AI Pedagogies of Contra-Technological Resistance

Societies carry existing violences and discriminations into the technological apparatuses they produce. Never mind how we use those machines, we are complicit. In this post, I suggest introducing students to (the critical philosophy of) AI through these ethical and political challenges. Furthermore, I suggest inspiring students to find creative ways to appropriate and resist these violent machines in the examples of Rosa Wernecke’s (2025) video work Before the Drop and Zach Blas’s (2015) video performance Contra-Internet Inversion Practice #1: Constituting an Outside (Utopian Plagiarism).

This post is based on four undergraduate seminars, titled “Beyond Representational Justice—Violence, Identity and AI,” taught from 2024 to 2026 and developed from my text “Beyond Representational Justice” (deLire 2023). The materials usually move, surprise, and inspire students. The technique should work for seminars on the (critical) philosophy of AI, but also (with adjustments) for courses on ethics of AI, political theory, critical theory, philosophy of race and gender, and queer theory. Depending on how much time you have, you may want to stretch this over two or three sessions. I typically assign Hito Steyerl’s text “Mean Images” and Sarah Lewis’ “Racist Photography” (a short New York Times article), but the class can also work impromptu if you just show the relevant passages or have the students read them on site. Literature for a follow-up session is included below.

Starting with Lewis’ article, I ask the students if they know how analog photography works (most don’t). Analog cameras are essentially dark chambers that contain hypersensitive photographic film. A convex glass lens converges incoming parallel light rays onto a precise focal plane, adjusting aperture depth to render distant objects sharply focused while excluding peripheral light scatter. Once opened (briefly), light burns into the film (called ‘exposure’), and with it an image of whatever the light was reflected by before falling into the chamber. Using various chemical processes, this film can be developed into photographic images.

Racism is built into the photographic apparatus in at least two ways. For one, differently sensitive film can be differently susceptible to different colors, which are essentially electromagnetic wavelengths. Until the 1970s, this film was geared towards the depiction of white people, for racist reasons, but also because white people were thought to be the primary clientele. So white people would tend to look better than black and brown people. It was only due to requests by industrial producers of chocolate and furniture in the 1960s and 1970s that companies provided film more susceptible to black and brown colors. Secondly, chemically developing the film requires a model with which to check whether the coloring is correct. That model used to be a “Shirley card,” showing a white woman with brown hair (called “Shirley”). The overall rule was that “if she looks good, the picture will look good.” Later on, white Shirley was joined by more diverse models for the same purpose. However, we can see that racism was built into the photographic apparatus both in the material film and in the measure used to develop the film. The technology is a product of the society that makes it—and that technology will inevitably mirror social conditions.

I then project the first paragraph of Steyerl’s text:

A while ago, science-fiction writer Ted Chiang described ChatGPT’s text output as a ‘blurry jpeg of all the text in the web’—or: as a semantic ‘poor image’. But the blurry output generated by machine-learning networks has an additional historical dimension: statistics. Visuals created by ML [machine learning] tools are statistical renderings, rather than images of actually existing objects. They shift the focus from photographic indexicality to stochastic discrimination. They no longer refer to facticity, let alone truth, but to probability. The shock of sudden photographic illumination is replaced by the drag of Bell curves, loss functions and long tails, spun up by a relentless bureaucracy. (Steyerl 2023)

I ask students what the distinction is that Steyerl bases her critique on, and we reconstruct it as follows:

	Photographic Images	AI ‘Images’
Source:	actually existing objects	statistics (database)
Mode of production:	indexicality	stochastic discrimination
Truth procedure:	facticity (correspondence)	probability
Temporality:	sudden illumination (causal connection)	bell curve

