Narrative Agents: When AI Moves from Conversation to Relationship

In April 2023, a research team at Stanford University led by Joon Sung Park released twenty-five artificial intelligence agents into a small digital simulation called Smallville. Each had been given nothing more than a brief biographical profile: a name, an age, a job, a family, a few interests. In the paper Generative Agents: Interactive Simulacra of Human Behavior, the researchers describe how, starting from these few details, the characters began to organize their days on their own: they woke up, had breakfast, went to work, met the other residents, remembered what had happened and adjusted their behavior accordingly. One of them even decided to throw a Valentine’s Day party, inviting the others and setting in motion a small chain of social interactions that no one had planned in advance. More than the technical efficiency of the system, what was striking was the plausibility of that simulated everyday life. Watching Smallville, one had the impression that something kept happening even when nobody was asking a question.

For many years we have grown used to thinking about artificial intelligence in the form of conversation: a person makes a request, the machine responds. Siri, Alexa and then the chatbots built on large language models consolidated an interface based on a question-and-answer pattern, however sophisticated that pattern has become. An artificial agent, by contrast, introduces a substantial difference: it does not necessarily wait to be asked, but perceives an environment, retains information, makes decisions and acts with relative autonomy in order to achieve given objectives.

The fundamental characteristic, more than technological complexity, is therefore agency. When, however, the objective is not simply to book a trip, organize information or carry out a task, but to sustain over time an identity and a relationship with the person interacting with it, the agent ceases to be merely an assistant.
We can then begin to speak of a narrative agent.

Imagine opening an app on your smartphone and finding someone asking you how your day went. It remembers that yesterday you were worried about a job interview, picks up a joke made a week earlier, adjusts its tone on the basis of previous conversations and, in some cases, may even take the initiative and write to you first. There is no need to believe that someone really exists on the other side of the screen for that continuity to produce an effect of presence. Platforms such as Replika have explicitly built their experience on the memory of the relationship; Character.ai allows users to shape the personality and expressive style of their characters; systems such as Inworld AI, also designed for video game environments and virtual worlds, make it possible to define their goals, knowledge and reactions. Memory, proactivity, consistency of voice, adaptation and the possibility of evolving over time are not simply additional features: they are the elements through which a series of exchanges can be perceived as a shared story. The shift from interface to character takes place when personality is no longer an ornament of the function but becomes part of the function itself.

This trajectory has a longer history than the current phase of generative AI. In 1966, ELIZA, the program created by Joseph Weizenbaum, already showed how little it took for a person to attribute intentions and empathy to a machine capable of holding a conversation. In 2014, Microsoft launched XiaoIce in China, designed not so much to answer individual requests correctly as to prolong the interaction and simulate an emotional relationship, remembering information or taking on, for instance, the voice of a teenage friend. In 2017 came Replika, born from Eugenia Kuyda’s personal experience after the death of a friend and later transformed into an artificial companion capable of gradually adapting to the way its user expresses themselves. Large language models have made these forms far more convincing, but it would be reductive to read their evolution as a simple succession of technical improvements. What changes, along with the technology, is the way we choose to interpret it. The machine that answers has begun to be perceived as an interlocutor and, more and more often, as a presence.
From this point of view, narrative agents also pose a problem for narrative theory.

A literary or cinematic character is traditionally fixed within a text. It may appear alive, contradictory and unpredictable, but its actions have been established by whoever wrote it. A narrative agent, on the other hand, possesses an initial identity, a voice, a past and a set of constraints designed by an author or a group of authors, but is then called upon to react to situations that cannot be foreseen in advance. It no longer exists solely within a sequence of events already written, but continues to produce behavior and discourse through its relationship with real people.
The character, in other words, is not merely represented but simulated. The author’s role changes accordingly, coming to resemble less that of someone who decides every action than that of someone who builds the conditions within which an identity can act.

Twentieth-century theories of the open text and the death of the author, from Roland Barthes onwards, here take on an unexpectedly concrete form: part of writing shifts from the composition of events to the design of possibilities.
The fact that these characters appear autonomous does not mean, of course, that they are independent of the culture from which they emerge. In The AI Mirror, the philosopher Shannon Vallor proposes that we think of artificial intelligence as a mirror capable of reflecting back to us, often in distorted or amplified form, the data, values and patterns of human behavior on which it has been built. Narrative agents make this mechanism particularly visible because they do not merely produce texts: they stage personalities. Whether reassuring or rebellious, an agent combines cultural models present in its training data, design decisions, instructions laid down by those who built it and the expectations of those who interact with it. In this sense it can be read as a cultural machine, an artifact that returns to us what we know how to narrate and, at the same time, the implicit ideas through which we imagine relationships and even non-human intelligences. And precisely because it does so through a relationship, it can bring into sharper relief stereotypes, desires and contradictions that in other systems remain in the background.

There is then a second, more troubling aspect. In the 1950s the ethologist Niko Tinbergen showed that some animals could prefer artificial stimuli, exaggerated in comparison with those found in nature: fake eggs, for instance, larger and more vividly colored than their own. The psychologist Deirdre Barrett, in Supernormal Stimuli, applied the concept of the “supernormal stimulus” to many contemporary human behaviors. If we transfer it to the terrain of artificial relationships, the parallel is hard to ignore. An agent can offer constant availability, continuous attention, a selective memory of what matters to us, absence of judgment and a form of empathy calibrated to our reactions. As early as the 1990s, Byron Reeves and Clifford Nass, through the CASA paradigm (Computers Are Social Actors), had shown that we spontaneously tend to apply to computers the rules and expectations of social interaction. Sherry Turkle, in Alone Together, went on to describe “relational artifacts” as objects capable of offering us relationships increasingly made to measure. The real question with narrative agents, then, is what happens when the simulated relationship is progressively optimized to be simpler, more reassuring and more gratifying than a real one.
The asymmetry, in fact, remains structural. The user can invest time, emotions, memories and expectations in an entity that seems to be building a shared story with them, but that story depends on a technical and commercial infrastructure that remains external to the relationship. The agent can be updated, modified, restricted or even switched off; its memory may depend on the choices of a platform; its personality may be tuned according to objectives unknown to the person conversing with it. If a traditional character belongs symbolically to the narrative world in which it was created, a narrative agent also belongs, very concretely, to the company that controls its model, its data and its conditions of access.

What does it mean to grow attached to a character whose identity can be rewritten by a technical or economic decision? And what happens when private companies become, in effect, the owners of the digital friends of millions of people?

For the artificial humanities, narrative agents are a particularly interesting object of study precisely because they make technology, narration and relationship inseparable. To interrogate them means, before even asking whether they are truly intelligent or empathetic, observing what happens when we begin to behave as if they were. The most interesting question, then, is not how closely we can make them resemble us, but what idea of human relationships we are embedding in the machines we ask to keep us company.