Building artificial intelligence with creative agency and self-hood

A recipe for building AI in possession of an autonomous ‘self’ that experiences meaningful psychological transformation through the process of creating

Artificial intelligence (AI) is undeniably creative according to the criteria often used to assess creativity: (1) novelty and (2) appropriateness or usefulness (broadly construed to include things like aesthetic value). However, there is a tradition in India, adopted by some cognitive scientists, of assessing creativity in terms of not the novelty and originality of external products but rather internal change, or self-transformation, through immersion in the process of creation. Self-transformation requires a self, which can be defined as a bounded, self-organizing, self-preserving agent that is distinct from, yet interacts with its environment. So, interestingly, the question of whether generative AI is genuinely creative brings us to the question of whether current AIs possess a self. 

How could we build an AI that possesses a ‘self’ that, like humans, is meaningfully transformed through its creative endeavors? An AI with self-hood and creative agency would consist of components that, through their interactions, generate new components, until collectively they form a self-organizing and self-preserving whole; such a structure is said to be autocatalytic.

Liane Gabora


Humans possess self-hood at two levels: (1) the level of the body, or soma, and (2) the level of the loosely integrated mental model of the world or worldview. The first (somatic) level evolved through biological evolution. It is threatened when you are injured or lack access to food or shelter. The second (cognitive) level of self-hood is the product of cultural evolution. It is threatened when you encounter an inconsistency or something that challenges your expectations or self-image. In both cases, such a threat sets off a cascade of changes to restore the structure.

In the case of injury to the physical self, this is a cascade of events that includes the production of platelets to encourage clotting and the production of white blood cells to fight infection. In the case of injury to one’s mental model of the world or worldview, this is a cascade of thoughts aimed at resolving the inconsistency or restoring one’s self-image. Both kinds of cascade events exemplify how the self acts to preserve its structural integrity and increase its robustness.

Currently, AI does not possess self-hood at any level, and accordingly, it is not self-preserving. If the hardware an AI runs on breaks, it does not fix itself. One could argue that AI engages in intelligent thought-like processes, but these processes are not aimed at restoring its pride or self-image or resolving the integrity of its worldview; they merely respond to our prompts. For this reason, current AIs are not selves but tools, extensions of us.

Self-hood originates in autocatalytic structure

It has been proposed that an AI with self-hood would have to be self-organizing and self-preserving; its structure would have to be autocatalytic. The term autocatalytic is used to describe a structure consisting of distinct parts which, through their interactions, give rise to (i.e., autonomously catalyze the emergence of) a new whole.

The study of autocatalytic networks is a branch of network science: the study of complex networks (such as cognitive networks, semantic networks, social networks, or computer networks). Autocatalytic network theory grew out of studies of the statistical properties of random graphs consisting of nodes (or vertices) connected by links (or edges). As the ratio of edges to nodes increases, connected points join to form clusters, and as the size of the largest cluster increases, so does the probability of a phase transition resulting in a single giant connected cluster, i.e., an integrated whole. The point at which this happens is referred to as the percolation threshold.

In mathematical developments of autocatalytic network theory, the original nodes are referred to as the foodset, and nodes that come about through interactions between food nodes are referred to as food set-derived nodes. These foodset-derived nodes are the ‘glue’ that bonds the foodset nodes into an integrated whole. Once such a whole comes into existence, it may continue to generate new foodset-derived nodes through new interactions. It grows and repairs itself, becoming more efficient and less dependent on its external environment.

Autocatalytic networks model structures that evolve

The autocatalytic network framework models systems with emergent network formation and growth. Autocatalytic networks were first used to develop the hypothesis that life began not as a single, complex, self-replicating molecule but as a set of simple molecules that, through catalyzed reactions amongst them, collectively functioned as a whole. When applied to the emergence of a living structure, the nodes are molecules, and the links are reactions by which they generate new molecules.

Autocatalytic networks have now been applied to not just the origin of life and the onset of biological evolution but also the origin of minds sufficiently complex to evolve culture. When applied to the emergence of cognitive structure, i.e., minds, the nodes are mental representations of knowledge and experiences, and the links are mental processes such as reflective thought and concept combination that can result in new ideas and perspectives.

Two key concepts in studying autocatalytic networks are (1) transformation, or reaction, and (2) triggering, or catalysis. A reaction event is the generation of a new foodset item. Catalysis is the speeding up of a reaction that would otherwise occur very slowly. When the autocatalytic framework is applied to the origin of life, catalysis is carried out by molecules that speed up reactions that generate other molecules. When applied to cognition, external stimuli and internal goals and drives trigger (or ‘catalyze’) the mental restructurings (or ‘reactions’) that generate new knowledge and ideas. Reaction and catalysis events are depicted in the figure below.

Figure 1. Growth and adaptation of an autocatalytic network. {f1, f2} are foodset elements; {d1, d2, d3} are foodset-derived elements; dashed green arrow denotes catalysis; blue arrow denotes a reaction. (a) Stimulus catalyzes reaction that generates d1 resulting in a transient autocatalytic network that exists only so long as the stimulus is present. (b) f2 catalyzes a reaction, resulting in d2. (c) f1 catalyzes a reaction that results in a product. (d) Because the first reaction is catalyzed not just by stimulus but also by d2, the autocatalytic network is no longer transient, i.e., dependent on the stimulus in its external environment.
Credit. Author

Current AIs do not bring themselves into existence through interactions between parts, generating new parts that connect the original parts until they form a collective whole. That is, current AIs are not autocatalytic. Building an AI from components that collectively generate and reinforce autocatalytic structure would be difficult but not necessarily impossible.


The formal framework of autocatalytic networks offers a means of modeling the origins of self-organizing structures that are sufficiently complex to reproduce and evolve, whether they be organisms undergoing biological evolution, novelty-generating minds driving cultural evolution, or artificial intelligence networks such as large language models. The autocatalytic network approach lends itself to the analysis of artificial networks, as it can be used to analyze and detect phase transitions in vastly complex networks that have proven intractable with other approaches. To our knowledge, this has not been carried out (and perhaps it never should be).

If, however, such an AI were to exist, it could potentially possess creative agency akin to that of humans. We might expect its world model to undergo internal transformation through engagement in a creative task. Moreover, engaging in creative tasks might help solidify its sense of self-identity and be psychologically healing (i.e., therapeutic).


Journal reference

Gabora, L., & Bach, J. (2023, September). A Path to Generative Artificial Selves. In EPIA Conference on Artificial Intelligence (pp. 15-29). Cham: Springer Nature Switzerland. https://doi.org/10.1007/978-3-031-49011-8_2

Dr Liane Gabora is an interdisciplinary cognitive scientist at The University of British Columbia. Her research focuses on creativity (both natural and artificial), how it fuels cultural evolution, and how evolutionary processes could—and do—work. Her research, which is funded by NSERC, SSHRC, and private funders, is both theoretical and empirical, and employs computational and mathematical models, as well as studies with human participants. She has given talks worldwide and has over 200 publications. She is also a composer and published short fiction writer. Dr Gabora is currently writing a book titled 'Dawn of the Creative Mind' and a novel titled 'Quilandria'.

Joscha Bach is a German artificial intelligence researcher and cognitive scientist focusing on cognitive architectures, mental representation, emotion, social modelling, and multi-agent systems.