In 2023, I wrote a short book on image generating AI — AI in art (visual) — with a collaborator which failed to find a publisher. We tentatively titled it Subprime Automation: Images of Human Generated AI. We stated its goal in our pitch document which I’ll reproduce in part to explain our intent:
The latest wave of AI image-generating systems, such as DALL-E 2, and Midjourney, launched last year, have propelled AI to the centre of public conversation. These systems can be confounding, obscured behind technical jargon or hyperventilating promotional spiel. In falling for the dazzling ‘novelty’ of these AI models and recycling shallow tropes, much of the recent commentary has, in fact, hindered our understanding and engagement with the issues at stake: how will these systems impact the production and reception of culture? What will they mean for workers and artists? How will they function in our cultural economy? We go beyond questions of artistic theft, which only strengthens notions of property exploitation, to understand the current economic position of the cultural worker and the impact AI could have.
I cannibalized parts of this for the first chapter of my self-published pamphlet A Cultural Reader of the Tarot, ‘We Can Generate It for You Wholesale: Tarot and AI’. A couple of the chapters more rightfully belong to my collaborator as standalone essays and I hope that they use them for something.
However, I can share one chapter here that was almost entirely written by me, faults and all, which has aged somewhat but may remain valuable as a snapshot into a certain period and the concerns of that period.
Who Benefits From Image Generating AI?
“…capital accumulated to the point that it becomes images.” – Guy Debord, Society of the Spectacle
‘Couple talking in bed’. Over 7,000 images match the search query (fig. 1). The couples are predominantly happy, young, and attractive, white and heterosexual, wearing white clothes, wrapped in white linen sheets in clean rooms suffused with light. The photos are well-composed, highly saturated, and lack any distracting background details. In most cases, the couples look into each other’s eyes, smiles reveal gleaming white teeth.
On top of the Shutterstock homepage, there is a ‘call to action’ banner: “Need a unique image? Create your exact vision in seconds with Shutterstock’s AI image generator.” Searching again for ‘couple talking in bed’, the results are very different. A kitsch geometric animation plays out and then we are presented with 4 quite different images (fig. 2). The first is a blue-tinged twilight scene of vague, misshapen body-parts on a bed in a 3D rendering style. The second looks much like the ‘real’ photographs in the previous search results, but with disturbing smeared faces, recognisable as the signature style of DALL-E 2. The final two images are a close shot of a smiling woman with a fragment of a man’s face above her, and another 3D bed with body parts.
Fig. 1, screenshot of Shutterstock ‘couple talking in bed’ image search results. 6 Feb. 2023.
Fig. 2, screenshot of Shutterstock AI-generated image search for ‘couple talking in bed’. 6 Feb. 2023.
Since the early days of advertising, it has been recognised that ad copy is more effective when lead by an image. Stock images are essentially decorative. Some of these images are singularly concerned with a subject, removing as much background and context as possible for maximum generality. Some are highly specific, even absurd. Most, though, attempt to strike a balance between the generic and specific, while retaining some character that marks out quality. This ‘character’ will often reference established trends and fashions in photography, film, illustration, and other art movements and popular culture.
Stock photography has its origins in the 1920s when American photographer H. Armstrong Roberts arranged a group portrait photoshoot and had the six models sign release forms so that he could sell the image. This became ‘Group of passengers waiting in front of ford tri-motor airplane’, one of the first images sold through his company, RobertStock. Later Otto Bettman saw the lucrative potential in stock images and organised his own large collection of photographs using the principle of a library card index, which enabled clients, such as advertisers, publishers and designers to find images efficiently. Bettman charged a fee for use of images in his archive, despite the fact that many of the images he sold were reproductions that he had not paid for, photographing the originals with his 35mm camera and selling his copies. Today, we see the ubiquity of stock images play out on social media, where users are bombarded with still and moving images in a torrent of advertising, news articles, blog posts, and memes. In the last 20 years, many of these images were bought and sold through Shutterstock, one the largest stock photography agencies in the world.
