The question of intelligence: a short history of AI art

In May 2023, more than 350 executives, researchers, and engineers working in artificial intelligence (AI) signed an open letter released by the nonprofit organization Center for AI Safety declaring that mitigating the risk of human extinction from AI should be made a global priority. The statement crystallized various tensions surrounding AI technologies at a pivotal moment in their evolution. From 2021 to 2023, several large-language models (LLMs), allowing the processing of vast textual data sets via AI and text-to-image models, generating visuals by means of natural-language prompts, had been made publicly available and had an impact on a broad range of sectors, from commerce and politics to art and entertainment. The tensions that lay the foundation for considering AI as a global societal risk are complex and have a long history. They include the dichotomy of seeing technology as the carrier of either human salvation or human extinction, the question of assigning responsibilities when it comes to establishing guardrails for technological developments, and the challenges in defining sentience, intelligence, and, consequently, what it is to be human. Over the decades, AI art has critically addressed these tensions and challenges in relation to AI technologies as they have emerged, testing and modifying systems and highlighting their characteristics and inherent values.

It is notable that the signatories of the open letter included top executives from leading AI companies, including Sam Altman, the chief executive of OpenAI; Demis Hassabis, the chief executive of Google DeepMind; and Dario Amodei, chief executive of Anthropic, founded by former members of OpenAI. The very companies that had developed the technologies they were cautioning against had decided to release and sell them despite their own concerns about their implications and scalability. This warning, therefore, needs to be considered as a calculated move in both positioning the companies as the potential savior, and negotiating responsibilities for AI’s potential consequences by asking for governmental and legislative regulations to protect corporate AI. Not surprisingly, concerns about the extinction of humanity were hardly mentioned when the Trump administration announced plans for the Stargate Project, a $500 Billion investment in AI infrastructure, in the first weeks of 2025.

Artists have used AI to experiment with its potential and impact since at least the early 1970s, fifteen years after the field was formalized at the 1956 Dartmouth Summer Research Project on artificial intelligence. They have probed its creative possibilities and engaged with its ethics and biases, as well as its effects on ecologies and labor, often by developing their own hybrid models and architectures.

Artistic practice in the field gained new momentum with the launch of OpenAI’s large-language-model-based ChatGPT, which launched on November 30, 2022, and text-to-image tools such as Open AI’s DALL-E (2021) and DALL-E 2 (2022), Midjourney Inc.’s Midjourney (2022), and Stability AI’s Stable Diffusion (2022). The hype surrounding text-to-image models immediately led to a polarized discussion, with the claim that AI would replace artists for good, on the one hand, and the dismissal of these tools as insta-kitsch engines that couldn’t produce anything of aesthetic value, on the other. While the former position seems to lack the aesthetic vocabulary necessary for evaluating art, the latter ignores artists’ more sophisticated engagement with these AI tools. Both tend to reduce AI art to visuals created by means of simple text prompts. To center the conversation, one needs to consider the evolution of AI art and trace the shifts in artists’ approaches to collaborating with, changing, torquing, and/or critiquing AI systems. At the core of such art lies the ability of humans and machines to acquire and apply skills and knowledge, raising the question of what the encoding of “intelligence” might mean for being human. In more recent years, artists also have looked beyond the relationships between anthropos, computer hardware, and computer software to consider how AI might move beyond anthropocentric models for knowledge creation. In his exhibition Distributed Consciousness (2024) at Gallery QI in San Diego, artist and creative technologist Memo Akten, for example, uses the cognition of cephalopods, which have the majority of their neurons distributed across their body rather than a central brain, as an inspiration for an AI-generated manifesto, juxtaposing decentralized natural with synthetic intelligence.

The term “AI art” is commonly understood to designate art that employs AI technologies in its creation, but this conception warrants further scrutiny. It fails to make distinctions between the use of AI as a tool as opposed to a medium and neglects the art’s conceptual engagement with AI. AI art can be defined as a subcategory of digital art or computational art, which uses digital tools and media to create and contextualize artwork: it incorporates technologies of artificial intelligence as both a tool and medium, engaging with it both practically and conceptually. The employment of a simple text prompt to generate a visual by means of a corporate text-to-image software does not automatically turn the resulting image into AI art. The use of AI as a medium — engaging its inherent systems and characteristics — in ways informed by a conceptual approach distinguishes a work as AI art. Computational art and its aesthetics are established and ever-evolving fields. As a subcategory of generative computational art, AI art requires a continuous reassessment of its models and expression. The following will focus on key moments in the history of AI art, tracing how it has developed in the context of technological developments and investigating the potential of art in general to contribute to the critical discourse that has developed around the aesthetics of — as well as the cultural, socio-political, and ethical impact of — AI technologies.

