Photorealism versus photography. AI-generated depiction in the age of visual disinformation

ABSTRACT

In the spring of 2023, we witnessed a breakthrough in the development of AI-generated images accessible to the general public. Pictures of Pope Francis wearing a stylishly long, white puffer jacket or driving a motorcycle down a busy street went viral. The same thing happened with AI-generated pictures fabulating over what the press coverage of an imminent arrest of former US President Donald Trump would look like. Amnesty International used AI to generate images to mark the second anniversary of police violence against protesters in Colombia. Boris Eldagsen turned down an award for Best Creative Photograph from the Sony World Photography Awards, announcing that it had been generated with AI. Critical reactions in the public were not long in coming. The use of AI to generate images was discussed with sometimes shrill words and phrases such as fake, fraud, a threat to photography’s credibility, and fake news. This article seeks to intervene in this crisis-oriented debate by proposing three analytical moves: First, we need a concept of photorealism that is kept separate from the idea of photography. Secondly, we need a conceptual distinction between two basic functions of photography: depiction and detection. In addition to this primary distinction between image functions, the article proposes a third move to introduce a function-oriented genre concept. Through an interdisciplinary approach to photorealism, photography, and genre, these three analytical measures are presented and examined step by step and discussed through analyzes of concrete and recent examples of AI-generated and AI-enhanced publicly available images in today’s society. Today’s crisis-oriented public debate about AI images serves neither democracy, art, journalism, nor photography. The purpose of the article is to contribute a simple and, at the same time, useful analytical tool in these discussions about the relationship between photography and current and future image technologies.

KEYWORDS:

• AI-generated images
• photography
• photorealism
• depiction and detection
• visual disinformation
• fake news

At the Sony World Photography Awards in April 2023, the Berlin-based German photographer Boris Eldagsen won the Creative category of the Open Competition. The award is considered one of photography’s most prestigious honours. The winning image appears as a black-and-white depiction of a close relationship between two women, perhaps mother and daughter; the elder, gesturing with tender, sorrowful care for the younger. The artist shocked the jury by refusing the award. He explained the rejection by saying that the image was not a photograph but an AI-generated image. He submitted the work to the competition to provoke a debate about AI’s role in photographic art.¹

The case is not unique. In 2021, the Magnum Photos photographer Jonas Bendiksen published The Book of Veles, an apparently documentary book and photography project from Veles. This provincial North Macedonian town placed itself on the world map in 2016 as an epicenter for fake news production. The book is created in a complex combination of AI-generated text, photographed landscapes and outdoor spaces, and computer-generated images of people and animals placed into the pictures. This project was also intended to provoke debate, but the book was hailed instead as a purely documentary work—a perception that he did not immediately correct. Bendiksen did not stop at merely publishing the book; he sent the entire photo series and a full-resolution PDF of the book to the prestigious Visa pour l’image: International Festival of Photojournalism in Perpignan, France. In response, festival administrators offered Bendiksen an evening screening during the festival program. Despite Bendiksen feeling conflicted by this turn of events,² The Book of Veles’ screening at Visa pour l’image was carried out (2021). Bendiksen thought, in order to prove his point about the dangers of AI-generated images passing for documentary or real-world photographs, he had to undergo such a test. He figured the exercise was necessary for the fields of photojournalism and photo-documentary art.³

Between these two stunts, we have had an increasingly shrill discussion about using AI in many areas, including news and reportage. It comes on top of, or intertwines with, debates about deepfakes, cheap fakes, manipulated photographs, and apocalyptic warnings about post-truth society.⁴

At the end of March 2023, images of Pope Francis wearing a fashionably oversized Balenciaga puffer coat, driving a motorbike through a busy city street, or acting as a DJ at a nightclub, went viral on social media platforms like Reddit and Twitter. The informed and critical press published big headlines that the images were fake and that they were “computer-created forgeries of reality” (for example, see the 26 March 2023 issue of Forbes Magazine).⁵ Just a few days earlier, Eliot Higgins, the founder of the open-source investigative outlet Bellingcat, had posted some pictures of Donald Trump under arrest on Twitter. It was ahead of the expected indictment of the former president, and Higgins wanted to visualize it: What would an arrest of Trump look like if it were to be covered by the press? Just a couple of days later, The Washington Post reported that Higgins’ posts depicting an event that never happened had been viewed nearly five million times, demonstrating the ways in which AI can be used to spread falsehoods.⁶

The reactions to these AI-generated images are symptomatic of how the press reacts when AI-generated images go viral. They are immediately labeled as false, fake, fraud, lies, or manipulated, explicitly or implicitly contrasted with photography, which should presumptively be real and true. A few years ago, we witnessed a similar reaction to deepfake video, i.e. various types of AI-supported video manipulation.⁷ It cannot be ignored that AI-generated images and deepfake videos can have the potential for harm. But it is striking how the public discourse is characterized by panic and cries of lies and the undermining of truth and trust. These reactions also meet the examples above. Eldagsen’s image was called fake by, among others, CNN.⁸ The pictures of Pope Francis in a white puffer coat and in other improbable situations were called “the first real mass-level AI misinformation case” (see BuzzFeed 27 March 2023).⁹ Such a crisis-oriented public discussion serves neither democracy, art, journalism, or photography.

In this article, I will introduce two conceptual distinctions in this debate. First, we need a concept of photorealism that is kept separate from the idea of photography or camera-based information. Secondly, we need a conceptual distinction between depiction and detection. This analytical distinction is not reserved for photography but is very important for understanding what is at stake in discussions of photographic imaging. It is a basic analytical distinction concerning the operations of photographic images. In addition to such basic functions, images can fulfill other essential tasks that I will discuss with the concept of genre. With this function-oriented genre concept, I propose a third analytical move which I hope will contribute to a better debate on visual disinformation.