Shutterstock was founded in 2003 by Jon Oringer. The company was not his first start-up attempt, having had a reportedly entrepreneurial spirit in his youth and 10 start-up attempts before his success with Shutterstock. Over the next 17 years, he developed the company into a publicly traded behemoth, valued at 2.7 billion (as of February 2023), and is widely considered the first ‘self-made’ billionaire of ‘Silicon Alley’, New York’s answer to California’s ‘Silicon Valley’.
The business model was lifted from a tried and tested blueprint: undercut the competition on price and take their business. The kind of cheap stock photography he provided came to be known within the industry as ‘microstock’, a term of endearment within a culture of ruthless efficiency. Coincidental improvements in technology were essential to this strategy—Oringer’s timing was fortuitous. The early 2000s saw the proliferation of affordable DSLR cameras that drastically improved the quality of photos any non-professional could produce, and it was around the turn of the century when ‘Photoshop’ became a verb. The barrier to entry had been demolished and, while industry incumbents like Getty Images strived to hold onto larger licensing fees, Shutterstock capitalised on these new and improving technologies, offering attractive prices.
Shutterstock succeeded in the market due to various competitive advantages. Workers who contribute to the platform must be vetted for quality and consistency of work, but once registered on the system, the submission process is quick and efficient. Their payment structure before 2020 also encouraged network effects, with payments based on the popularity of a contributor’s images. Shutterstock offers a subscription for access to their images (which they do not own, but license), simplifying the payment process while building up a dependable revenue stream. Over the years, they have also struck many lucrative deals, for example, integrating into Facebook and Google’s advertising products and beating rival Getty Images in a partnership with the parent company of entertainment journal Variety.
In July 2021, Shutterstock announced its acquisition of image AI companies Shotzr, Pattern89, and Datasine through a new subsidiary called Shutterstock.AI. The purchases signalled the company’s increasing focus on AI. The platform had, however, been working with AI-enhanced features from as early as 2016 when they first offered users the ability to search using an uploaded image (known as a ‘reverse image search’). This had already become a feature in Google Search. The AI-specialised team behind Shutterstock’s new feature reportedly started working in 2015. By 2023, they had almost a decade of AI expertise to draw on.
In 2022, Shutterstock announced it would use OpenAI’s DALL-E 2 image-generating AI. The integration promised to provide generated images for sale to users alongside the human-made stock photography that has, until now, defined their business. Journalist James Vincent reported in The Verge that this was another step in an ongoing partnership with OpenAI, which had been granted access to use Shutterstock licensed images as part of its machine learning model training. In response to accusations of betraying their contributors, Shutterstock announced in October 2022 that they would compensate artists whose images were used in their AI training via a ‘compensation fund’. Rather than a profit-sharing or collaborative mechanism or even a worker employment contract, this compensation fund can be seen as a restitution of damages. Indeed, artists were not informed of their inclusion in an AI training dataset before the fact.
It may seem odd that Shutterstock should partner with a company that appears to compete with it. By the end of 2022, we could already see images generated by Midjourney, for example, used prolifically on the blogging platform Medium, where writers have typically used generic stock photography from Shutterstock and others. Photographers are taunted by sensationalist headlines such as “Shall I throw away my camera? These images are NOT photographs.” (Digital Camera World)
It may seem a case of “if you can’t beat ‘em, join ‘em”, but Shutterstock has, in fact, been instrumental in creating its would-be successor, training it on the corpus of high-quality, generic and—most significantly—popularly rated images it controls. It is a hedged bet against market shifts that threaten the core of its business. Rather than trying to compete with AI for stock image generation, it has embraced it.
In the end, Shutterstock has paid heed to the business adage to ‘sell the hole, not the drill’. Oringer stepped down as CEO from Shutterstock in 2020 but before leaving, he commented on the subtly changing focus that would become important to the company as it expanded its use of AI:
I think we’ve evolved into more of a tool over time. We are with you and helping you find the right image and drive the right outcome with that image, helping to discover content that will be more on-brand for your business and do things you were unable to do before. (Silicon Republic)
Shutterstock routinely remarks on its blogs that “AI will never be able to truly replace human creativity” (Source) and commits to building ethical AI. However, the ethical questions dealt with tend to be focused on diversity and coded bias rather than labour relations. The impact on workers is significant.