I then offer them an extended version of this distinction, which differentiates kinds of media by their mode of production into symbolic, indexical, and AI media. In symbolic media (drawing, painting, musical notation), a human actor, functioning themselves as the medium, actively generates the representation. For example, a painter may draw a sketch of a tree or a person. In contrast, indexical media (photography, audio recording, fingerprints) are produced by direct causal interaction between the represented object and its representation. As discussed earlier, in a photographic image, the photographic film is exposed to light that was reflected by an object beforehand (“sudden illumination”). In this sense, an indexical medium points to its object (the Latin index means pointer, hence “index finger”). The aspiration towards “true” representation stems from this direct correspondence between object and representation. Lastly, AI is based on vast amounts of data. “Training” the AI means to point it to an image and ask: “Is this a cat?” and similar questions, then assign the answer with a yes or no tag (Steyerl describes this in her text as well, so you can refer back to it if you like). From droves of images thus labelled, those labelled as cats most often will constitute the peak of a bell curve.

Once prompted to produce an image of a cat, the AI will pick out that peak. Consequently, even though the representation may resemble a cat, there is no such thing in reality. Steyerl’s critique of Chiang’s description of AI as a “poor image” is thus: “Visuals created by ML tools are statistical renderings, rather than images…” (Steyerl 2023, my emphasis). This is a good time to allow students a few minutes to discuss among themselves whether AI “images” are images at all, or what we should understand under the term “image.” Calling it an “image” misses the technical transformation of representational media. In order to give students an impression of what this means, I tell them that before the fingerprint, criminal investigation often relied on eye witnesses or on torture for material that could be used in court. Likewise, Nazi Propaganda in Germany made ample use of the radio, calculating that it would establish a direct connection between the “Führer” and the people (Weckel 2023, 14–15). In both cases, the alleged (!) correspondence between representation and object brought about massive cultural and political transformations. Consequently, what the technological transformation of AI will mean for culture and politics at large is still anybody’s guess, but its scale can already be envisioned.

Of course, these forms mix. Analog photography requires manual development, which reintroduces symbolic aspects into the image. Color grading in digital film is often done manually yet again through computer software. In digital photography, aspects of the incoming light are automatically adjusted, such as the white balance—which can of course also be done by AI. Most importantly, the choice of framing and contextualization which imbues the image with meaning is often done by humans, but can also be generated by AI. The distinction is thus not absolute, but is nevertheless helpful to attune students to the differences between various media and the specificities of AI.

I then move back to Steyerl. Under the heading “the means of mean production,” she details the material conditions of the production of AI technology: “As Time magazine reported in January 2023, underpaid workers in Kenya were asked to feed a network ‘with labeled examples of violence, hate speech and sexual abuse’” (Steyerl 2023). “Is this a cat?” is an AI training question—but so is “is this rape?” Some people must feed an AI with their emotional suffering, provoked by violent content. This work is typically poorly paid and outsourced to minoritized people in or from the global south, a result of the depletion of resources and the impoverishing repercussions of colonial oppression. Factually, AI is based on both emotionally- and economically-exploited racialized, postcolonial labor. Although AI technologically differs from indexical media, it structurally regenerates the material and structural racism that we had seen in the case of photography earlier. Again, technology is a product of the society that makes it, and that technology will inevitably mirror these conditions and feed them back into its host society.

I then show them Rosa Wernecke’s five-minute video art piece Before the Drop (2025). Its first half is a screen recording, the second half is AI-rendered “lesbian porn.” The video starts with a prompt to an AI tool to generate an image of “two lesbians kissing.” It refuses. The artist then turns to ChatGPT to ask (in writing) why that is. We then learn how structural discrimination is built into AI: While allegedly ‘protecting’ viewers from indecent imagery, content regulation policies in fact suppress queer content and favor heterosexuality. Wernecke then has ChatGPT explain “lesbian erasure” as the “systemic ignoring, minimizing, or invalidating of lesbian identity,” hence prompting an AI to give a critique of AI policy. Next, Wernecke asks ChatGPT to script a lesbian porn that will not be censored by content regulation, which it does. The second half of the video then presents the resulting “lesbian porn,” which largely consists of AI renderings of flowers and fluids moving in surprising directions, followed by the words “It wasn’t gravity. It was desire,” to indicate that the depicted movements were driven by attraction.

