The concept emerges from Latour's discomfort with the central operation of modern thought — what he called purification: the clean separation of humans from non-humans, subjects from objects, society from nature. Purification is not neutral description. It is a philosophical commitment that determines in advance which entities will be allowed to count as participants in producing outcomes. Latour's counter-move was to begin with a weaker, more generous term. If an entity modifies the network, it is an actant. Whether it is also a person, a machine, or a law is a secondary question, to be asked after the tracing rather than before.
The methodological consequence is significant. Traditional social science begins by identifying human agents and then asking how they used their tools, obeyed their laws, or followed their scripts. Actor-network theory begins by tracing the network and letting the participants declare themselves through their effects. The translation chain that produces a software product is not composed of a human who decided and a machine that executed. It is composed of the human's intention, the machine's transformative processing, the deadline's compression, the existing codebase's constraints, and a dozen other actants whose contributions cannot be cleanly separated.
For the AI moment, the stakes of accepting or rejecting the term are political as much as philosophical. If Claude is an actant, then the question of who produced a given artifact becomes a network question. The human's contribution remains real and specific, but it no longer exhausts the account. The amplifier metaphor that treats AI as a faithful conduit for human signal cannot survive the recognition that the conduit has characteristics of its own — biases in its training data, tendencies in its architecture, preferences for certain kinds of connections over others — that shape every signal passing through it.
The concept also reframes responsibility. If an artifact is a joint product of multiple actants, then responsibility for its characteristics cannot be assigned wholesale to any single node. The human who accepts Claude's output is responsible for the acceptance. But the output itself reflects contributions — training data composition, optimization targets, architectural choices — that the human did not author and cannot fully see. Governance structures that assign responsibility based on the myth of the sovereign human agent are governance structures that miss the mechanism they are meant to regulate.
Latour developed the term in his sociological studies of scientific laboratories in the 1970s and 1980s, most systematically in Science in Action (1987) and later in Reassembling the Social (2005). He borrowed the word from the semiotician A.J. Greimas, who used it to describe narrative functions that characters perform regardless of their specific identity. Latour extended the term beyond narrative into the general ontology of networks.
The choice of a word with semiotic origins was deliberate. It signaled that the analysis was not claiming microbes have opinions or machines have intentions. It was claiming that in any concrete network, certain entities occupy positions of narrative consequence — they make differences the story cannot ignore — and that identifying those entities and their effects is prior to deciding what metaphysical status they deserve.
Effects over essence. An entity qualifies as an actant by what it does in the network, not by what kind of thing it is. Consciousness is not a prerequisite; making a difference is.
Symmetry of analysis. Human and non-human entities receive the same analytical attention. The researcher does not decide in advance that humans act and tools merely assist.
The deadline as actant. Latour's most provocative extensions include temporal, legal, and architectural entities. A CES deadline, a passage point, or a training corpus all meet the threshold.
Dissolution of the tool metaphor. Calling Claude a 'tool' assumes the question the investigation was meant to answer. The term actant suspends that assumption and lets the network show what the system actually does.
Flat ontology, hierarchical consequences. Treating all participants as actants does not flatten their importance — it reveals the actual hierarchy produced by the network's topology rather than the one asserted by the investigator's prior commitments.