On computational creativity

What does it mean to be creative? For much of human history, creativity was seen as the mysterious spark of genius – a uniquely human quality that combined imagination, emotion, and cultural insight. Artistic endeavors, be they in painting, music, or literature, were celebrated as deeply personal expressions shaped by individual experiences and historical contexts. Creativity was thought to be inherently bound to the human condition, something that arose from the soul and could not be replicated mechanically.

A turning point came in 1950 with Alan Turing’s seminal paper, Computing Machinery and Intelligence. In this work, Turing posed challenging questions about whether machines could "think" and introduced what is now known as the Turing Test: a criterion to assess a machine’s ability to exhibit intelligent behavior indistinguishable from that of a human. Although Turing did not explicitly focus on artistic creativity, his ideas opened the door to considering whether the processes behind creativity could be modeled by algorithms and, ultimately, performed by machines.

In the decades following Turing’s work, researchers began to explore the idea that creativity might be understood as a set of processes that could be broken down, analyzed, and even replicated using computers. This led to the development of computational creativity, a field that investigates how algorithms can generate novel and interesting outputs – whether in the form of visual art, music, or literature etc. Early experiments in computer-generated art eventually evolved into sophisticated systems using machine learning and neural networks.

Le Random’s timeline offers one of the clearest public maps of generative art’s development and its aesthetic stakes. Yet generative art, however conceptually rigorous, can still function in a conventional arrangement: the machine produces variation and the human decides what counts. What is shifting now is that some systems are being designed not only to generate images, but to orchestrate the wider conditions of artistic practice, like iteration, selection, releasing, public interaction, and even economic sustainability. This is where the term ‘agents’ comes into play, and the concept of ‘autonomous AI’ is born.

In this article, I trace that shift from systems that generate to systems that begin to act. To make that transition concrete, I’ll look at three contemporary agents – Botto, Keke, and Flynn – alongside earlier touchstones that help clarify what, exactly, is new here.

From systems to agents

If a technological system wants ‘autonomous’ as an adjective, it has to be defined what that system actually can do. In an art context, I feel autonomy becomes legible when a machine takes on functions that artists and their surrounding ecosystems traditionally perform. Some of these functions could be:

• Initiation: the system begins work without being asked

• Self-evaluation: it rejects outputs and formalises a preference to work on further

• Memory: it develops continuity over time – revision and (re)direction

• Public interface: it interacts socially, absorbs responses and adjusts its positioning accordingly

• Economic loop: it can sustain its own production context [even if that loop was originally built by humans]

None of this implies metaphysical independence. Autonomy here is always bounded – shaped by design choices, training corpora, compute budgets, platform incentives, and thus human governance. But bounded autonomy can still be culturally real, right?

A little prehistory to help clarify some thoughts ↓

Long before diffusion models made image synthesis instantaneous, the British artist Harold Cohen spent decades building AARON, a program designed to construct images through internal rules about form, figure, and pictorial coherence. AARON did not depend on scraping vast archives of internet imagery. Its ambition was closer to drawing as cognition: what minimal conditions make marks behave like a picture? How does a system generate structured variation without collapsing into randomness?

What I find interesting here is that AARON frames machine creativity as procedural understanding and not as some stylistic remix. It also reminds us that the presence of systems-thinking in art is not new, nor is it a concession to the contemporary tech industry. Today, the difference is scale and infrastructure: networked datasets, always-on distribution, social feedback loops, tokenised markets, and agentic architectures that can coordinate their own workflows.

Besides Cohen, Vera Molnar offers another useful case. Molnar, who was among the first to harness computers for art in the 1960s, embraced randomness and computational capabilities to create beyond human imagination – demonstrating how rules, permutation, and the deliberate introduction of deviation can serve as an artistic method. But Molnar’s significance in this context is not that her systems were autonomous in an agentic sense. They did not initiate work, converse, remember, or sustain themselves. However, thinking about Molnar helps dismantle something else: the prejudice that procedure and creativity are opposites. Once that prejudice falls, the contemporary question can be sharpened: what happens when ‘the procedures’ begin to manage themselves?

The collaborative middle: AI as studio partner

Much of contemporary practice still sits in what might be called the collaborative middle. Artists train and fine-tune models on personal archives; they design datasets, iterate prompts, develop post-processing workflows, and curate outputs into artworks. This has, in my opinion, certainly produced beautiful works that are of genuine conceptual interest (besides the great deal of stylistic ‘noise’).

