“Deepfakes” and the End of the Photographic Age

Note: This paper was written at a time when the state-of-the-art of image generation used GANs, not the diffusion or diffusion-transformer models that dominate in 2026. I have summarized the paper as it was published, but none of the key philosophical points is affected by a difference in model architecture. Claims to novelty should also be interpreted relative to 2020.

Generative Adversarial Networks (GANs) enable computers to generate images that are visually indistinguishable from photographs. The popular press calls such images “deepfakes” and has recognized the threat they pose to photographs as evidence. But little has been said about their potential aesthetic value, or about how true photographs will fare once the two cannot be told apart. I argue that GAN images are indeed an epistemic threat, but also a medium with significant potential for artistic expression. If GAN images and photographs become indistinguishable, photographs will lose their epistemic authority but continue to be valued aesthetically, much as paintings are today.

A GAN is a competing pair of…

A GAN is a competing pair of algorithms. A ‘generator’ turns random numbers called ‘seeds’ into images resembling those in a ‘training set’; a ‘discriminator’ tries to tell the generator’s images from the training set’s, and both improve by trial and error. Seeds can also be chosen deliberately to shape the image: one application turns written descriptions of birds into seeds for matching images.

The term ‘deepfake’ often refers to images…

The term ‘deepfake’ often refers to images or videos of people created for deceptive or salacious purposes. I use ‘GAN image’ more broadly, for all photorealistic images generated by GANs, regardless of their purpose or content.

Photographs are commonly thought to occupy dual epistemic and aesthetic roles. The average smartphone camera roll might contain pictures of important documents (epistemic tool) as well as colorful sunsets (artistic expression). But the assumptions behind this common view reveal a tension. Photographs are thought to be epistemically valuable because they are ‘mind-independent’: the appearance of a photograph is causally determined by the state of the world, independent of anyone’s intentions. Yet on a widespread view, art must be intentionally produced. So if photography can be art, photographs must be in some degree mind-dependent.

Two skeptical arguments resolve the tension by denying photographs any mind-dependence, and so any aesthetic interest of their own. On what I call transparency skepticism, most provocatively argued by Roger Scruton, a photograph is by nature a transparent depiction of reality: a reproduction of how something looked, not an interpretation of it. On Nigel Warburton’s mechanical skepticism, photography’s automatic processes interfere with the transmission of a human intention, making it less mind-dependent. Dawn Wilson (publishing as Phillips) instead proposes that a photograph is ontologically dual: a photograph comprises a mind-independent photo-image, fixed by the photographic process, and a mind-dependent photo-picture, a product of human intentions. On her view, a skilled photographer uses the camera and the photographed objects to create images with the visual properties they intend, so a photograph can fulfill a photographer’s intentions as much as a painting can fulfill a painter’s. Both skeptics are right only about the photo-image.

Scruton claims that “with an ideal photograph…

Scruton claims that “with an ideal photograph it is neither necessary nor even possible that the photographer’s intention should enter as a serious factor in determining how the picture is seen”. On his view, an ideal photograph is merely a means of seeing what it depicts, so an interest in it is an interest in its subject.

What about GAN images, then? On the epistemic question, Jonathan Cohen and Aaron Meskin hold that photographs carry information about what they depict. GAN images carry no such information: the faces at thispersondoesnotexist.com vary in age, gender, and skin tone, yet provide information about no actual person. But carrying information is distinct from being believed to carry it. On Cohen and Meskin’s account, what we believe about an image depends on the salient type under which we categorize it. We deem a photograph reliable because we categorize it as a photograph. But even though a veridical portrait painting carries information, we are inclined to categorize it as a ‘portrait painting’ and deem it unreliable. Which types are salient is, presumably, contingent. If GAN images become indistinguishable from photographs, photographs will cease to be a salient category, and both will likely fall under ‘photorealistic image’. Since that category contains information-devoid GAN images, it cannot be said to reliably carry information. True photographs would still carry information, but would no longer be believed to.

Cohen and Meskin define information as a…

Cohen and Meskin define information as a “probabilistic, counterfactual-supporting, connection between independent variables”. A photograph of a face will show a smile if and only if the person is smiling, whatever the photographer or viewer believes. This holds even if the photograph is taken by a computer-controlled security camera.

On the aesthetic question, I extend Wilson’s duality to GAN images, whose making involves three human choices, each with a photographic analogue. The training set fixes the output’s possible range of appearances, as a photographer’s choice of subject does (though much more narrowly). The seed fixes its specific appearance, as vantage point and camera settings do. And the curation of outputs is identical to a photographer’s curation of their best work: a wedding photographer may capture a thousand photographs but find only a few dozen satisfactory. If the photographic process is mind-dependent enough to fulfill artistic intentions, these parallels suggest that the GAN image process is too.

Transparency skepticism falls flat against GAN images, since its premise is that a photograph reproduces reality, and GAN images do not carry visual information or resemble any existing object. Mechanical skepticism applies with little to no modification, since GANs are even more automated than cameras. But so does Wilson’s reply: image-making processes, GANs included, can help artists realize their intentions, not just limit them. A skilled individual can create not just a GAN image, but a GAN picture. Barring totally novel objections, GAN images may soon join photographs and paintings on the walls of the art gallery.

Taking inspiration from André Bazin, I call the period in which photographs have occupied their dual epistemic-aesthetic role the Photographic Age. Before it, on Bazin’s account, painting “was torn between two ambitions: one, primarily aesthetic… the other… to duplicate the world outside”; photography “freed Western painting, once and for all, from its obsession with realism”. The impending shift in the role of photography will likely be similar to the effect that photography itself had on painting. The Photographic Age began with the invention of photography, and it may end with the invention of the GAN image. Two practical problems remain, one for each role. First, the loss of photography as a trusted epistemic tool must be mitigated. Second, for the sake of artistic innovation, we ought to keep our definitions of aesthetic value permissive.

A film theorist, who called the invention…

A film theorist, who called the invention of photography “clearly the most important event in the history of the plastic arts”. On his account, painters afterwards abandoned their imagined obligation to realism, ushering in a golden age of creativity.

Written in 2020, before the Cambridge B.A. and before diffusion models. Published in Aporia, the undergraduate journal of the St Andrews Philosophy Society, vol. XXI, 2021, where it received the top paper prize.

Questions, objections, or a better example? Write to me: patrick.pan@patrickpan.com.

“Deepfakes” and the End of the Photographic Age.