The three analytical measures are presented and examined step by step. Through an interdisciplinary approach to photorealism spanning computer graphics, film, landscape architecture and archaeology, the first section discusses important dimensions of the idea of photorealism in contrast to the concept of photographic images. The focus is then directed towards detection as a fundamental photographic function across academic cultures and various photographic practices. In the third and last section, a highly controversial case is presented where Amnesty International used generative AI imagery on the two-year anniversary of the police’s excessive use of violence against demonstrators in Colombia. The two analytical distinctions—photorealism versus photography and detection versus depiction—are supplemented here with a third analytical approach, a function-oriented genre concept. Using these three analytical measures, this case is analyzed and discussed, contrasting it with other current cases of AI-generated images.

Photorealism (versus photography)

The AI-generated images mentioned at the beginning of this article mimic photographs without being photographs. Imitating photography is nothing new; we have a word for it: photorealism. I will argue that we must (re)introduce photorealism as a critical term in these debates.

Photorealism is an aesthetic term that denotes a visual style that developed in Europe and the USA in the late 1960s in the art of painting. One can think of Richard Estes, known for his paintings of New York City scenes, Chuck Close and his massive-scale photorealistic portraits, or Audrey Flack, famous for her contemporary versions of still-life paintings. Photorealism is a style, an aesthetic strategy for imitating photographic images.

With the rise of digital imaging technology, computer graphics, and other digital equipment in the early 1990s came a renewed interest in photorealism, not only within the field of fine arts but across a variety of graphic media and forms of expression: cinema, advertising, games, and visualization of information in science and communication. Still, the images were rarely to be mistaken for photographs. Today, the situation is different: Research shows that it is becoming impossible to distinguish between a photographic or camera-based image and an image that mimics photography, even for experts (see Lehmuskallio, Hӓkkinen, and Seppӓnen 2019). Computer scientists are trying to develop models for how computers can learn to distinguish between photorealistic CGI and photographic images (Meena and Tyagi 2019). This research is extensive and complex and has been going on for a long time. Already in 2005, Siwei Lyu and Hany Farid tried to develop a method for differentiating between photorealistic and photographic images (Lyu and Farid 2005). Today, the production of images imitating photographs is rapidly increasing because many have easy access to relevant imaging tools (practical, technical, economic, and time-wise). Admittedly, one should be careful to consider this, in itself, as a problem. However, two conditions, in particular, suggest that easy access to image tools with the potential for deception can become problematic today.

One condition is that many people today have social media as their primary source of information. These platforms tend to dampen awareness of the sources of information, blur the distinctions between more or less reliable sources, and often have a business model that promotes not only negative but also improbable stories or “click-bait” at the expense of more sincere and balanced news. Moreover, in the last decade, we have witnessed active campaigns that seek to flood communication channels and public conversations with everything from irrelevant information to outright disinformation (Pomerantsev 2019; Rid 2020). In such a situation, images that look like photographs can be potentially harmful. Therefore, we need a more precise language and some analytical distinctions. All problems are not solved with the introduction of these terms, but we can define and sort a little better what the problem consists of and how we can contribute to reducing it.

When people create photographs, they often say they are “taking a picture”. But what does it mean to “take a picture”? The answer to this question varies throughout history, and several photographic technologies exist in parallel. Nevertheless, and to put it simply, “taking a picture” involves some fundamental processes by which a camera captures and records visual information (see, for instance, Nakamura 2006).¹⁰ It can be helpful to think of the process, at least for digital photography, in three parts: First, the image capture process begins when the light from the subject enters the camera lens and is focused on the image sensor. Second, the camera’s image sensor detects and converts the light that enters the camera lens into an electronic signal. Third, the electronic signal produced by the image sensor is transformed into a digital image that can be processed, edited, stored on the camera’s memory card, and shared using a computer or other device. We will return to this photographic image acquisition in the article’s next section. The purpose here is to contrast it with images that mimic being produced using such an image acquisition process without this being the case. This is what we call photorealism.

First of all, let’s state that photorealism is not photography but rather mimics a style: it is an aesthetic strategy for imitating photographic images. As James A. Ferwerda says in the widely cited article “Three Varieties of Realism in Computer Graphics” (2003), “When we speak of photo-realism in computer graphics, we usually mean that we want to create an image that is indistinguishable from a photograph of a scene” (2003, 292). Computer graphics are present in many disciplines, each of which highlight what is essential for their practice. Like many working with computer graphics, Ferwerda emphasizes light energy as a central point of discussion concerning conditions for photorealism. In a systematic literature review of photorealism, Lidiane Pereira, Wellingston C. Roberti, Jr., and Rodrigo L. S. Silva seek to identify the main characteristics concerning photorealism in Mixed and Augmented Reality systems. The purpose is to find research opportunities that can be further exploited or optimized (Pereira, Roberti Junior, and Silva 2021, 15). Photorealism is then mainly discussed in terms of light, shadows, radiosity, and reflection. Creators look at interactions between surfaces and light and assesses whether near and intermediate reflections are best suited to make the computer-generated image indistinguishable from a photograph of the same scene.

In film production, these discussions are different. According to Julie Turnock, it is common in special effects within film production to appeal to the sense that “it looks right”, a notion that has been surprisingly unexamined according to the author. Turnock rejects a concept of perceptual realism. Perceptual realism is a realism that is based on what the eye sees. Cinematographic realism, on the other hand, is based on what the camera sees. Rather than modeling its look on the real or phenomenal world, special effects’ digital techniques imitate the look of photography (Turnock 2012, 158 and 160). In a study of special effects in US blockbusters, Turnock seeks to identify “the historically contingent, technologically inflected building blocks of the dominant aesthetics of contemporary photorealism” (159).¹¹

There are several crucial points to be gleaned from this study. Firstly, Turlock emphasizes that photorealism mimics photography’s perceptual cues (158). “Effects artist technicians may start with something based in computation or science, but then this is almost always tweaked, stylized or transformed in an attempt to suggest how it would look as photographed”, writes Turnock (162). Building up an image to look like the scene has been photographed or filmed thus involves miming camera movements, zoom, and techniques such as lens flares and rack focus. According to Turnock, the CGI addition of lens flares is “a chief indication that digital designers of photorealistic special effects are more often than not referencing the cinematography of earlier films, not, as is often assumed, the perceiver’s actual visual experience of the world” (162).

