An arXiv preprint proposes a way to describe the kinds of agency advanced AI systems might exhibit, but it does not establish that any AI system is a moral or legal agent. The framework organizes agency around three dimensions—nature, mode and locus—and combines them into eight possible agency instantiations. It then groups those configurations as conventional, contested or controversial.
The paper’s central question is how to determine the type of agency exhibited by advanced AI, especially when the form under discussion is not moral agency. It is a conceptual analysis, built by synthesizing established theoretical positions in philosophy, ethics, legal theory and sociology. No participant sample or empirical dataset is described.
A framework built from theory
That choice of method sets the boundaries of the result. The paper is not measuring how often systems act in a certain way; it is building a vocabulary for describing possible agency configurations. In the framework, nature, mode and locus are kept distinct, so a question about an AI system’s legal status need not automatically become a claim about its moral status.
Put more plainly, the framework asks what kind of agency is at issue, how it operates and where it is located. That last dimension accommodates individual and collective forms, which appear throughout the proposed categories. The three dimensions are combined into eight possible configurations.
Eight combinations, three levels of dispute
Three configurations are described as conventional: individual moral human, individual legal human and collective legal non-human. The paper gives the corporation as the paradigmatic example of the last category. Three others are classified as contested: collective moral human, collective legal human and collective moral non-human.
The remaining two are controversial: individual legal non-human and individual moral non-human. The labels identify the parts of the framework where the underlying philosophical or legal questions are unsettled; they do not amount to a finding that a specific AI system belongs in either category.
Legal and moral questions split apart
Legal agency is treated separately from moral agency. In this context, the legal question is whether an entity’s acts could eventually have a recognized place in law, including possible liability or accountability; the moral question is whether it meets the criteria for moral agency. The framework allows the first question to be discussed without presupposing the second.
For AI, individual legal non-human agency is classified as controversial. The paper asks whether legal affordances could eventually be extended to an AI system and whether that system could be held legally liable or accountable. It does not determine that any AI system should receive legal agency.
Individual moral non-human agency is also classified as controversial. The author says it is unlikely that any AI technology fulfils the relevant criteria, even under the paper’s enactivist perspective, one of the philosophical approaches it considers. That is a theoretical judgment, not an empirical assessment of particular systems.
The governance problem is attribution
The framework’s practical pressure point appears in governance. The author argues that the distinction becomes increasingly relevant when instrumental goal pursuit makes it difficult to attribute an AI action to a particular human actor. The issue, in other words, is how responsibility should be discussed when attribution becomes complicated.
The paper is careful about the examples used to motivate that concern. Its introduction characterizes the discussed deception as anecdotal evidence of instrumental goals and describes other reported incidents as instructive but not direct evidence. Those examples are therefore part of the case for asking better attribution questions, not empirical validation of the typology.
An additional layer of accountability
That caution carries into the proposed legal response. Legal agency is presented as a possible additional accountability layer, not a replacement for responsibility held by developers, owners or deployers. Under the proposal, considering legal agency would not replace the responsibility of those human roles.
The mechanics remain open. The paper does not settle what rights, duties, liabilities or enforcement mechanisms would attach to a system with legal agency. Nor does it decide whether legal affordances should ever be extended to AI; it frames that as a future question about whether an AI system could be held liable or accountable.
What the preprint leaves open
Because the work is a conceptual synthesis, its central output is a typology rather than an empirical estimate. The analytic material is theoretical positions in philosophy, ethics, legal theory and sociology, and no participant sample or dataset is described. The paper therefore provides a structure for reasoning about agency, not evidence that a particular system has been assigned one of the eight configurations.
The supplied analysis also identifies a practical gap: the paper does not provide an operational test for placing a real AI system in one of its categories. Future work would need to examine whether the three dimensions can be applied consistently to specific frontier AI systems and legal settings.
Other questions follow from the accountability proposal: how rights, duties, liabilities and enforcement would attach, and how persistent autonomous behavior should be distinguished from actions attributable to human developers, owners or deployers. The paper leaves these questions unresolved rather than supplying a legal design.
A preprint, not a verdict
The document is marked arXiv:2608.20041v1 and dated 20 Aug 2026. Its end matter reports use of large language models including ChatGPT and Claude for language editing and critical feedback, while stating that substantive arguments and final editorial decisions were made and verified by the author. Funding information is not reported in the supplied end matter.
For general readers, the paper’s conclusion is narrower than the debate around AI agency might suggest. It offers a conceptual map for separating questions about moral agency, legal agency and human accountability; it does not show that advanced AI systems are moral or legal agents. Its unresolved forward-looking question is whether an AI system could ever qualify as an individual legal non-human agent—and, if so, how liability and accountability could be structured without replacing the responsibility of developers, owners or deployers.
Paper data and sources
Original title: A three-dimensional typology of agency for advanced AI systems
Authors: Willem Fourie
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text