I then ask students to discuss the video amongst themselves for a few minutes in the context of Lewis’s and Steyerl’s texts. In the ensuing conversation, I mark the various ways in which the anyways-discriminatory technology produces discriminatory content: AI relies on data, so one way to manipulate outcomes is to feed the AI only with a particular data diet. Another is to train it so as to recognize or reproduce only particular outcomes (hence to adjust the peak of the bell curve). A third is to prohibit certain outputs through filtering. In this way, AI regenerates existing injustices. Technology is again, and yet differently, a product of the society that makes it—and that technology will inevitably mirror these conditions (see deLire 2023). Yet it can also be used against itself. Wernecke enlists an AI as a collaborator to undermine the discriminatory logic of AI. Nevertheless, the AI imagery we see of plants and fluids is still stereotypical, which can be used in class to revisit the production of AI imagery through statistics as opposed to photographic images. But its framing, both visually and contextually through the introduction in the video, gives it a critical spin. Not all is lost.

To end the session, or as homework or a separate session, I recommend that you show your students Zach Blas’s six-minute video performance work Contra-Internet Inversion Practice #1: Constituting an Outside (Utopian Plagiarism) from 2015. The video is another screen recording, accompanied by the Le Tigre song “Get Off The Internet!” We see Blas copy-pasting portions from various revolutionary, feminist, and anti-capitalist texts into a one-page document. He then adjusts the text. For example, he replaces all occurrences of the term “capitalism” with the term “internet.” Curiously, the ensuing text still makes perfect sense—a fact that often startles students, who think of capitalism as tendentially negative, but of the internet as tendentially good and liberatory. Eventually, we hear a digital voice reading out the resulting “Contra-Internet Manifesto,” a call for digital network communication beyond the internet, a possibility that most students will not have entertained. Ensuing questions could be: What could digital philosophy look like, given the artistic strategies that Wernecke and Blas suggest? What would a contra-internet, contra-photography, or contra-AI look like? And how would they affect existing societies? I usually leave these questions to students for brief written reflections of one to two pages.

Sources and Suggested Readings

• Blas, Zach. 2015. Contra-Internet: Inversion Practice #1: Constituting an Outside (Utopian Plagiarism). New York: Rhizome.
• deLire, Luce. 2023. “Beyond Representational Justice.” In Trans Perspectives, ed. Christian Liclair, Luce deLire, Antonia Kölbl and Anna Sinofzig. Berlin: Texte zur Kunst.
• Lewis, Sarah. 2019. “The Racial Bias Built Into Photography.” New York Times Lens, April 25, 2019.
• Steyerl, Hito. 2023. “Mean Images.” New Left Review no. 140/141 (March–June).
• Weckel, Ulrike. (2023). Introduction: Media and their Users in Nazi Germany. In Audiences of Nazism: Using Media in the Third Reich, ed. Ulrike Weckel. Berghahn Books.
• Wernecke, Rosa. 2025. Before the Drop. Video artwork. Berlin: Goethe-Institut SOL Solitude festival. (Alternate link).

Suggested Readings for a Second Lesson

If you would like to use Blas’s work (and text) as a separate session, you can assign the texts that he uses for his video performance and discuss the wider theoretical framework of a possible contra-internet and contra-AI theory:

• Gibson-Graham, J. K. 2006. The End of Capitalism (As We Knew It): A Feminist Critique of Political Economy. Minneapolis: University of Minnesota Press. (pp. 1–11, 35).
• Jameson, Fredric. 1998. The Cultural Turn: Selected Writings on the Postmodern, 1983–1998. New York: Verso. (pp. 50–56).
• Marcos, Subcomandante. 2001. Our Word Is Our Weapon: Selected Writings. Edited by Juana Ponce de León. New York: Seven Stories Press. (p. 47).
• Preciado, Paul B. 2018. Countersexual Manifesto. Translated by Kevin Gerry Dunn. New York: Columbia University Press. (pp. 20–24).

---