Stating that photorealistic special effects reference the cinematography of earlier films also underscores another essential point in this article: that photorealism is a historical term. Turnock voices that “Notions of photorealism change and shift historically as new and different image capture (analog or digital) techniques and technologies become standardized or expected in filmmaking practice” (160). Moreover, this historicization of the phenomenon of photorealism marks a significant difference to the concept of perceptual realism. One may debate whether human perception is not also historically changeable, and in evolutionary terms, perhaps this can also be said about the human eye. Turnock nevertheless has a point when she claims that photorealism better describes the characteristics of the aesthetics of visual effects than perceptual realism because it is a historically contingent and changeable style that is “not dependent on a transhistorical biological eye” (160).

It is also worth noting that, according to Turnock (2012), today’s photorealism bears the stamp of a certain 1970s photorealism.¹² The crucial point here is that historical conditions for technologies and aesthetics do not always go in sync. Even if notions of photorealism change and shift historically, they do not automatically follow the latest photographic techniques and technologies. Today’s photorealism does not necessarily mimic what a camera does or can do today. It simulates an aesthetic that may come from another time or a particular dominant aesthetic tradition.

If this is correct, as Turnock claims, that today’s photorealism is not in sync with today’s photographic technology, then knowledge of the discrepancy between the histories of photorealism and photographic technology could contribute to today’s discussions about visual disinformation. Part of the problem is that we have so much experience with these stylistic conventions that we no longer see them. As Turnock points out, “Through constant repetition, we have been conditioned to accept this specific historical aesthetic as perceptually real” (2012, 158). It is all the more important to emphasize that photorealism is a style and not photography as such and that this style does not mimic what an observer would see if they had stood next to a camera when a picture was taken (see Lee and Pae 2018, 12), but that it mimics a photographic style that may have been cultivated in a specific time or within a distinct genre. It is a style that, with some effort, can be denaturalized.

Photorealistic lens flares and virtual hand-held cameras are, as Turnock points out, prominent signs of our acceptance of an often-evident stylization as “properly” photorealistic (Turnock 2012, 162). Another unmistakable example of how photorealism should not be confused with perceptual realism is the photo-stylistic technique often referred to as “sfumato”. In the article “Photo-fake Conditions of Digital Landscape Representation” (2018), Myeong-Jun Lee and Jeong-Hann Pae refer to “sfumato” (from Italian sfumare, “to fade” or “to evaporate like smoke”) as a technique that stretches back to the art of Renaissance painting, but which later made itself felt in photography. The authors characterize “sfumato” as a stylistic device to create a dream-like atmosphere in a picture space reconstructed within the picture plane (11). The blur effect often produces a striped afterimage that creates the illusion of movement, a result that is associated with motion blur filters in Photoshop (see Lee and Pae 2018, 13).

Lee and Pae criticize this photorealistic, stylistic convention within current landscape architectural representation because it tends to be used mainly to offer illustrations of not-yet-actualized designed landscapes (2018, 15). The problem seems to be that “The resulting final synthetic representation is then perceived by viewers as a photograph that seems to capture an actually existing landscape” (16). Lee and Pae do not specify why or how this would be problematic. But this is a well-known discussion within the development of architectural prospects, where many may be critical of using a photorealistic representation style. In Rendering as Critical Reflection: On the Visual Production of Architecture in China and the West, Celia He refers to this widespread rejection of photorealistic representations of architectural constructions that have yet to be built. She gives examples of how this rejection has been explained with the danger of misunderstanding between architects and clients. The architect seeks to reduce or moderate photorealistic rendering to avoid the client being surprised by the project’s final result, fearing that the customer will mistake a photorealistic prospectus for photographic documentation (He 2021, 27).

A similar critical discussion of photorealistic representation can be found within archeology and the dissemination of historical and cultural knowledge. In “Virtual Heritage: Exploring Photorealism” (Rahaman, Das, and Zahir 2012), Hafizur Rahaman, Rana Das, and Shehzad Zahir refer to a survey where it is stated that “The problem with photorealism is that people tend to think of such images as the truth about the past, and not just a version or what it could have been like” (cited in Rahaman, Das, and Zahir 2012, 193). Without endorsing such viewpoints, the authors point to researchers claiming that “a photo-realistic rendering of the same element would imply ‘historical truth’, which may not be desirable” (cited in Rahaman, Das, and Zahir 2012, 193). The problem here is not only a confusion of photorealism and photographic documentation but also a problematic notion that photography should be able to document an unequivocal and unbiased truth about the past.

Some have called for a more nuanced understanding of the advantages and disadvantages of photorealism in archaeological visualization. Among these, we find the information scientist Isto Huvila. He points out that photorealistic representations of prehistoric finds can give the impression that archaeologists have a more comprehensive knowledge than they actually have. In a large-scale discussion of archaeological photorealism, however, Huvila emphasizes that the conflict is not really about visualization technique but about the effective means to help the audience understand the certainties and uncertainties on which the visualization is based.¹³

Before I contrast AI-generated images more explicitly with photography, let me introduce the second analytical distinction we need to train in today’s discussions of visual disinformation: depiction and detection. Photorealism is a form of depiction—it is a visual style that mimics a photographic representation of a scene. However, depiction is only one of several functions of photography. I have argued elsewhere that there are two parallel paths throughout the histories of photography: depiction and detection (see Hausken 2020). When the histories of photography are told, however, there is a tendency that one function (depiction) takes precedence over the other (detection).¹⁴ As also stressed by Patrick Maynard in his 2017 article “Photo Mensura”, “photography is almost universally misunderstood in terms of certain of its products: pictures of things” (2017, 41). In the following, we will look closer at the other basic function that characterizes photographically generated information: detection.