Collaboration is at the core of the earliest artificial intelligence program for artmaking — one of the longest-running, ongoing projects in contemporary art, Harold Cohen’s (1928-2016) AARON. An established British painter, Cohen began exploring the potential of software for artmaking in 1968 when he became a visiting lecturer at the University of California San Diego (UCSD). He officially named his program AARON in 1973, after being invited to the Artificial Intelligence Lab at Stanford University. ¹ Contrary to today’s statistical AI, trained on large data sets of images, AARON was symbolic AI that encoded “knowledge” about drawing and composition on the basis of rules that Cohen wrote in programming languages. AARON does not entail any of the standardization, averaging, and optimization used in the current models, which have been trained on massive data sets of existing images. Instead, AARON was shaped by the aesthetics of Cohen, who kept developing the software until his death and experimented with shifts in the style of work — from simple evocative shapes to figures and jungle-like environments — and a move from monochrome to color output. From the 1970s to 1990s, Cohen built his own drawing and painting machines that plotted and painted AARON’s creations, and in the 2000s, he switched to purely screen-based presentations.

After the first artistic explorations of AI represented by AARON, AI art evolved in three major phases from the 1990s to the early 2020s, shifting its focus to explore and investigate technological developments as they emerged.

In the 1990s and the early 2000s, artists created AI systems that critically engaged with the emergence of software agents — applications running automated tasks on the Internet for filtering or imitating humans—as well as chatbots. The problematic aspects of the software-driven filtering of information and encoding of human communication, as well as the “personality” of bots became an active area of exploration. Lynn Hershman Leeson’s chatbot Agent Ruby, released in 2000, explored chatbots as essentially social beings — autonomous characters with a life of their own — while Peggy Weil’s MrMind (1998–2014) was created with the specific intent of highlighting the differences between humans and machines. Rebecca Allen’s The Bush Soul series (1997–1999) explored communication between users and autonomous creatures in a virtual environment and laid the groundwork for more recent projects involving AI-driven life forms, such as Ian Cheng’s artworks.

The 2010s saw a shift in artistic AI practice, responding to a new stage of big-data analysis and neural networks — dating back to the 1920s and now benefitting from big-data processing — as well as the emergence of Generative Adversarial Networks (GANs) and Generative Pre-Trained Transformers (GPTs). Artists increasingly addressed the biases in big data sets and the ethical issues resulting from algorithmic processing. The socio-political dimensions of pattern recognition and apophenia, the perception of a meaningful pattern between unrelated or random things,² played a major role in critical discourse and art exploring large-scale data sets. Stephanie Dinkins decidedly countered the idea of the benefits of big data with a decidedly small data set and examined AI in the context of race with Not The Only One (N’TOO) (2018) – a sculptural AI storyteller trained on data supplied by three generations of women from an African American family, drawing attention to a drastically underrepresented data set. Dinkins also explored data sets and the process of an AI’s learning in Conversations with Bina48 (2014 – present). The work documents the artist’s ongoing conversations with Bina48 (whose name derives from “Breakthrough Intelligence via Neural Architecture, 48 exaflops per second”), an intelligent “social” robot modelled after a Black woman and built by the Terasem Movement Foundation. Dinkins’ Conversations with Bina48 explores what identity, race, and kinship mean to an artificial intelligence and whether we can form sustained relationships with the increasing number of non-human entities surrounding us.

While AI art’s engagement with Generative Adversarial Network (GANs) and Generative Pre-Trained Transformers (GPTs) can vary significantly in its focus, the key issue of these forms of statistical AI is always the automation of image and text generation through data sets with their embedded biases. GANs are learning neural networks in which generative algorithms trained on a specific data set generate new original images with the same characteristics as the training set and are then evaluated by discriminative algorithms that, based on their own training, judge whether the newly produced data looks authentic. GANs, in particular, led to an explosion of art projects exploring their potential for image generation on the basis of specific training sets and aesthetic goals. Less successful works stayed on the level of the “imitation game”, probing GAN aesthetics and the capabilities of AI software to reproduce images in a familiar period style. By contrast, Mary Flanagan’s [Grace:AI] Origin Story (2019) focuses on the aesthetics of using a deliberately gendered data set. Flanagan trained a GAN on thousands of images of paintings and drawings by female artists only, then tasked the software to create its “origin story” by looking at 20,000 online images of Frankenstein’s monster and producing its portrait. [Grace:AI] both alludes to Mary Shelley’s feminist critique of artificial life and male-dominated creation in Frankenstein and explores whether a gendered training data set produces a distinctive style.