The application of technology to ever more facets of human life has created a pervasive ‘solutionism’, the political strategy that assumes all social problems have a technological solution. Social critic Evgeny Morozov coined this neologism in his 2013 book To Save Everything, Click Here. One of these social problems is the problem of undesired labour, resulting in the invention of many time- and labour-saving devices. In the mid-twentieth century, the ultimate futuristic domestic utopia was imagined as a world where robots would cook, clean, and play with the children—a society served by an army of mechanical ‘Mary Poppinses’, perhaps looking a little like Star Wars’ C3PO. In other words, the masses were encouraged to dream of attaining the life of the upper classes, where human servants, or even enslaved people, were replaced with tireless machine workers performing domestic labour. This kind of domestic work is, of course, highly gendered in the long cultural, social, and political history of ‘women’s work’.
More recently, the futurist image has been the magic of the automated supply chain, as goods and services are ‘streamed’ into the home or carried by flying drones. The origin and story of these products are ignored, since they are all made to a standard specification and are interchangeable. Our mass produced ‘ready meals’ defer cooking labour in the home to some unknown factory, so the idea of a robot humanoid cook in the kitchen is less relevant. We are entertained by watching television and playing video games that are piped in as data over the internet. Just as in the fantasies of robotic servants, the technophilic imagination obscures the low-skilled and precarious work that supports such services in reality today.
What labour would we not want to replace? While many people do, of course, love cooking and caring for their children, it is undeniable that the reduction of labour by modern outsourcing in both these areas has allowed adults to put their energies into other areas. Amazon’s voice assistant Alexa, for example, released an AI-generated bedtime story feature for children that “helps solve a problem many parents face at bedtime: a lack of creativity.” (ZDNET) That a generic synthetic voice from an electronic speaker would be a welcome substitute for reading bedtime stories—that most caring and intimate of activities between a parent and a child—is an indication of how alienated and lifeless the ideology of tech solutionism can be. It is another example of the creation of non-existent problems on which to apply developing technology.
Can we say any of this for the labour of the artist? Is artistic work the kind of work one would rather not do, given the choice? Conceptions of timesaving—of removing waste and increasing efficiency—are economic ideas, as they are imperatives within economies that require continual growth. As with many commercial drives, they exist in tension with artistic practices. Exceptional from other forms of commodity production, the creation of fine art resists conceptions of economic efficiency and ‘waste’—of materials, time, human labour, and space. In an apocryphal tale, a Japanese potter is asked how long it took him to make a ceramic bowl, and replies: “Thirty minutes and thirty years”.
Image-generating AI has not been designed to save time for the artist, or even to replace them. This AI was created first and foremost because it was possible. It is the unintended twin, the mirror image, of image classification software. Now, having been created, it is a solution in search of a problem. Techno-solutionists are trying to solve a problem that does not exist for artists—to replace their work with a terse text prompt—while claiming they are the intended beneficiaries.
Who, then, stands to benefit from image-generating AI?
Shutterstock is not alone in its integration of image-generating AI. Low-skill design tool Canva and Microsoft’s Designer for Office 365 have both integrated OpenAI’s tools. Picsart, another low-skill design tool, has also integrated an in-house image-generating AI into their product, with others sure to follow suit. (PetaPixel)
What is striking about these integrations is how the functionality is presented: as a search. Shutterstock, for example, interfaces with one set of its customers, those wishing to buy free or paid images, by providing a search engine of its licensable images. One way to present text-to-image functionality is as a text search, as we have been accustomed to since the earliest days of search engines, a service Google has had a monopoly on for some time, habituating us to this mode of user interaction.