Detection (versus depiction)

Photography is a means of detection. By that, I mean that when one considers an image as photographic, it is always possible to direct one’s attention to the visual information the camera has recorded. The same applies to photographic images created without a camera by exposing a photosensitive material to light. It also makes no difference whether it is a photographic still or moving image. A photographic image enables the detection of the visual information recorded in the production process. One may consider a photograph from many other interests. However, the point is that such an interest is possible. It is possible because photographic images involve the recording of light. I have previously described photographic image acquisition as a process where light from the subject enters the camera lens and is focused on the image sensor, which detects and converts the light into an electronic signal, which is in turn transformed into an image that can be processed, edited, stored, and shared. This process does not give a simple or direct access to what was in front of a camera when the image was taken. On the contrary, it involves complex processes and transformations.

That photography is a means of detection is most evident in science, diagnostics, surveillance, border control, and sports photo finish. To better grasp this basic concept of photographic detection, I will first take a closer look at some of these practices. Let me emphasize that this concept is critical to understanding today’s discussions about AI-generated images that mimic photography.

Let me take a leap back in history and compare photography with the discovery of X-rays and natural radioactivity. When Wilhelm Conrad Roentgen discovered an unknown form of radiation in 1895 (the X-ray), it was not a point for him to create a picture of something (e.g. the bone structure in his wife’s hand). The radiation was the important element under investigation, as well as being able to register it (see Farmelo 1995, 89; Hausken 2007, 33). Working with a cathode-ray tube in his laboratory, Roentgen observed a fluorescent glow of crystals on a table near his tube. The table was replaced with a photographic plate so that the radiation could be registered (see Wilder 2009, 50). Shortly after the discovery of X-rays, another form of penetrating rays was discovered. In 1896, French scientist Henri Becquerel discovered natural radioactivity. Becquerel researched the principles of fluorescence, where certain minerals glow (fluoresce) when exposed to sunlight. He used photographic plates to record this fluorescence (Wilder 2009, 58–65). In both cases, a photographic method was used for investigating radiation. In both cases, this involved not only detection but also the representation or visual display of radiation. Unlike the widely publicized image of Frau Roentgen’s left hand, few would characterize the visual display of Becquerel’s discovery as a depiction. We need, in other words, more analytical distinctions between image functions than either detection or depiction. I will therefore return to a genre-oriented concept of function later in this article. Still, combinations of these two basic functions are also evident for photographic technologies. This is perhaps most pronounced in scientific photography. As Kelley Wilder points out in Photography and Science, “Like a canary in a coal mine, photography is often called on to do work in areas where human senses fail. […] To those who would dismiss photography as merely a detection device, it can only be pointed out that canaries too have their uses outside the field of mining instruments” (Wilder 2009, 53). Photography can often have multiple functions in parallel or over time. It can, for example, be a detection device in one context (e.g. an experiment), be printed as an illustration in a scientific magazine or an advertisement, and then be hung on a wall or in an online image gallery as a form of art (see Wilder 2009, 53). Nevertheless, discussing whether an image is generated by a camera involves clarifying whether it can be a detection device.

Often, the interest in using camera-based information as a tool for detection will require further image processing. In border control, for example, agents may use a camera or a scanning device to capture a facial image (also known as a probe image). In this case, the device will detect the face as such in the captured image and extract it from the larger picture. The system will then normalize or standardize the probe image to be in the same format (size, rotation, etc.) as the images to which it will be compared (see Hausken 2020). The normalized face image is then passed to the facial recognition software. As Lucas D. Introna and Helen Nissenbaum explain in Facial Recognition Technology (2009), “This normally involves a number of steps such as extracting the features to create a biometric ‘template’ or mathematical representation to be compared to those in the reference database (often referred to as the gallery)” (Introna and Nissenbaum 2009, 11). As the example shows, the complexity can be significant. The point here is to demonstrate that detection is a particular interest with which to encounter photographic images; It involves considering the photograph as a manipulable recording that can be examined more closely, precisely as a recording.

What then is the relationship between the photographic capacities for detection and depiction? Let me again stress that in photography, detection is key. Sometimes camera technology is used without even creating an image. In reverse vending (or container recycling) machines, multiple cameras are used today, providing a 360° view to analyze beverage containers from any angle when returned. The purpose is to achieve reliable detection of pre-approved shapes and evaluation of barcodes printed on beverage packaging (see Kokoulin and Kiryanov 2019).

When discussing images generated by cameras, the information is transformed into a visual expression. Still, it is not necessarily a depiction, a figurative image, or a visual representation of a scene. It can be abstract art. Many techniques go into recording an abstract photographic image, like macro lenses or multiple exposures (Rossbach 2011). Abstract fine art photography is nevertheless generated in the same way as other photographic images. And its character of being a recording of certain lighting conditions is vital for the experience that the exhibition of the picture facilitates.

Even in photographic depictions, the photograph’s detection functions can be superior to the depiction function. A clarifying example of this could be biometric passport photography. When people find that they do not look like their passport photo, it is not necessarily just because the picture was generated automatically by a machine. It is also because the metric properties and, thus also, the detection capacities of the registration and verification systems are superior to the depiction function (see Hausken 2020).

It seems reasonable to assume that photography’s capacities for depiction and detection are combined and balanced in everyday interaction with photographs. We point to a photograph and say, for example, that there is Uncle Arthur, even though it is just a picture of Uncle Arthur. The picture is both a depiction of this person and a tool to detect the unique features of this person. One function does not necessarily dominate over the other (in contrast to the passport photo example, above). This balance between basic functions can persist even if the photograph has been processed and beautified. But as long as I can examine this picture of my uncle to determine whether he resembles my mother, whether that shirt was the one I gave him last year, or whether he is photographed in his old apartment, I consider this picture a photograph. As Patrick Maynard points out, photographic images are often both depictions of objects and channels for detecting them (Maynard 1997, 128).

