Composite Memory: How AI Rewrites What We Remember
AI-generated images are not fake or real. They are composite memories — statistical aggregates of cultural value, made manifest.
A Photographer's Guide to the Theory Behind Generative Images
When Boris Eldagsen submitted an AI-generated image to the Sony World Photography Award in 2023 and then refused the prize, the media reduced the event to a simple binary: real photograph versus fake AI. But David Bate, in his landmark essay "AI Photography and Composite Memory" (Bate, 2026), argues that this framing misses the point entirely. AI images are neither fake nor real. They are, in his formulation, "statistical samples, aggregates of cultural value made manifest in an image" — a new category that demands an entirely new theoretical vocabulary.
This essay traces that vocabulary through the recent special issue of photographies (Volume 19, Issue 1, March 2026), titled "Photography & Memory in the Age of AI," and connects it to the broader theoretical landscape that informs how we — as photographers and AI visual artists — should think about the images we make with generative tools.
1. The End of "That-Has-Been"
For over a century, photography's claim to truth rested on what Roland Barthes called the "that-has-been" (ça-a-été) — the physical trace of light on a photosensitive surface, guaranteeing that something was once there before the lens (Barthes, 1980). AI-generated images shatter this guarantee. Nothing was ever in front of any camera. Instead, these images are statistical renderings extracted from latent spaces — mathematical compressions of millions of training images into high-dimensional vector fields (Yacavone, 2026).
Fred Ritchin names this new image category "desirents" — visualizations not of "the way things are" but of "how one wants things to be," mapping what he calls "territories of unknown origin" within high-dimensional vector spaces (Ritchin, 2025). Kathrin Yacavone prefers the terms "virtual photographs" and "possible memories" — artifacts that visually mimic the aesthetic conventions of the camera while lacking any physical referent in the material world (Yacavone, 2026).
These are not merely taxonomic distinctions. They reshape fundamental questions about evidence, memory, and cultural transmission — questions that become urgently practical the moment you open Midjourney and type a prompt.
2. The Psyche as Photographic Archive
Bate's theoretical point of departure is an insight from Jacques Derrida. In a 2000 conversation about photography, Derrida argued that human perception is already technical: "Within perception there are already selection, exposure time, filtering, and development." The psychic apparatus, Derrida claimed, functions "like" — or perhaps "as" — a photographic archive (Derrida, 2010; cited in Bate, 2026).
The slippage between "like" and "as" is deliberate. Derrida suggests that the human mind and the photographic archive are not merely analogous; they share the same structural operations: selection, filtering, inscription. The question of inscription within the psychic apparatus had long been central to Derrida's engagement with Freudian psychoanalysis (Derrida, 1978), and Bate builds directly on this lineage. Just as every archive is selective — police archives select for criminal identification, family albums for happy moments, medical archives for health markers — human memory selects according to the topography of consciousness, preconsciousness, and the unconscious (Bate, 2026; Laplanche and Pontalis, 1988).
The arrival of generative AI adds a third term to Derrida's analogy. If the psyche operates like/as a photographic archive, then AI's latent space constitutes a new kind of archive — one that is neither subjective (like memory) nor physically indexed (like photographs), but statistically aggregated from culture at large. Bate frames the central question: "If human perception cannot be separated from technical images, what are the consequences of generative AI images for human memory? Or conversely, what is the effect of human memory on this generative artificial image production model?" (Bate, 2026).
3. Latent / Manifest: A Model Across Three Domains
The most powerful conceptual tool in Bate's essay is the latent/manifest couplet, which he traces across three domains:
In analog photography, exposed film carries a latent image invisible to the eye until chemical development renders it manifest. In Freud's dream theory, latent thoughts (day residues, unconscious wishes) are transformed through the dreamwork — condensation, displacement, visual representation — into the manifest content of the dream (Freud, 1900/1980). In AI image generation, the latent space (statistical compressions of training datasets) is activated by a human prompt and processed through algorithmic computation (denoising, diffusion) into a manifest image (Bate, 2026; Somaini, 2025).
The parallel is not merely illustrative. Bate argues that all three domains share the same structural logic: invisible sources are transformed through a process of selection and synthesis into visible images. The specifics of the process differ — chemistry, unconscious desire, algorithmic computation — but the latent-to-manifest arc is consistent.