However, this is a search of images not yet created, a kind of search of a vast space of potentiality. The technology does not explicitly conform to this conception, as it is a novel method for profiting off the massive, contextless ‘output’ of artists whose work has been made available online. But presented as a search, it is a search that cannot fail to find a result. This must constitute something of a holy grail in the business of search engines. Oringer indicated as much, discussing the challenges of searching Shutterstock’s massive image database. For him, it has always been central, that “[f]rom day one, search was a problem, and it got more difficult as it went.” (Source)
One popular way to test this search space promise is to combine unlikely terms. The results can be funny and unexpected. We could use the term ‘a giant monster with apples for eyes’ and be confident the subject of the image would match our expectations in some way. What most pundits commenting on this wave of low-effort surrealism failed to notice was that stock photography had already been doing this for years. A 2012 article in The Atlantic read:
…there may well be one occasion when one will find oneself seeking an image of a cat in smart clothes with money and red caviar on a white background. This being the Internet, actually, there will probably be two or three. For such occasions, when they arise, your best bet is to turn directly to an image service like Shutterstock. … Stock photographers specialise not just in imagery, but in sentiment prediction: They anticipate people’s needs before they become needs in the first place. (The Atlantic)
The year before that, blogger Edith Zimmerman sparked a meme with her post Women Laughing Alone With Salad, revealing the uncanny absurdity of the sanitised, highly gendered space of corporate imagery of modern stock photography. In 2015, this meme was used as the basis for a theatrical play that explores the negative impact of aspirational advertising, the male gaze, and power and empowerment. (VICE)
Some of the most compelling stock photography is by photographers that have learned this ‘sentiment prediction’ to exploit the cultural zeitgeist. One successful stock photographer described their process as striving for “a story in picture form. We could have popular photos if we were hitting on something that was already going on in the media.” (Vox) The AI system appears to delve into a similar corporate imagination to create images. However, this novelty overshadows the issue of quality and the politics of the image. We have our result, it matches the subject and is ‘good enough’; why should we look any further?
The economic imperatives associated with successful search results, where the search user finds what they were looking for, have spawned the large and fragile ecosystem of SEO—search engine optimisation techniques. SEO amasses an ever-changing set of practices to manipulate search engines and bring a particular business’s results ‘closer to the top’ of the ordered search list, traditionally arranged in order of a nebulously defined ‘most relevant’. These practices include adding metadata, such as popularly searched keywords, to webpages or, in the context of image search on Shutterstock, tags and keywords in titles on images. This push for relevance suits search economics as designed by the search provider since, if a user does not find an adequate result, they may leave the website for another provider, thus losing a potential customer. Indeed, the search term ‘a giant monster with apples for eyes’ did not result in any photography search matches on Shutterstock at the time of writing. Having failed to fulfil our strange request, might we look elsewhere? Now we can use the AI ‘search’, and while its generated images are of questionable quality, we will not now be confronted with an empty search page.
What will happen to the second set of Shutterstock customers—the creators of the stock photography the platform has built its profits upon—when they must compete with AI-generated image results? We might expect that the most ingenious creators will find gaps in what AI can currently provide to stay relevant and sell their work. But their work will only be fed back into the systems again to train the next iteration of data models, further closing off viable niches—bread-and-butter stock photographers will have been co-opted into training their own replacement.
What effect will AI image generation as search have on the perceived power of these search engines? The AI functionality can be trained on more input image data, but a crucial part of its training is the sorting of acceptable generated images from the bad. Shutterstock and others can refine their AI based on the selection of AI-generated search results, following in the footsteps of Google. In the broader context, then, we can expect to see some platforms dominate the market in a pattern now typical of what has been dubbed ‘platform capitalism’. (See Nick Srnicek’s book of the same name.)
In fact, with the landscape now clearly illuminated, we can see that while there is some genuine technological progress and novelty in the tool of AI image generation itself, there is no innovation in its application. What we see is the logic of surveillance capitalism (the corporate collection and commodification of personal data) and platform capitalism (the business model based on creating intermediary digital infrastructures, such as Google, Microsoft, Uber, and Airbnb), already by now very well analysed by authors Shoshana Zuboff and Nick Srnicek. Given this, it is more important than ever to see past the technophilic hype to the core dynamics of the issues facing workers and consumers.