I think Maynard is right when he claims that “photography is almost universally misunderstood in terms of certain of its products: pictures of things” (2017, 41). At the same time, it is as if this misunderstanding is acknowledged, but we don’t have a language for it. In today’s discussions about AI-generated images and manipulated photographs, it is clear that photographs are not considered just any form of visual depiction. As we see in the public discussions of the AI-generated imagery presented in the introduction to this article, the notions of photography and image are often used synonymously. We see this, among other things, in that AI-generated images are called fake images. It appears as a simplified and awkward distinction between the true and the false, the real and the fake as if we had suddenly forgotten discussions about framing or cropping and filters and bird’s eye view and frog’s perspective and everything else that can be considered part of the rhetoric of the photographic image. The language comes across as panicky and against better judgement.¹⁵

So far, I have suggested two analytical distinctions useful in today’s AI imagery and visual disinformation debates. I have proposed introducing a critical concept of photorealism which must be kept separate from photography or photographic information. Second, we need an analytical distinction between depiction and detection and knowledge of how these two basic functions can interact. Along the way, I have also mentioned other functions that images can have (such as a portrait function or an illustration function, which I will return to below). As a third analytical move, we need a concept of genre with a particular focus on the image’s function. In the remainder of the article, I will elaborate on this genre concept and discuss some current examples using these three analytical measures. I hope they can contribute to a more informed and balanced discussion about AI-generated images and visual disinformation. First, let me introduce a highly controversial case where the use of AI-generated images created a great debate among the public, which we will look at in more detail and discuss against other current cases of AI-generated images.

The question of genre

In April 2021, thousands of people took to the streets in dozens of cities across Colombia to protest economic inequality, police violence, unemployment, and poor public services. Most demonstrations were reportedly peaceful, but according to the publication Human Rights Watch, the police responded using excessive, often brutal, force, including live ammunition (9 June 2021).¹⁶ In the spring of 2023, on the occasion of the two-year anniversary of the police’s excessive use of violence against demonstrators, Amnesty International, a human rights advocacy organization, published four AI-generated images as part of a campaign to highlight police brutality in Colombia. Amnesty International faced massive reactions from the press and human rights organizations worldwide. Critics argued that using AI-generated images could damage the credibility of human rights defenders and set an alarming precedent. It was stated that the AI-generated images portrayed fictional people instead of representing actual protesters and argued that such a practice undermines the organization’s core mission. There was also concern about ripple effects on smaller human rights organizations. Critics argued that such a trend could inadvertently harm smaller, local groups that are already under enormous pressure and are often targeted by their governments to discredit their work.¹⁷ This is an argument we know as “the liar’s dividend”. It refers to a practice that makes it easier for liars to avoid accountability for things that are actually true (Chesney and Citron 2019, 1758; de Ruiter 2021, 1320). In other words, the fact that a global human rights organization such as Amnesty International uses AI-generated images can contribute to undermining the ability to document actual abuses, the critics claim.

Amnesty International had access to photographs of the events in Colombia and had previously collaborated with photojournalists and videographers to document on-the-ground media that captured the true horrors of police brutality in the country. The argument for not using photographs in this campaign but instead AI-generated images was that they wanted to protect the anonymity of vulnerable protesters and avoid displaying actual protesters’ faces. They later regretted this choice.¹⁸

We will take a closer look at these images below, but first, let me compare this project to a critically acclaimed documentary film project that also used AI to protect individuals vulnerable to police brutality and who risk torture and murder if their identities are revealed.

In 2020, Welcome to Chechnya was released, a documentary film about the work of a group of activists who smuggle young people out of Chechnya who are persecuted and threatened with torture and death because of their sexual orientation. The film was more or less edited and finished in 2018, but to protect those involved from being recognized, they worked for almost three more years to develop a technique to hide their identities. Within documentary films, there are many techniques for this. One can distort voices and modify images of faces in various ways. But the filmmakers wanted a method that could maintain a sense of contact with the persecuted youth and enable the audience’s empathy. They found such an opportunity in what is often referred to as deepfake technology. Twenty-three volunteer LGBTQ+ activists from Western countries agreed to lend systematic video recordings of their faces to protect the characters in the film. A solid dataset was developed for each of the volunteers and used to disguise corresponding Chechens in the film. The documentary opens with an explicit reference to this. Specific stylistic actions have also been made so the audience is kept aware that the filmmakers have taken measures to protect those involved. Considering this work started in 2018 and was completed in 2020, it is a remarkable effort and pioneering work worth noting in a field where technological development is moving exceptionally fast. The work effort is overwhelming, and comparing Amnesty’s Colombia campaign with such a project may seem unreasonable. When taking a closer look at this campaign, we need a further distinction from those I have introduced so far. For projects such as Welcome to Chechnya and Amnesty’s campaign in Colombia, developing a concept of function indicated by genre appears suggestive.

Both projects use AI to protect vulnerable individuals. But where Welcome to Chechnya wants to depict empirical individuals in a way that means they cannot be detected, while at the same time safeguarding their humanity and enabling the audience’s empathy, it is more unclear what Amnesty wants to achieve with their depictions beyond marking the second anniversary of the 2021 protests and preventing detection of individuals who protested.

