Generative AI, which can now produce entire visual universes from a few typed words. Yet within this process lies a paradox: is the artist still the creator, or has the machine become an equal partner in the act of creation? The Museum of Artificial Art (MoAa) describes AI as a “site of resistance, speculation, and cultural imagination,” suggesting that the human-machine relationships is no longer merely technical, but also aesthetic and even ontological: a new way of seeing, a new way of dreaming.

And as we stand at this crossroads, we must ask ourselves: who is really doing the dreaming?
The answer may be more complex than we are ready to admit.

The idea of a machine as a creative partner is far from new. As early as the 1960s, British artist Harold Cohen (1928–2016) developed one of the first computer programs devoted to artistic production: AARON. The name, inspired by the biblical messenger, was no coincidence – AARON was not a mere tool, but a true collaborator. Cohen programmed it to generate images by following rules that mimicked human decision-making processes. Over time, an authentic dialogue began to emerge between artist and software, reshaping long-held assumptions about authorship and creativity.

A new phase began in the 2010s with the advent of deep learning and neural networks. These technologies allow systems to learn from vast datasets and to generate increasingly complex images. In 2015, Google released DeepDream, an image-processing program that, through a convolutional neural network, detected and amplified hidden patterns within images – a process akin to pareidolia, the human tendency to perceive familiar forms in abstract stimuli. The results were striking psychedelic visions, landscapes populated by eyes, figures, and ever-morphing patterns – dreams seemingly produced by an artificial mind. DeepDream went viral and led to the creation of Google’s Artists and Machine Intelligence (AMI) program, culminating in the 2016 exhibition DeepDream: The Art of Neural Networks.

From that moment on, AI-assisted art moved beyond experimental circled and entered in the global art market. In 2018, Christie’s auctioned Portrait of Edmond de Belamy, created by the French collective Obvious using a Generative Adversarial Network (GAN). The portrait, somewhere between a Renaissance nobleman and a fading apparition, bore no signature, only a line of mathematical code. The following year, Sotheby’s presented Mario Klingemann’s Memories of Passersby I, an AI-based installation that generated an endless stream of synthetic portraits in real time. These were not reconstructions of real people, but visions of beings who had never existed – appearing and dissolving like fragments of a machine’s dream.

The sale of Portrait of Edmond de Belamy made explicit the core tensions in the debate on AI and art: the legitimacy of algorithmic creativity, the question of authorship, and the aesthetic value of machine-generated work. A discussion that echoes the reflections of Walter Benjamin in the 1930s and John Berger in the 1970s, both of whom suggested that the artistic value of a work could transcend its artisanal origin. In this light, the use of a machine does not represent a threat, but rather a natural component in the ongoing evolution of what we call art today.

For a long time, the relationship between humans and AI has been framed as a one-directional: we train the machine so that it serves us. Feeding on data that describe human artistic expression, AI processes this information without ever truly seeing what it absorbs. When it “paints” a dream, it is, in fact, assembling a mosaic made of billions of visual fragments from human dreams, filtered through our prompts.

But the relationship is far more complicated. In her 2023 essay Mean Images, art historian and filmmaker Hito Steyerl offers a piercing critique. She argues that AI-generated images cannot truly be considered creative. Instead, they are statistical results derived from the data on which they are trained. She calls them mean images, where “mean” refers not only to the mathematical average but also to their aesthetics mediocrity, and at times, their unsettling nature. They reflect what is most common in the dataset, not what is most meaningful.

This condition makes them revealing: they expose our collective biases, dominant aesthetics, and social norms, yet reveal almost nothing about individual experience or personal vision. Lacking intentionality, emotion, and historical context, these images are, as Steyerl describes, post-representational, they do not depict anything specific, but rather the statistical average of everything.

From this arises an uncomfortable realization: in our attempt to teach AI to “see”, we may have trained it to see in the most generic way possible – privileging the predictable over the particular, the common over the singular. Tools like Stable Diffusion do not generate likeness, they generate a likeliness, constructed fron millions of aggregated images.

Such images, therefore, are never neutral. They reproduce and amplify biases, stereotypes, and power structures embedded in their data. They mirror how society sees itself—through distorted lenses that often confuses the popular with the banal, the ordinary with the oppressive.

It is within this context that the mission of the Museum of Artificial Art (MoAa) becomes vital.
Rather than celebrating AI as a mere technological novelty, the MoAa explores its cultural and critical implications, supporting artists who use it to question its very foundations. It promotes a practice that goes beyond the instrumental use of technology – one that probes its limits through conceptual inquiry, social critique, and visionary experimentation.

The power of art does not lie in what machine generates by default, but in the human ability to direct it against its own statistical nature. It is within this liminal space that the creative act is born.

A clear example is Joy Fennell, who employs AI to rewrite history and imagine futures that defy the mean image. Her practice challenges stereotypical representations of Blackness in visual culture, subverting the algorithmic biases toward homogeneity and restoring complexity, identity, and intention to the act of imagination.

What happens when AI begins to speak in a language we no longer fully understand? Perhaps the most surprising aspect of generative AI is not its efficiency, but its capacity to invent. When faced with ambiguous or impossible concepts, the machine does not fail, it hallucinates. It blends references in unpredictable ways, creating images that are both alien and oddly familiar, not merely imitating but synthesizing.

This dynamic is explored in the MoAa’s inaugural exhibition, I Miei Ricordi: AI Dreams, Hallucinations, and Memories. Among the featured works, Maddy Minnis generates painterly scenes that resemble machine hallucinations, dreamlike visions that evolve into entire ecosystems and imaginary species. These are not reproductions of anything that ever existed. They are something else: fragments of possible futures, dreams the machine has learned to tell. 

This changes how we understand creativity. The machine appears to become visionary, while we – artists and viewers alike – turn into apprentices, exploring its strange and seductive landscapes. This is the new dialogue: we offer AI the data of our past, and it returns images of a future we could not imagine.

This is not about passive consumption. It is about co-creation: a collaboration where artists don’t just use AI but actively shape its visions, channeling its capacity to hallucinate, foreshadow, and rewrite timelines. In this space, art becomes not a product of prediction, but a process of a dream unfolding.

So, are we teaching AI to see—or is it teaching us to dream?

Perhaps the answer lies not in choosing one or the other, but in recognizing the reciprocal transformation at play. As we guide machines to imitate our imagination, we learn to see ourselves – and the world – through a new lens. Not always clearer, but undeniably different. And in that difference, a new kind of vision is emerging.