Image-generating AI are services and are consistent with the ‘as a service’ business model that has proliferated in recent years. When products become services, this allows continuous rents to be collected on the assets that ultimately support services offered by their owners. For example, Microsoft used to sell their popular word processing software as a product, with a once-off purchase of a license to use the software. Today, Microsoft offers Word on a subscription basis—that is, as a service. The shift is in some ways subtle; where once it was considered a discrete word processing package, now it provides the service of word processing that users can access for a limited time. In hosting their service, Microsoft rents access to the infrastructure on which the software is sustained and delivered.
Similarly, the stock photography platform Shutterstock does not normally charge per photograph but operates on a subscription system. Even so, the company pays its contributors per image download, touting on its submission webpage that “[o]ver the last 15 years, we’ve paid out a billion dollars to our worldwide community of contributors.” (Source) This statement reveals the scale of the potential cost savings offered by automating visual content production while reducing the role of artistic labour. Image-generating AI is the ultimate ‘as a service’ system since it automates the production of images on demand. Once trained on source material—those artworks and images rendered as data—the original artists have no legal claim on the data model as their intellectual property. The AI systems and the data models they are based on can be proprietary, as in the case of DALL-E 2 and Midjourney, and thus effectively protected as a trade secret, though not all are, such as Stable Diffusion, which was publicly released. This leads to the question: from what asset is this service extracting value? Is it the disparate images it is trained on or the resulting data model?
Nick Srnicek describes this shift in his book Platform Capitalism:
[Platforms] have enabled a shift from products to services in a variety of new industries, leading some to declare that the age of ownership is over. Let us be clear, though: this is not the end of ownership, but rather the concentration of ownership. […] Far from being mere owners of information, these companies are becoming owners of the infrastructure of society. [Emphasis added] (p. 92)
As with Microsoft and their renting of Word, Shutterstock rent access to their infrastructure. But within this infrastructure is encoded the creative work of many photographers and artists who have lost any claim of ownership by this process of laundering. If ‘information wants to be free’, then platforms and services built on that freely available information want to be paid.
Some AI ethics researchers draw our attention to something missing from the conversation that is hiding in plain sight:
…there is no room [in contemporary AI ethics discussions] to discuss one of the central questions with AI—whether or not it should actually be built in the first place. Or should a problem which can be fixed with AI, actually be fixed with AI, or should it perhaps not be fixed at all… There is a dominant view among AI positivists that technological innovation always constitutes social progress. Yet this view ignores the politics of design and production. (Source, pp. 12-16)
Artists and photographers selling their work online for general use do not need image-generating AI. On the contrary, it competes with them, and their struggle trains it to beat them at the economy they are forced to participate in, as constructed by the tech platforms. By adopting AI to remain competitive, artists will confirm their place as individuals fighting over meagre budgets rather than as a craft profession acting collectively in solidarity to protect their livelihoods. Derivative image aesthetics will grow as artists apply only finishing touches to the products of the AI. As AI tools become commonplace, it will be significantly harder to opt-out of the use of these tools as an artist.
Consumers do not need image-generating AI, but they will quickly become used to it. Now that I can get an image of a giant monster with apples for eyes, I do not have to modulate my request, I just need to find the AI that gives the best results. Unknowingly, users continue to be used in training their systems to yield better results, as they have done for decades, raising the quality of these tools and causing further brand loyalty and lock-in. In the long term, this loyalty supports monopolistic strategies within the economic system, as the spoils go to the victor, and beyond that limits the scope for systemic change.
It is only the interests of capital that need this technology; the business owners, stock market gamblers and entrepreneurs—and their managers who are empowered by capital to work labour at the lowest possible cost. While many may value the human relationships they have with their artists on staff or contract, it will only take a hard economic downturn to make cutting these workers—who can now be seen as non-essential—seem like the only way to maintain profitability. As Martin Ford writes in Rule of the Robots, “[h]istory shows that the vast majority of job losses from the adoption of labour-saving technology tend to be concentrated in economic downturns.” (p. 179) In the boom-and-bust cycle of capitalism, downturns occur frequently, and workers cannot afford to be complacent. Unless a comprehensive set of protections are put in place, workers will be the last to benefit from the benefits of automation, and the first to suffer the downsides.