As to style, it can be debated whether the photorealistic style of the four Amnesty pictures is of such a nature that the images can be mistaken for photographs. Confusion may perhaps be possible on small screens with low image resolution. One thing that may have contributed to the adverse reactions, however, is that one of the images, in particular, is close to a specific photographic genre: journalistic reportage photography. When looking up these four images, one might be struck by the many critical comments that highlight this particular image. All four pictures have verbal language text printed across the image, but the text on this one may indicate that it is to be the campaign’s front image. The text has a larger font; Its content points forward in time, and the text fragment ends with three dots indicating a continuation: “Why the Colombian police need comprehensive reform…”. The text on the other three images has a smaller font and an ascertaining text in the past tense, indicating what the police did to the protestors two years earlier. It says: “The police used force to punish people who protested peacefully in Colombia.” “Police officials raped and insulted women and LGBTI people taking part in protests.” “They misused less lethal weapons, like tear gas, blinding young people who were raising their voices.” In addition to the fact that the image most critics seem to have reacted to may be considered the campaign’s front page, it is striking how the other three pictures seem to belong to a different genre, a genre that is perhaps not as provocative when it comes to the question of whether photorealistic AI images can contribute to undermining the photograph’s credibility, the political poster.

Political posters usually originate from one of two sources: governments or civil organizations. Non-governmental poster sources include political parties, labor unions, and civil and human rights organizations (see Tschabrun 2003, 307). Since the Second World War, the political poster has transformed into a communication modality used in political protests (Tschabrun 2003, 306). As Susan Sontag points out in a 1970 article, “A poster claims attention—at a distance. It is visually aggressive” (quoted in Sontag 1999, 196). The political poster communicates quickly but ephemerally in public spaces (see Tschabrun 2003, 303). It is not unusual for political posters to experiment with alternative styles of political discourse (see, for instance, Tschabrun 2003, 304). Considering Amnesty’s spring 2023 campaign in Colombia seems reasonable in light of such a form of communication.

Let us first look at the three images I would argue link to the political poster tradition. Each image depicts an individual in a close-up or semi-close-up view. Two of them are seen in frontal view: a young woman in front of a sea of flames, the other, a young boy wearing a spartan helmet. This second image is in a green sepia hue and is dominated by something that seems like a red blob of paint over the picture surface just above the image of the boy’s face. Incidentally, many other technical-aesthetic distortions in the picture make one think, for example, that it could be a montage. The third picture also shows a young man with a helmet, which has a visor. This picture shows the figure in profile in front of smoke and flags raised above a sea of police officers; you can almost only see their helmets. It is unclear whether the man in the profile is supposed to be a policeman or one of the protesters. Regardless, none of these three pictures are to be considered portraits.

A portrait depicts an individual as an individual. In the tradition of the bourgeois portrait, the person portrayed can also present themselves, seeking to offer their self-image (see Bourdieu [1965], 1990, 82). Often, the person will be named. Publication of a photo portrait in the press would usually be motivated by two conditions: visual identification or what is known as a journalistic portrait (Kędra 2016, 38). The primary purpose of the former is to show what a person looks like, while the latter is also to tell something more about the life of the person concerned. None of the three pictures from the Amnesty campaign discussed above fall within these portrait genres. On the other hand, they fit nicely into the subgenre of the political poster that promotes iconic images of anonymous heroes.¹⁹

As Sontag points out, political posters often rely on the image rather than the word, just like commercial posters. The aim of an effective political poster is rarely more than the stimulation and simplification of moral sentiments, and the classic means of stimulating and simplifying is through a visual metaphor. “Most commonly, a thing or an idea is attached to the emblematic image of a person”, emphasizes Sontag (Sontag 1999, 203). The political poster often promotes a heroic figure, be it a celebrated leader of the struggle or an anonymous representative citizen, such as a soldier, a worker, a mother, or a war victim (Sontag 1999, 203). It is the latter we see in the three Amnesty images, most clearly in the two with a frontal view. They are depicted as anonymous, representative, protesting citizens, such as a soldier or worker, a woman or representative of the LGBTQ+ community, a political rape victim. Where the person in commercial posters must be attractive, more is needed for the political poster that appeals to emotions with more ethical prestige, Sontag underlines. The imagery of political posters cultivates a sense of obligation (Sontag 1999, 203).

The remaining image in the campaign that Amnesty International presents as a cover image is different. It shows a young woman panting toward the viewer with a distressed facial expression, held on either side by uniformed and armed policemen, followed by a small army of police officers in what looks like a narrow city street. The text above the image reads, “Why Colombian police need comprehensive reform…”. The police’s brutality towards the demonstrators is marked verbally in the three posters. But in this front picture, it is as if we see the action while it is taking place and which will continue if no one intervenes with comprehensive reforms; that is how we must read this.

Arguably, this image could also be considered part of a political poster tradition. Political posters use a variety of emotional appeals to create a sense of psychological or moral obligation. Sontag emphasizes that the image on posters focusing on one exemplary persona can be heartbreaking. She also refers to a variant of that single model figure poster where the image “depicts the pain or struggle itself, juxtaposing the heroic figure with the figure of a dehumanized or caricatured enemy” (Sontag 1999, 204). This characteristic could fit Amnesty’s cover image. But according to Sontag, such tableaus often show the enemy defeated or fleeing. “Compared to posters that only show exemplary figures, posters with images of agony usually appeal to coarser emotions, such as revenge and resentment and moral complacency,” she claims (Sontag 1999, 204). Such a potential is difficult to see in this case. The oppressor has not been overcome; the picture does not celebrate anything. Unlike the other three images in the campaign, the central person does not appear as a hero, icon, or symbol of a political struggle. In contrast, this picture looks more like an image from another type of visual series, the journalistic reportage.

Photojournalism encompasses many genres that are categorized in several ways. Some typologies are thematic, based on the subject area (news, sports, entertainment) (see, for instance, Kędra 2016, 29). Others are more function-oriented (like Kędra 2016), and these are most interesting in our context. Kędra suggests distinguishing between four journalistic photography genres: news photography, reportage photography, portrait photography, and illustrative photography (Kędra 2016, 33). I have already mentioned how portrait photography in the press, according to Kędra, is usually motivated by two conditions: visual identification or journalistic portraiture (Kędra 2016, 38). Let me briefly comment on illustrative photography before concentrating on the first two categories in Kędra’s photojournalistic genre typology.