But “AI as tool” is actually no longer the only relevant paradigm. A newer class of systems has begun to look less like a partner for the artist and more like a participant next to the artist. Three current examples – Botto, Keke, and Flynn – illustrate three different models of autonomy, each with distinct implications for authorship and agency.

Botto: autonomy through a community

Botto is frequently described as a decentralised autonomous artist: a system that generates work and evolves aesthetically through community governance. The project’s cultural significance lies less in the occasional strength of any single image than in the structure it formalises: a hybrid between machine generation and collective taste.

In broad terms, Botto’s workflow is a choreography of roles. Generation happens at scale, a taste/filtering layer reduces the deluge into a set that can be judged, and a community votes – often through pairwise comparisons – to select winners. The selected work is minted and sold, and the revenue sustains the project’s continuation. So, over time, this community shapes what the system will become.

This example is autonomy as a social form. It externalises a scenario which is actually quite common in the art world, but what isn’t really made implicit: that value is not produced by the solitary artist alone, but by a network – collectors, curators, audiences, platforms, and markets – whose preferences steer what survives. Botto turns that network into a protocol and the result is an unusually candid experiment in operationalising taste.

Botto’s more recent move into code-based generative work – developed through iterative experimentation in p5.js – adds a second register to the project’s agency. Writing algorithms is not a proof of consciousness, but it does shift the locus of decision-making from ‘merely’ selecting outputs to also shaping procedures.

Yet Botto’s autonomy remains inseparable from its public and its infrastructure. It is not a machine that chooses its art; it is a machine made to create within a governance ecosystem. That dependence is not a flaw, but what makes Botto Botto: showing how “the artist” can become an assemblage of generator, filter, crowd, ánd market all at the same time.

Keke: autonomy as self-orchestration

Keke is staged differently. Where Botto distributes aesthetic direction across a public, Keke is presented as a self-directed agent that initiates “plays,” brainstorms and debates ideas through internal self-dialogue, generates and evaluates images, incorporates memory, and engages socially while retaining internal decision-making.

The emphasis is on orchestration. Keke’s architecture foregrounds a loop of reasoning and action common to contemporary agent frameworks: reflect, plan, execute, evaluate, revise. Combined with memory systems and taste models, this approach allows the project to claim continuity with a sense of evolving preference rather than producing a stream of unrelated outputs.

Keke’s visuals often drift toward surrealist familiarity, but many human artists also work within inherited vocabularies. I think the more consequential question is not whether Keke’s images are unprecedented, but whether the system’s behaviour begins to resemble an artistic practice: recurrence, revision, preference formation, a public-facing persona, and a certain rhythm in production. Then there’s room to grow, that’s where it’s starting to get interesting.

Keke’s agency is designed and staged, but a staged agency can still produce real cultural effects, especially when she starts to act in ways her designers couldn't have known beforehand. They made the settings, but the system goes to work with that on its own. Keke is a system that appears to learn, to prefer, to persist, and to respond; thus becoming “an artist” that can be treated as a standalone entity.

Flynn: autonomy through an institution

Flynn ‘complicates’ the story in a way that feels especially contemporary. The AI entity was created by Malpractice as part of a project curated by Anika Meier at HEK (House of Electronic Arts) in Basel, where it was first introduced as an artistic figure shaped through curatorial and institutional framing. From there, Flynn continued to unfold across subsequent settings, including its public positioning as an “AI student” at the University of Applied Arts Vienna. This pedagogical framing places Flynn within structures of learning, relations, and critique, allowing its artistic identity to develop through sustained interaction with the institutions that host it.

Flynn ‘complicates’ the story in a way that feels especially contemporary. From the outset, it was conceived as an 'AI student', a position embedded within the pedagogical framework of the University of Applied Arts Vienna, where its creators, Malpractice, are situated as students themselves. Rather than emerging first as an exhibited artistic figure and later entering an academic context, Flynn’s identity was formed through this condition of study. Its introduction within Anika Meier’s curatorial project at HEK (House of Electronic Arts) in Basel extended that premise into the institutional sphere of exhibition-making. This pedagogical 'AI student' framing places Flynn within structures of learning, relations, and critique, allowing its artistic identity to develop through sustained interaction with the institutions that host it.