Artists not only investigated the aesthetics of AI tools, they also started responding to the rise of paradigms of environmental management, engineering, and strategic intervention. Tega Brain’s Deep Swamp (2018) humorously critiqued environmental optimization in the form of a triptych of semi-inundated environments of wetland life forms governed by artificially intelligent software agents with different goals; while Asunder (2019) by Tega Brain, Julian Oliver, and Bengt Sjölén tested the potential benefits and pitfalls of an AI-controlled, fictional “environmental manager” that proposes and simulates future alterations to the planet to keep it safely within the boundaries of Earth.

Starting in 2021, AI and AI art entered the mainstream with the launch of ChatGPT and text-to-image tools. The latter models use vast data sets of images with associated text and, by means of deep-learning methodologies, generate digital images with different styles and attributes via users’ text prompts, aka natural language descriptions. While non-specialist media outlets debated whether these capacities would render human artists obsolete, actual digital art practice revealed both the flaws and potential of AI tools, highlighting the intense work and rigorous processes required to create sophisticated works. The text-to-image phase of AI art has arguably constituted the biggest shift so far, wherein tools such as DALL-E, Midjourney, and Stable Diffusion make images subordinate to language classification. These tools deeply fuse visuals with a lexical register, drawing on pre-existing dependencies. They produce their visual output on the basis of the textual classification of the training set and source data. Their visual creations are then determined by the alignment of users’ prompts with the text pre-associated with images, thereby building output on layers of existing taxonomies. Artists have begun exploring the potential and problematic aspects of these new semantic frameworks in medium-specific ways, assessing its impact on painting, photography, and film. Bennett Miller, for example, established parallels between the dawn of photography and the early days of text-to-image models’ transformative power in his 2023 exhibition of prints at New York’s Gagosian Gallery. Occupying a delicate threshold between the familiar and uncanny, the images remain eerily detached from a graspable subject. They are suspended in an alternate reality that both captures the essence of a distinctive stage of AI and highlights the differences between photographic processes and images generated in a photographic style by text-to-image models. While today’s AI can be seen as the result of a long evolution of “machine learning” rather than a new kind of technology, it also radically questions traditional definitions of media forms.

The focus of artists’ engagement with text-to-image AI programs covers a range: some use the software as more of a tool in the creation of projects that rely on multiple digital technologies in their creation process; others make it a focal point of aesthetic and conceptual explorations. While discourse about AI art in the mainstream media has focused on the dangers of artists being replaced by AI “creators”, the artists critically addressing AI technologies have been investigating the problems that the inherent classification, standardization, and optimization of AI tools pose to creativity due to their normative foundation. Text-to-image models use trillions of existing images to which corporations have access — many of them stock images — which means that they are operating within an echo chamber from the start. The text associated with these images was often originally created for marketing purposes, inscribing a specific agenda. Artists and other creators are currently not compensated for the use of their images, while the text-to-image tools generate an increasingly diluted version of their style. According to data published by AI Secrets in late 2023, AI generates roughly 34 million images per day, which then feed back into the training data sets, bringing us ever closer to the state of “model collapse”, where AI will be trained only on images of its own creation.

Digital art has always been at the forefront of engaging critically with the technologies it employs, and AI art can play a crucial role in assessing the aesthetic and socio-political impact of the tools that are shaping our future. Public discussions, as well as those in the industry, are frequently resorting to polarized narratives of extinction versus salvation, often driven by a commercial agenda or simplistic understandings of intelligence and sentience. AI art often provides a much-needed reality check, challenging facile assumptions and adjusting or breaking systems to raise more profound questions about human and other forms of intelligence.

Originally published in October, no. 189 (Summer 2024). The MIT Press. DOI: <https://doi.org/10.1162/octo_a_00533

1. Cohen kept developing the software until his death in 2016 and, over the decades, experimented with shifts in the style of work, and a move from monochrome to color output. From the 1970s to 19990s he also built his own output devices — from drawing machines to a Turtle, a robotic drawing device originally used in mechanical engineering, and a painting machine — and, in the 2000s, switched to a purely screen-based presentation.

2. See: Hito Steyerl (2016, April). “A Sea of Data: Apophenia and Pattern (Mis-)Recognition”. e-flux Journal, no. 72 [online]. Available at: https://www.e-flux.com/journal/72/60480/a-sea-of-data-apophenia-and-pattern-mis-recognition/

Recommended citation: PAUL, Christiane. The question of intelligence: a short history of AI art. Mosaic [online], April 2025, no. 203. ISSN: 1696-3296. DOI: https://doi.org/10.7238/m.n203.2502