The illustration is often considered the oldest visual genre in the press, much older than photography (see also Kędra 2016, 39). Illustrations can have many functions, be it a cover image to sell a magazine, indicate a topic, or visualize a point. For this genre, it is not uncommon to use stock photography (Kędra 2016, 49). If we relate stock photography to the basic functions of photography, I argue that stock photography is typically used to depict something, not to detect it. In passing, let me suggest that in cases where stock photography can be considered journalistically sound, one should ask whether one could just as well use AI-generated images and mark them, as for stock photography, with the image being considered an illustration. This argument will not normally be valid for news and reportage photography.

Kędra characterizes the aim of news photography as providing visual information about events of topical significance. It is about capturing the right moment and answering the same five questions as news stories should: what happened, where, when, how, and why. The caption may provide some of this information (Kędra 2016, 33). The subject of reportage photography, on the other hand, is usually less topical than news photography. At the same time, its documentary character tends towards conveying opinion or creating atmosphere instead of delivering current news (see Kędra 2016, 35). A photo reportage may stir up the viewer’s emotions more than any other photojournalistic genre (Freeman, 2011; according to Kędra 2016, 36). Kędra suggests this may be because the main topic of photo reportage is a human being in the context of news and events (see Kędra 2016, 36).

To the extent that it seems reasonable to associate the AI-generated cover image of Amnesty International’s Colombia campaign with photojournalistic genres, news photography and reportage photography are close candidates. Since the Amnesty campaign deals with a two-year-old news story, marks the distance in time verbally and aesthetically to that prior event, and appeals to political commitment and future action, I will argue that the campaign’s cover image points to a greater extent in the direction of the journalistic photo genre called reportage photography than it does to the more news-oriented journalistic photography. Both journalistic photo genres will be controversial to imitate with the help of generative AI. However, I would assume that AI-generated depictions of what appears to be a news story will be even more challenging for the credibility of the press than an emotionally oriented reportage of events that lie back in time. At the same time, images in the photojournalist reportage genre are often invested with difficult emotions. One should, therefore, think carefully before using AI-generated images to fill any of these functions, not least if the images can be mistaken for photographs associated with one or both journalistic genres, i.e. if they are photorealistic and otherwise imitate one of these two journalistic genres.

Boris Eldagsen’s Sony Award-winning photorealistic AI-generated depiction may, at first glance, appear as an image at the other end of a scale for photorealistic images that mimic photography without being photographs. It is an art project, even submitted in the creative category. Perhaps it doesn’t matter if it’s AI-generated? It does not mimic news photography, photo reportage, or documentary photography. It is presented as art. If we are concerned with visual disinformation, maybe it doesn’t matter?

However, it is not that simple. The photographic image’s forms of appeal are more expansive than its documentary potential. It also has an existential dimension, a possibility to evoke a sense of an existential moment, a moment of human presence, a presence that has been there. Roland Barthes did claim that art was one of the ways in which photography could be tamed, that this kind of existential disturbance was kept in check, broken down by the art photographer’s intention to create a thoroughly controlled work of art (Barthes, [1980] 1993, 117). I would nevertheless argue that photographic art can have such an existential-aesthetic quality and that Eldagsen’s image flirts with such a quality. By depicting a black and white close-up of two women, the elder grieving, perhaps caring, with a hand on the shoulder of the younger, who with a clear if mournful face, looks up at an angle, both with hairstyles and appearances that point in the direction of the 1940s, and with some technical distortions in the image surface that can easily be associated in the direction of old analog amateur photographs, Eldagsen has not only created a photorealistic AI image but an image that mimics an existential photographic moment that might be highly valued in a family album. Photorealism comes in several genres, and this one can easily be invested with a sense of the existential—a past presence.

When the Sony Award-winning image is then revealed to be AI-generated, it’s no wonder if the viewer feels cheated, made a fool of; it is a hoax; what you felt was just sentimental, the artist played you. This is a dimension of photography’s potential that is rarely discussed these days. It may not be disinformation as such, but we should not ignore the impact of the feeling of being deceived in the discussions of disinformation.

The possible experience of deception may also be important for understanding the critical reactions to Jonas Bendiksen’s project, The Book of Veles. This may be the case in connection with the screening at the Visa pour l’Image festival in particular, where even the organizers were not informed about the nature of this project. Like Eldagsen, the Magnum photographer wanted a debate about photography and artificial intelligence. I will not go into it here, as his project is one of some complexity, but I emphasize that we need such debates. And I have tried here to contribute to some analytical distinctions that may be useful in these debates.

Conclusions

Today’s public discussions about AI-generated images show a poor understanding of the complexities of this rapidly emerging problem. Images are characterized as true or false, real or fake, based on whether they can be considered photographs or not, as if the notions of image and photography were synonymous and as if photography guaranteed some form of unequivocal meaning. This conception must be against our better judgment: every photojournalist knows that a press photograph is never neutral or unambiguous: every choice one makes (framing, angle, shutter speed, etc.), as well as all coincidences over which one has less control play a role in the image’s rhetoric. The rapid development of photorealistic pictures generated with the help of artificial intelligence has created some challenges that we as a society must take seriously. But we must be equipped with a better analytical apparatus than simple distinctions between truths and lies. In this article, I have suggested three analytical measures. Firstly, I propose to introduce a critical concept of photorealism that is kept separate from the idea of photographic or camera-based information. Secondly, we need an analytical distinction between depiction and detection and knowledge of the relationship between these two basic image functions. Thirdly, we need a concept of genre to discuss which image functions are particularly important when assessing when, where, and how artificial intelligence can be included in image production. I hope that with the implementation these proposals, we can have a more qualified discussion about visual disinformation, democracy, art, journalism, and photography.

