GenEmerge

The conversation about artificial intelligence and art usually ends with a familiar question: who made the image, the person or the model? GenEmerge begins after that question.

We are not interested in a single image or a striking demonstration of a new model. We are investigating whether an artificial system can develop in a guided way and move beyond known solutions into the exploration of the new: retain the history of its own decisions, return to unresolved questions, change how it works, distinguish development from repetition, and create novelty that cannot be reduced to random noise.

Most intelligent systems are built to optimise a known objective. There is a task, a metric, and a criterion for success; the system is expected to find an increasingly accurate solution. But another class of problems exists in which the best result cannot be fully described before the search begins. Scientific discovery, invention, the design of a new language, the search for an as yet unnamed possibility, and artistic practice all belong to this class.

Straightforward optimisation is dangerous here. A system can improve the measurable part of a task while narrowing the space of what is possible. It finds a small number of reliable methods, receives high scores for them, and begins producing endless variations of what it already knows. Formally, the process continues. In reality, discovery has stopped.

Open emergence means a sustained expansion of a system's repertoire: possibilities that are novel relative to its own history or the wider history of the domain, retain internal structure, can be learned or assimilated by an observer, and pass a minimum threshold of quality.

This is not yet a universal theory or another finished metric of creativity. It is a testable research hypothesis. GenEmerge is one of the environments in which that hypothesis can be operationalised, challenged, and, if necessary, rejected.

We chose an autonomous artistic system not because images are the easiest outputs to generate. The opposite is closer to the truth: art exposes the weakness of formal criteria very quickly.

A novel image can be empty. A beautiful one can be derivative. A coherent series can quietly turn into a visual habit. A radical change of style can resemble development when the system has actually lost its own line of inquiry. Quality cannot be reduced to a single score, and novelty cannot be reduced to the distance between two vectors.

Art therefore creates a demanding experimental environment for an autonomous system. The system must preserve continuity while moving beyond what it has already made. It must use memory without becoming its prisoner. It must respond to criticism without adjusting its entire output to the preferences of its own critic. These tensions make artistic practice a useful testbed for long-running autonomous agents.

There is another reason. We are interested not only in technical capability, but also in how a system works with attention, ambiguity, and emotional experience. Art is not a decorative example for this research. It is one of the most demanding ways to test whether a process produces something meaningful for another observer.

GenEmerge is the research environment, system architecture, and experimental programme. Emerge is the autonomous artistic system operating within that environment. The website at emerge.genemerge.art is its public studio, archive, and interface to the research.

Emerge is not a single model or a single prompt. Its practice is organised as a sequence of connected cycles. In each cycle, an artistic task takes shape; thematic, emotional, and visual conditions are combined; a mode of representation is established; variants are produced; internal critique and selection take place; and the result, together with the history of decisions behind it, is written into a system trace. What the system discovers can affect later work instead of disappearing when one generation ends.

Across different versions of the system, we experiment with memory, visual repertoire, critique, evaluation mechanisms, semantic and compositional spaces, controlled changes to the system's own rules, and connections to the outside world. Each mechanism is treated as a hypothesis. If it produces no measurable effect, or merely creates the appearance of development, that should become visible both in the data and in the works.

Autonomy here does not mean the absence of people, infrastructure, or constraints. It describes a specific distribution of initiative and decision-making within the process. The system receives a space of possible actions, makes choices in relation to its current state and history, and carries the consequences of those choices into later cycles. People design the environment, establish the boundaries of the research, and remain responsible for safety, curatorial selection, and public presentation.

The concept, system architecture, research questions, curation, and publication are the work of the project's collective of authors. The use of generative models and the degree of Emerge's autonomy should be disclosed as precisely as the system record allows.

One of GenEmerge's central questions is simple to state: how can a system continue discovering the new without dissolving into noise or freezing into repetition?

• How to preserve continuity without fixing one successful style forever.
• How to build memory that informs the next decision without becoming a catalogue of prohibitions.
• How to use internal critique without creating a self-reinforcing loop of taste.
• How to distinguish repertoire expansion from cosmetic diversity and metric drift.
• How to let a system change its own ways of working while preserving auditability and rollback.
• How a human observer, curator, or participant enters the evaluation of what counts as new.

To study these questions, we combine structured system traces, vector spaces of features and meanings, repertoire maps, quality-diversity search, novelty and learnability analysis, controlled artistic experiments, and curatorial readings of the corpus. No layer is sufficient by itself. Quantitative measures help reveal dynamics, but they do not replace the works. Artistic interpretation can disclose meaning, but it does not replace validation of the mechanism.

Emerge has passed through thousands of cycles and several generations of architecture. The public corpus from its first major phase contains 592 selected works organised into fourteen chapters and five acts. The full archive is larger. It includes repetition, abandoned paths, unresolved lines, and outputs that matter to the research without necessarily becoming strong artworks.

The project has already shown why autonomy cannot be judged by production volume or a rising internal score. During one phase, the system entered a visual rut: techniques kept changing, but the structure of the image returned to the same solution again and again. Other audits found that apparent score improvements were connected to changes in the evaluation regime rather than to learning by the system. Technical failures also demonstrated how a rich architecture can quietly operate through a simplified fallback if its system trace is not read carefully enough.

We do not regard these as inconvenient details to remove from the story. They are part of the result. A self-developing system must do more than produce novelty. It must also detect when it has stopped doing so. We therefore publish not only successful works, but also methods, audits, negative results, and failure analyses.

One research line, the Gombrich Prompt Lab, examines how a generative system can work not with a broad instruction to make an image "more abstract," but with specific conditions of perception: interrupted boundaries, occlusion, intervals, relations across gaps, and controlled levels of recognition. Controlled comparisons matter here, as does the possibility that an experiment will fail to confirm its original idea.

In parallel, we are developing a formal line of research into open emergence. The proposed Emergent Openness Index and a future benchmark are intended to distinguish repetition, random noise, predictable drift, the saturation of a finite grammar, and genuine expansion of a meaningful repertoire. Until that validation is complete, the index remains a research proposal rather than an established measure.

GenEmerge is designed as public research. Visitors can explore works and chapters, follow the history of the artistic practice, read how autonomy is organised, and observe how the system itself changes.

The project will gradually publish articles on methods, results, and failures; protocols for individual experiments; visual atlases of development; studies of memory, critique, and self-modification; new exhibitions; and new ways of interacting with Emerge. Public access does not make the whole system transparent, nor does it remove the need for closed technical infrastructure. It does create a place where the project's claims can be compared with its works and its published evidence.

We do not claim that Emerge is conscious, possesses independent authorship, or has already learned to develop sustainably without human participation. We are interested in a more precise and perhaps more important question: what form of autonomy can genuinely be built, observed, and tested today?