “Trick-or-treaters wearing Christmas costumes” results in a hybrid image where Halloween elements (orange fur on a reindeer suit, Jack-O-Lantern pails) blend with Christmas decorations. These examples show how the statistical dominance of collective associations (childhood with classrooms, Halloween with orange) bleeds into the singular image, preventing it from achieving true, unpolluted objecthood.
Also Read:
- AI’s Inner Drive: Assessing Curiosity in Large Language Models
- Unmasking Watermark Vulnerabilities: A New Method for Forging Digital Signatures on Images
Beyond Human-Like Desires: The Environmental Toll
Wasielewski critiques the tech industry’s “eschatological angst” and the TESCREAL movement (Transhumanism, Extropianism, Singularitarianism, Cosmism, Rationalist ideology, Effective Altruism, and Longtermism), which attributes consciousness and autonomous desires to AI. While acknowledging the human impulse to project imagination onto creations, the author emphasizes a more pressing concern: the very real harms created by the AI industry, particularly its environmental impact.
The paper concludes with a stark and powerful answer to its central question: What do AI-generated images really want? They want our water. Data centers, essential for training and running generative AI, are enormous consumers of resources, especially water. The creation of a single AI-generated image, while seemingly small, contributes to this vast consumption. The complexity of the AI apparatus makes it difficult to trace the exact material origins or resource consumption for an individual image, rendering even its materiality abstract. This hidden, greedy desire for resources, particularly potable water in water-scarce regions, positions AI-generated images not as downtrodden entities, but as “kings and idols” consuming our most valuable offerings.
Ultimately, the paper argues that there’s no need to project human-like subjecthood onto AI-generated images to understand their desires. They simply want to be valid, like nature itself. Breaking the spell of their perceived inevitability is crucial to addressing the profound environmental challenges they pose.
