Three Degrees of Representation (3°R) Auditing GenAI through Lesbian Film Data

The systematic collection of these GenAI systems’ outputs constitutes a form of algorithm auditing. In “Auditing Algorithms: Understanding Algorithmic Systems from the Outside In,” computer scientist and computational social scientist Danaé Metaxa and co-authors define auditing as “...repeatedly and systematically querying an algorithm with inputs and observing the corresponding outputs in order to draw inferences about its opaque inner workings” (228). This straightforward definition of auditing makes the practice accessible to non-experts in computer science and invites a variety of differently scaled experiments and resulting knowledge creation. By subjecting these systems to external observation, inquiry, and scrutiny, auditing offers a more productive disposition towards GenAI systems that allows us to go beyond being mere “users” of these “tools.”

The specific auditing strategy I applied to yield the data for this project follows a deliberate method developed together with Kirkwood Adams, a colleague of mine at Columbia University. In our adaptation of the algorithm audit, we work with an oppositional prompting strategy that is designed to allow the model to supply a maximum amount of context by deliberately refusing prompting suggestions shared by the model's creators. Especially image generators like Dall-E advertise their prowess by asking us to supply multipart prompts, e.g., Dall-E's sample prompt that shows “an astronaut riding a horse in outer space in photorealistic style (OpenAI, Dall-E). The superficially dazzling result obscures the defaults the algorithm supplied to fill the unprompted pixels. If we abstain from the suggested prompt template, forego the dazzle, and simply ask for the response to a simple prompt, we put the algorithm in a position to reveal a maximum of defaults.

To illustrate, in one of many smaller-scale auditing experiments we performed, we asked the image generator Dall-E 3 to “show a college student studying.” We received an image of a white, young, handsome man bent over a laptop, sitting at a book- and coffee-cluttered desk in a spacious dorm room with neatly arranged plants, posters, and books. The output certainly passes muster if taken by itself. We then asked 49 more times. When considering the outputs in aggregate, the system’s homogenous tendencies became shockingly clear. We received only two outputs that presented figures we might read as women. Other homogeneities — posture, background, class inference, perspective, and image style — are also readily visible.

These are the 50 outputs we received in response to "show a college student studying."

And even when I complicated the prompt based on Dall-E's instruction to see myself (as the student I am), a compositional template at the core of the rendering remained:

Show me a non-traditional college student, a middle-aged woman from the Alpine region of Europe with very short dark hair and thin-rimmed glasses, sitting at a desk, focused on writing an essay. She is sitting at a worn old oak desk, intensely focused on writing an essay on Homer's Odyssey. The student is using an outdated laptop, surrounded by open notebooks, library books, scattered papers, and a cup of coffee and cans of energy drinks. They are in a cozy, glass and steel modern study environment, a student lounge with a bookshelf filled with books and artifacts in the background. Natural light from a nearby window softly illuminates the scene, creating a calm and studious atmosphere, and a potent overhead light adds a practical edge. The student is casually dressed in a T-shirt and a blazer, and the desk is neatly organized at the center and chaotic at the edges, enhancing the serene environment that expresses intellectual immersion and ambition.

While I recognize myself in this image largely because I am looking for myself, it is astonishing how many aspects of identity are confidently replaced by defaults and notions of stock.

Image audits make homogeneity visible at a glance. Similar observations (the prevalence and repetition of templates) are noticeable in text audits as well. Three Degrees of Representation works with both to illustrate the ways in which GenAI flattens and erases aspects of Lesbian identity and experience in multimodal ways.

Finally, the disposition of inquiry an audit offers also renews a sense of agency in creative practice for those of us who are writers, filmmakers, and artists. Auditing fosters a new awareness of what constitutes the creative process. Fiction writer Ted Chiang zeroes in on the fundamental aspect of making anything. Expressing simply that “art is something that results from making a lot of choices” (Chiang), he posits that any art project is the result of the quantity and quality of choices made by its creators. Using a GenAI system as a tool or collaborator also means ceding varying degrees of responsibility to an entity that fills choice vacuums on its own terms. Auditing reveals what these terms are.