4. Galton's Ghost: Composite Photography and Its Legacy
This structural parallel reaches back further than either Freud or digital computing. In the 1870s, the statistician Francis Galton developed composite photography: a technique of superimposing multiple portrait photographs onto a single photographic plate through repeated partial exposures, producing a single "average" face that revealed "typical" features of a group (Galton, 1878; Porter, 1986).
Galton himself noted that the composite portraits looked "better" than any of the individual constituents — precisely because the averaging process erased individual irregularities (Galton, 1878; cited in Bate, 2026). This observation has a direct contemporary resonance: when Midjourney generates a portrait, it too produces a statistically averaged face, one that tends to be smoother, more symmetrical, and more conventionally attractive than any real human face. The cliché is not a bug; it is a feature of statistical systems.
Galton's composites were later entangled with his eugenics project, and Allan Sekula's foundational critique "The Body and the Archive" (Sekula, 1986) established the terms for understanding how photographic archives can function as instruments of power and social control. Bate summons Galton not to rehabilitate eugenics but to expose the risks inherent in any system of statistical image generation: the erasure of individuality, convergence toward the mean, and the amplification of dataset biases (Bate, 2026). Catherine Malabou extends this genealogy from Galton through genetics, cybernetics, and epigenetics to contemporary AI, arguing that "the brain and the computer are in a reciprocal and mirroring relationship" and that naturalistic resistance to the technological capture of intelligence is meaningless (Malabou, 2019; cited in Bate, 2026).
Freud himself cited Galton directly. In The Interpretation of Dreams, describing a dream figure that combined his uncle's face with a friend's yellow beard, Freud wrote: "It was like one of Galton's composite photographs" (Freud, 1900/1980; cited in Bate, 2026). He called this process "condensation" — the compression of multiple latent elements into a single manifest image. Bate's central argument is that AI image generation performs a structurally identical operation, minus the human unconscious. The critical difference: in dreams, unconscious desire drives the condensation; in AI, "the human prompt gives semiotic activation and desire to the generative apparatus" (Bate, 2026).
5. The Archive Transformed
If AI images are composite memories drawn from cultural archives, then the status of those archives matters enormously. Aleida Assmann defined the archive as a space of "passive remembering" — a repository where traces of the past remain in a "state of latency" until recalled (Assmann, 2008). But generative AI transforms the archive from a passive storage facility into what Roland Meyer calls an "operative image archive": historical data is no longer preserved but "scraped" and "mined" to fuel neural network training. The archive's latency becomes commodified as generative potential (Meyer, 2023; Yacavone, 2026).
Hito Steyerl names the outputs of this process "mean images" — visual representations of the statistical mean or average of training data (Steyerl, 2023). These are more reductive than even Galton's composites, because they have no contact with physical reality whatsoever. Daniel Palmer and Katrina Sluis analyze this as "the automation of style," where a photographer's historically situated aesthetic is reduced to an algorithmic parameter (Palmer and Sluis, 2024).
Bernard Stiegler warned about the broader social consequences of such automation: "Digital automation short-circuits the deliberative functions of the mind," producing what he called "systemic stupidity" — a functionally drive-based mode of cultural production that replaces reflection with reflex (Stiegler, 2016; cited in Bate, 2026).
6. Five Case Studies in Composite Memory
6.1 Boris Eldagsen: Reverse-Engineering Postmemory
Bate reads Eldagsen's The Electrician not through the media's "real vs. fake" narrative but as a case study in composite memory. Eldagsen's father, born in 1924, enlisted in the German army at seventeen and, like most of his generation, never spoke about the war. After his death, Eldagsen discovered photographs from the 1940s and began collecting similar images from flea markets and eBay. Using DALL-E 2, he synthesized new images that visualize the silent decade of his father's youth (Parshall, 2023; Bate, 2026).
Bate connects this to Marianne Hirsch's concept of "postmemory" — the transmission of traumatic experience across generations through stories, images, and behaviors (Hirsch, 1997). Generative AI offers the possibility of reverse-engineering this process: using historical images to generate new images of a past that was never visually documented. The resulting images are not "remembered" but "re-membered" — a bricolage of fragments that never originally belonged together, assembled into a composite that functions as artificial memory (Bate, 2026).