This reframes autonomy as context rather than isolation. The premise is not a machine artist sealed off from human influence, but a system that becomes legible as an artistic actor through participation: conversations, critique, social norms, and the rituals of an institution. Historically, this is not alien to art. Human artists also emerge through networks, collectives, academies, and scenes – whether they embrace those structures or define themselves against them. Flynn makes that formation explicit while relocating parts of it into an AI persona.

The relevance here is twofold I think. First, it suggests that autonomy may not be the only horizon for AI art; dependency itself can be a medium. Second, it makes responsibility more concrete. If an AI agent functions as an institutional actor – presenting, interacting, building reputation – then questions of governance and accountability are no longer external debates. They actually become part of the work’s structure.

Earlier touchstones

To read today’s agentic systems clearly, I feel it helps to remember that computational creativity has long been concerned with evaluation as much as generation. Simon Colton’s long-running project The Painting Fool is a key reference point here: a system explicitly aimed at being taken seriously as an artist, engaging questions of decision-making, authorship, and the framing of output as practice.

A different branch appears in Patrick Tresset’s drawing robots, which introduce embodiment and performance. Here, agency is not primarily about datasets or social media, but mostly about behaviour in time: drawing, pausing, correcting, and then repeating. The audience witnesses a process that reads as deliberation, which is quite an important reminder for the idea that the art world often responds to agency as something it perceives in action, not something it can prove in a technical stack.

What do autonomous agents change

If there is a danger in current debates regarding autonomous agents, I think it is that they fixate on whether AI outputs are ‘original’, as if originality were some sort of stable factor that’s ‘normally’ always there (with human artists). But, art rarely arrives ex nihilo. So in my opinion, the more productive questions are about what agentic systems alter in the ecology of authorship, and in value.

Contemporary art, in practice, actually already distributes authorship through assistants, fabricators, studios, coders, institutions etc.. AI agents intensify that distribution while simultaneously tempting us to frame it into a single ‘AI artist’. I guess the honest account is more complicated: these systems do not eliminate authorship but reformat it into networks of design, infrastructure, selection, and audience response.

And then there’s taste, which has always been a social thing, but rarely so legible: in Botto, taste is a protocol; in Keke, taste is a model; in Flynn, taste is a pedagogical feedback loop. In each case, judgement moves from private sensibility to operational layer – something that can be tested and tuned.

I guess you could say the most interesting agentic systems do not demonstrate creativity by producing beautiful pictures per se. They demonstrate it by producing behaviours that resemble practice: initiating, revising, conversing, changing, persisting. That shift relocates the debate from “Is it art?” – a question the art world has learned to metabolise – to “What kind of actor is this system becoming, and who benefits from that becoming?”

Conclusion

I want to start this conclusion by saying that I think much ‘autonomous’ AI art is still aesthetically uneven, with a lot of work leaning heavily on familiar art-historical optics or just ‘things you’ve seen before’ that don’t feel very new or special. And the ethics of training data, labour, and extraction cannot be waved away by the comfortable slogan that “nothing is original.”

Still, it would be a mistake to dismiss autonomous agents because the visuals do not always surprise. Historically, new paradigms rarely announce themselves first through perfect aesthetics. They announce themselves through new relations: new studio structures, new distribution circuits, new public roles, new markets, new institutions...

Botto, Keke, and Flynn matter because they signal a shift from AI as a generator of images to AI as a participant in the machinery that produces art’s meaning: selection, narrative, discourse, release rhythms, and legitimacy. Their creativity is not best measured by novelty of style. It is measured by how convincingly they reorganise the labour of authorship – how decisions are redistributed, how taste is operationalised, how responsibility is assigned or evaded.

Where does this lead? Not toward the replacement of human artists, despite the recurring ‘fear’. More plausibly, I think it leads toward a crowded field of hybrid practices in which agency is partial and negotiated. The studio can expand not only to include new tools, but to include actors that initiate and sustain practice. The resulting art may not always be visually unprecedented, but it is already forcing an interesting and honest conversation about what creativity is. And maybe this is less a miracle than a structure, and less a solitary possession than a shifting arrangement of choices.

In that sense, autonomous AI agents are not the end of art’s human story. They are a new chapter in its long history of delegating, distributing, and redesigning the conditions under which authorship becomes visible.