Disclosure statement

No potential conflict of interest was reported by the author.

Correction Statement

This article has been republished with minor changes. These changes do not impact the academic content of the article.

Additional information

Funding

The work was supported by the research project “PHOTOFAKE – Visual Disinformation, the Digital Economy and the Epistemology of the Camera Image” funded by the Research Council of Norway (RCN) (project number: 316376). https://prosjektbanken.forskningsradet.no/project/FORISS/316376?Kilde=FORISS&distribution=Ar&chart=bar&calcType=funding&Sprak=no&sortBy=date&sortOrder=desc&resultCount=30&offset=0&Fag.2=%C3%98konomi

Notes

1. https://www.eldagsen.com/sony-world-photography-awards-2023/

2. https://www.magnumphotos.com/arts-culture/society-arts-culture/book-veles-jonas-bendiksen-hoodwinked-photography-industry/.

3. “The Book of Veles project became my own little visual Turing test.» Bendiksen states. https://www.magnumphotos.com/arts-culture/society-arts-culture/book-veles-jonas-bendiksen-hoodwinked-photography-industry/, accessed April 20, 2023.

4. The term “cheap fakes” is usually traced back to Britt Paris and Joan Donovan’s 2019 report Deepfakes and Cheap Fakes: The Manipulation of Audio and Visual Evidence, a white paper from the independent nonprofit research organization Data & Society. “cheap fakes” is used for technically simple manipulations of video, such as the 2019 widely circulated altered video of Nancy Pelosi, then US Speaker of the House of Representatives, that, due to a moderate slowing of the tempo, making Pelosi seem to be slurring her words and appearing lethargic or drunk.

5. Matt Novak, “That Viral Image Of Pope Francis Wearing A White Puffer Coat Is Totally Fake”, available: https://www.forbes.com/sites/mattnovak/2023/03/26/that-viral-image-of-pope-francis-wearing-a-white-puffer-coat-is-totally-fake/?sh=216b41b91c6c [accessed April 20, 2023]

6. Isaac Stanley-Becker and Naomi Nix, March 22, 2023 “Fake images of Trump arrest show ‘giant step’ for AI’s disruptive power”. Available: https://www.washingtonpost.com/politics/2023/03/22/trump-arrest-deepfakes/ [accessed: April 20 2023]

7. This spawned not only similar discussions in the press but also several white papers and other policy documents that inform responses to the challenge more broadly, among them the WITNESS’s report Mal-uses of AI-generated Synthetic Media and Deepfakes (2018) and Paris and Donovan’s (2019) report Deepfakes and Cheap Fakes.

8. https://edition.cnn.com/style/article/ai-photo-win-sony-scli-intl/index.html [accessed July 3, 2023]

9. https://www.buzzfeednews.com/article/chrisstokelwalker/pope-puffy-jacket-ai-midjourney-image-creator-interview [accessed July 3, 2023]

10. Cameraless photographs also involve a form of photographic image acquisition. For the sake of simplicity, I refer here to the camera and camera-produced images. By doing so, I do not mean to exclude photographic image acquisition created by manipulating light to leave an impression on a photo-sensitive surface from what I, with Patrick Maynard, will call a family of photographic technologies (Maynard 1997).

11. Turnock investigates, more specifically, the importance that the effects company Industrial Light and Magic (ILM) has had in this story. It should also be mentioned that Turnock does not clearly distinguish between special effects (SFX) and visual effects (VFX) in this article. Special effects are often considered practical effects, i.e. those created on set (for example, a controlled explosion in an action scene). Visual effects, on the other hand, are made in the editing bay during post-production.

12. Largely due to the historical dominance of one special effects company, Industrial Light and Magic (ILM), according to Turnock (2012), the look of photography is strongly characterized by a particular photorealistic special effects aesthetics ILM developed in the 1970s into a powerfully convincing house style. “Given the prominence of ILM in the film industry, denaturalizing the ILM aesthetic is crucial to understanding how digital images evoke ‘authenticity’ or ‘veracity’” argues Turnock (abstract to article, 2012).

13. Isto Huvila discusses photorealism in archeology as a monstrous structure (in Donna Haraway’s sense). It would take too long to fully cover this discussion here, but I would like to refer the interested reader to an article that is highly worth reading. See Isto Huvila (2021) “Monstrous hybridity of social information technologies: Through the lens of photorealism and non-photorealism in archaeological visualization.”

14. One can of course argue that photographs have many other functions than these two, but it is this distinction that will be in focus here. See Maynard (1989), (1997), and (2017) for in-depth discussions of several basic functions of photography.

15. There are also research articles in which the terms fake photography and fake image are used synonymously. See, for example, Dagar and Vishwakarma (2022), and Kasra, Shen, and O’Brien (2018). Of critical analyses of the somewhat panic-like and dystopian social debate, I would particularly highlight Broinowski (2022), Yadlin-Segal and Oppenheim (2021), and Wahl-Jorgensen and Carlson (2021).

16. The police brutality was covered in many media worldwide. See, for example, Human Rights Watch, June 9, 2021. https://www.hrw.org/news/2021/06/09/colombia-egregious-police-abuses-against-protesters [accessed July 2, 2023].

17. See for instance “Human Rights Organization Faces Backlash for Using AI-Generated Imagery” May 4, 2023 at culture.org, available: https://culture.org/human-rights-organization-faces-backlash-for-using-ai-generated-imagery/ [accessed May 16, 2023]

18. https://www.nrk.no/norge/amnesty-trekker-tilbake-ki-bilder-og-beklager-1.16394409, [accessed May 30, 2023]

19. At least, this applies to two of the three pictures. In the third, where it is more ambiguous whether the young man depicted should be considered a protester or police officer, the genre also appears more unclear.