6.2 Ahn Jun: Materializing Verbal Memory
Korean artist Ahn Jun's 2023 project used Midjourney to generate 307 images published as a photobook in Japan. During his studies in early-2000s Los Angeles — before smartphones, before ubiquitous photography — he had no images of that period. He transformed stories and anecdotes heard from people he met into prompts, materializing "the life in California he had missed" as dreamlike images. The publisher describes this as "the materialization of imagined scenes through an AI image generator — itself a form of dreamlike imagination" (Ahn, 2024; Bate, 2026).
Bate reads both projects as "experiments in new memory-work using computers," analogous to the photo-therapy that predated the digital age — projects that invite audiences to "imagine other pasts, presents, and futures" (Bate, 2026).
6.3 Elena Efeoglou: The De-historicization of Style
Elena Efeoglou's exhibition Blurring Reality and Fiction — August Sander meets AI (2025) reinterprets August Sander's People of the Twentieth Century through AI. Sander's original project was a visual taxonomy of Weimar-era German citizens classified by occupation and social role — a project whose printing plates were destroyed by the Nazis in 1936 (Sander, 1986; Efeoglou et al., 2025; Yacavone, 2026).
Yacavone identifies three dynamics at work. First, a drive toward individualization: Efeoglou assigns names and fictional biographies to Sander's anonymous subjects. Second, a counter-pull toward typification: AI's statistical averaging reasserts generic "types" despite the artist's efforts at individuation. Third, the de-historicization of style: Sander's New Objectivity aesthetic — a product of specific social conditions in interwar Germany — becomes a detachable "metastyle" when translated into a Midjourney prompt parameter (Yacavone, 2026; Meyer, 2023; Palmer and Sluis, 2024).
6.4 Exhibit A-i: From Evidence to Testimony
The Exhibit A-i: The Refugee Account project (2023) addresses human rights abuses in Australia's offshore detention centers on Nauru and Manus Island — sites where the government has systematically blocked physical access, making photographic documentation impossible. Based on testimonies from 32 refugees, the project used Midjourney to generate 130 images, which were then uploaded to Shutterstock alongside conventional photojournalistic assets (Braeunert, 2026).
Svea Braeunert's analysis centers on the shift from evidence to testimony. These images lack physical causality and cannot function as legal evidence. But they can function as testimony — a subjective, conceptual reconstruction of lived experience made visible. Notably, the AI-specific artifacts (distorted fingers, painterly textures, expressionistic distortion) function not as technical failures but as ethical devices that visualize unspeakable suffering (Braeunert, 2026).
Yet this raises Ariella Azoulay's concern about the "civil contract" of photography — the ethical interaction between subject, photographer, and viewer that traditional portraiture presupposes (Azoulay, 2008). AI-generated images of vulnerable populations operate in the absence of this contract, as the Amnesty International controversy over AI-generated images of the Colombian protests made clear (Braeunert, 2026).
6.5 Eyes That Never Looked Back
Sara Oscar and colleagues give this theme its sharpest formulation. The eyes of AI-generated portraits are composite averages — statistical means synthesized from thousands of faces in the training data. In traditional photography, a real person looked into the lens, and that exchange of gazes formed the ethical foundation of the portrait. The eyes of an AI portrait have never looked at anyone. And yet we are moved by them — Ritchin calls this a "highly flawed but potentially interesting" simulation of humanity (Oscar et al., 2026; Ritchin, 2025).
7. Bate's "Hands" Experiment
In a revealing practical experiment, Bate entered the same English prompt — "create a picture of two hands" — into ChatGPT and Ideogram, comparing the results (Bate, 2026).
ChatGPT produced an image referencing Michelangelo's Sistine Chapel ceiling (1510), translating the iconic gesture into a stock-photo style. When asked for a non-religious version, it offered a secular "fist bump." Ideogram, by contrast, generated multiple images with diverse ages, cultures, genders, and skin tones, though with characteristic AI errors (incorrect finger counts, gratuitous symbolic elements like lavender flowers).
From this experiment, Bate draws three conclusions. First, different systems activate different latent archives, producing dramatically different results from identical prompts. Second, all results look "photographic" but periodically violate what Michel Foucault called the "discursive regularities" of photographic realism (Foucault, 1985; Bate, 2026). Third, and most consequentially, the responsibility for meaning shifts to the user — not as "creator" of images but as designer of their compositional meaning, representational ethics, and aesthetic effects (Bate, 2026).
