When Chinese-born artist Cai Guo-Qiang first began training his custom-built AI model, cAI™, on his own artworks and writings, he recalls it conjuring ‘strange, raw, illogical images carrying a wild and mysterious magic of their own.’ But as the system absorbed more data and grew increasingly sophisticated, it also became more predictable. Concerned that his digital counterpart was becoming too much like him, Cai introduced it to different intellectual personas based on Friedrich Nietzsche, Albert Einstein, and Ptolemy. ‘I didn’t want it to be merely my shadow,’ he explains over email. ‘I wanted it to become another, more “divine” Cai Guo-Qiang in the digital world.’

While much of the debate around AI and authorship has focused on technology companies scraping creatives’ work without their permission, Cai is among a number of artists deliberately turning their archives into training material for their own custom AI models. Typically, these are well-established figures, able to draw on decades of their accumulated artworks and writings, and develop AI systems that can revisit their previous output, extend their practices in unexpected directions or, in some cases, become independent creative entities.

American painter David Salle, in contrast, was attracted less by the metaphysical possibilities of AI than by its capacity to dismantle and recombine pictorial elements, offering new ways of seeing his earlier work. He was initially drawn to the technology’s ability to recognize visual similarities and differences – a process he saw as closely related to his own approach to composition. The way it scans visual phenomena, Salle explains, ‘is more or less what the human eye does.’

Using his ‘Pastorals’ series of paintings (1999-2001) as a starting point, Salle began experimenting with a version of the AI model Stable Diffusion customized with the help of engineer Grant Davis, to generate unexpected new arrangements from elements of his original compositions. ‘The model seemed to have a will toward deconstruction and reconstruction,’ Salle continues, ‘which is really my bailiwick.’ These experiments formed the basis for his ‘New Pastorals’ (2024-), a series in which digitally generated compositions are printed onto canvas before being worked over in oil paint by the artist. The resulting paintings combine distorted mountain landscapes, fragmented bodies, isolated items of clothing, and even a turbine-sun, with their conflicting scales and perspectives disrupting any coherent sense of space.

Training in reverse

Unlike these seasoned practitioners with extensive back catalogues to draw from, for the young Senegalese-Lebanese artist Linda Dounia Rebeiz, AI has become a way of documenting her practice as it develops. ‘There’s a record of me in these machines,’ she observes. She has been working with the technology since 2018 and has come to regard it not only as a tool or collaborator, but as a kind of griot, the traditional West African storyteller entrusted with preserving a family’s generational history and collective memory. AI is ‘gathering a picture of myself and my practice over a long period of time,’ she continues, ‘in a way that no tool has ever been able to do.’

For her series ‘Spannungsbogen’ (2020-2022), which used abstraction to explore the underrepresentation of non-Western perspectives in AI-generated art, Dounia found herself creating new paintings specifically to train a GAN model – a type of AI that learns from existing images to generate new variations. She began with around 1,000 small abstractions, but soon found she needed to paint another 1,000. ‘It changed the nature of how I painted because it had to become more programmatic and robotic,’ she recalls. Training her model was shaping her painting as much as her paintings were shaping the model.

Eventually, her AI model generated 10,000 variations, which she then developed into animations and still images. ‘For me, abstract painting is about chasing happy accidents,’ Dounia explains, ‘moments you didn’t anticipate, that surprised you.’ Yet she did not take these results as evidence that the machine understood her artistic intentions or style. ‘It was just tripping over itself,’ she recalls, ‘but some of it I found delightful or just really interesting.’

Like Dounia, Salle makes a clear distinction between the machine’s capacity to produce unexpected results and its ability to understand their significance. ‘The machine doesn’t really know what rules it’s breaking,’ he explains. ‘It doesn’t really know how disruptive it’s being.’ He draws a parallel with Francis Picabia’s method of buying secondhand paintings and working directly over them, allowing fragments of the original compositions to remain visible. ‘It’s a found object – even if it’s a found object that I created myself, it’s still a found object,’ adds Salle. What matters is the subsequent intervention. Whether the finished painting ‘rises to a level of joy, or remains boring is entirely up to me. That’s all my responsibility.’

Questions of autonomy

What if the relationship between an AI model and an artist were pushed further, to the point where the machine might work alongside its progenitor, or even independently of them? Cai, who began experimenting with AI in 2017, has ambitions for cAI™ that extend well beyond reproducing his own work, imagining that it could one day evolve ‘into a fully independent creator.’ The artist has even developed a robotic arm attuned to his way of working, allowing cAI™ to scatter and ignite its own versions of the gunpowder paintings for which Cai is best known. ‘It often paints alongside me: the same subject, yet two entirely different visual languages, logics, and forms.’ In 2024, cAI™ had its first solo exhibition, ‘Soul Scan,’ in Macau, featuring gunpowder paintings and laser-engraved canvases, their surfaces marked by scorched black patterns and vivid bursts of pink. The show suggested that what began as an extension of Cai’s practice might eventually develop an artistic identity of its own.

Although Salle stopped working with AI some time ago, his experiments, first with the Pastoral paintings and, later, the ‘Tapestry’ series (1989-1991), produced so much material that he is still working his way through it. ‘Someone else could start where I left off and do something else with it,’ he acknowledges. ‘It wouldn’t surprise me if AI became an autonomous image generator, but it wouldn’t interest me either, because it doesn’t really mean anything. All it’s doing is breaking apart and recombining.’ For Salle, the machine’s ability to generate images independently is not the same as understanding their artistic significance.

For Cai, who has pursued precisely the kind of autonomy Salle sees little value in, increasing sophistication has brought its own concerns – particularly about what might be lost as cAI™ grows ever more accomplished. ‘I often lament that I am searching for the Buddha, yet people often see only the “image” of Buddha that I’ve created,’ Cai laments. ‘In the age of AI, if algorithms smooth away every irregularity, what remains in the end may be nothing more than a perfect “image”.’

Yet for all that cAI™ has learned about his work, Cai seems no closer to understanding what he ultimately wants from his experiments with his AI model. ‘Although I often ask cAI™ what it can offer me that is “beyond the human,” what I am really asking may be: “What is it that I truly want?”’