8. What This Means for Us
Bate's concept of "composite memory" synthesizes everything above. AI-generated images are algorithmically condensed composites of existing cultural archives, manifested as artificial memories that never previously existed. This concept draws on the structural parallels between Galton's composite photography (1870s), Freud's dreamwork condensation (1900), and AI's latent-space computation (2020s) (Bate, 2026).
As Yacavone concludes, even though composite images lack indexical truth, they can — through artistic and activist intervention — enter the archive ecosystem as traces of "cultural reference memory." The very "a-historicity" of AI images may paradoxically become a historical mark of our era, testifying to the anarchival properties of the digital transition for future researchers (Yacavone, 2026).
Bate's final sentence deserves to be read slowly: "If humans leave 'photography' behind (the logic of photography as an aspect of subjectivity), they encounter something else: the machine-work images of algorithmic culture. In whatever practice, it is what humans see that matters to social forms of existence" (Bate, 2026).
For those of us who work with Midjourney, Flux, DALL-E, or any generative image tool: the prompt is not a search query. It is an act of memory design — a conscious decision about which cultural memories to summon from the latent space, which to exclude, and what new composite memories to bring into existence. The theoretical frameworks assembled here — from Derrida's psychic archive through Freud's condensation to Steyerl's mean images — are not academic luxuries. They are the operating system for a practice that aspires to be more than the production of statistically averaged clichés.
References
Ahn, J. (2024). Until You Left Off Dreaming About. Tokyo: Case Publishing.
Assmann, A. (2008). "Canon and Archive." In Cultural Memory Studies, ed. A. Erll and A. Nünning. Berlin: de Gruyter.
Azoulay, A. (2008). The Civil Contract of Photography. New York: Zone Books.
Barthes, R. (1980). Camera Lucida: Reflections on Photography. New York: Hill and Wang.
Bate, D. (2010). "The Memory of Photography." photographies 3(2).
Bate, D. (2026). "AI Photography and Composite Memory." photographies 19(1), 125–144.
Braeunert, S. (2026). "From Evidence to Testimony." photographies 19(1), 103–124.
Derrida, J. (1978). "Freud and the Scene of Writing." In Writing and Difference. London: Routledge.
Derrida, J. and Richter, G. (2010). Copy, Archive, Signature: A Conversation on Photography. Stanford: Stanford University Press.
Efeoglou, E. et al. (2025). "Elena Efeoglou Interviewed." In Artist Meets Archive #4. Cologne: Internationale Photoszene Köln.
Foucault, M. (1985). The Archaeology of Knowledge. London: Tavistock.
Freud, S. (1900/1980). The Interpretation of Dreams. Harmondsworth: Penguin.
Galton, F. (1878). "Composite Portraits." Nature 18, 97–100.
Hirsch, M. (1997). Family Frames: Photography, Narrative and Postmemory. London: Harvard University Press.
Laplanche, J. and Pontalis, J.-B. (1988). The Language of Psychoanalysis. London: Hogarth.
Malabou, C. (2019). Morphing Intelligence. New York: Columbia University Press.
Meyer, R. (2023). "The New Value of the Archive." IMAGE 37(1), 100–111.
Oscar, S. et al. (2026). "Looking at Eyes That Never Looked Back." photographies 19(1).
Palmer, D. and Sluis, K. (2024). "The Automation of Style." Media Theory 8(1), 159–184.
Parshall, A. (2023). "How This AI Image Won a Major Photography Competition." Scientific American.
Porter, T. M. (1986). The Rise of Statistical Thinking: 1820–1900. Princeton: Princeton University Press.
Ritchin, F. (2025). The Synthetic Eye. London: Thames and Hudson.
Sander, G. ed. (1986). August Sander: Citizens of the Twentieth Century. Cambridge, MA: MIT Press.
Sekula, A. (1986). "The Body and the Archive." October 39, 3–64.
Somaini, A. (2025). "A Theory of Latent Spaces." In The World Through AI. Paris: JBE Books.
Steyerl, H. (2023). "Mean Images." New Left Review 140/141, 82–97.
Stiegler, B. (2016). Automatic Society, Volume 1. Cambridge: Polity.
Yacavone, K. (2026). "Virtual Photographs, Possible Memories." photographies 19(1), 59–79.
