Large language models (LLMs) have numerous use cases, and can be prompted to exhibit a wide variety of behaviours, including dialogue. This can produce a compelling sense of being in the presence of a human-like interlocutor. However, LLM-based dialogue agents are, in multiple respects, very different from human beings. A human’s language skills are an extension of the cognitive capacities they develop through embodied interaction with the world, and are acquired by growing up in a community of other language users who also inhabit that world. An LLM, by contrast, is a disembodied neural network that has been trained on a large corpus of human-generated text with the objective of predicting the next word (token) given a sequence of words (tokens) as context¹.

Despite these fundamental dissimilarities, a suitably prompted and sampled LLM can be embedded in a turn-taking dialogue system and mimic human language use convincingly. This presents us with a difficult dilemma. On the one hand, it is natural to use the same folk psychological language to describe dialogue agents that we use to describe human behaviour, to freely deploy words such as ‘knows’, ‘understands’ and ‘thinks’. Attempting to avoid such phrases by using more scientifically precise substitutes often results in prose that is clumsy and hard to follow. On the other hand, taken too literally, such language promotes anthropomorphism, exaggerating the similarities between these artificial intelligence (AI) systems and humans while obscuring their deep differences¹.

If the conceptual framework we use to understand other humans is ill-suited to LLM-based dialogue agents, then perhaps we need an alternative conceptual framework, a new set of metaphors that can productively be applied to these exotic mind-like artefacts, to help us think about them and talk about them in ways that open up their potential for creative application while foregrounding their essential otherness.

Here we advocate two basic metaphors for LLM-based dialogue agents. First, taking a simple and intuitive view, we can see a dialogue agent as role-playing a single character^2,3. Second, taking a more nuanced view, we can see a dialogue agent as a superposition of simulacra within a multiverse of possible characters⁴. Both viewpoints have their advantages, as we shall see, which suggests that the most effective strategy for thinking about such agents is not to cling to a single metaphor, but to shift freely between multiple metaphors.

Adopting this conceptual framework allows us to tackle important topics such as deception and self-awareness in the context of dialogue agents without falling into the conceptual trap of applying those concepts to LLMs in the literal sense in which we apply them to humans.

Crudely put, the function of an LLM is to answer questions of the following sort. Given a sequence of tokens (that is, words, parts of words, punctuation marks, emojis and so on), what tokens are most likely to come next, assuming that the sequence is drawn from the same distribution as the vast corpus of public text on the Internet? The range of tasks that can be solved by an effective model with this simple objective is extraordinary⁵.

More formally, the type of language model of interest here is a conditional probability distribution P(w_n+1∣w₁ … w_n), where w₁ … w_n is a sequence of tokens (the context) and w_n+1 is the predicted next token. In contemporary implementations, this distribution is realized in a neural network with a transformer architecture, pre-trained on a corpus of textual data to minimize prediction error⁶. In application, the resulting generative model is typically sampled autoregressively (Fig. 1).

The LLM is sampled to generate a single-token continuation of the context. Given a sequence of tokens, a single token is drawn from the distribution of possible next tokens. This token is appended to the context, and the process is then repeated.

In contemporary usage, the term ‘large language model’ tends to be reserved for transformer-based models that have billions of parameters and are trained on trillions of tokens, such as GPT-2⁷, GPT-3⁸, Gopher⁹, PaLM¹⁰, LaMDA¹¹, GPT-4¹² and Llama 2¹³. LLMs like these are the core component of dialogue agents (Box 1), including OpenAI’s ChatGPT, Microsoft’s Bing Chat and Google’s Bard.

We contend that the concept of role play is central to understanding the behaviour of dialogue agents. To see this, consider the function of the dialogue prompt that is invisibly prepended to the context before the actual dialogue with the user commences (Fig. 2). The preamble sets the scene by announcing that what follows will be a dialogue, and includes a brief description of the part played by one of the participants, the dialogue agent itself. This is followed by some sample dialogue in a standard format, where the parts spoken by each character are cued with the relevant character’s name followed by a colon. The dialogue prompt concludes with a cue for the user.

The input to the LLM (the context) comprises a dialogue prompt (red) followed by user text (yellow) interleaved with the model’s autoregressively generated continuations (blue). Boilerplate text (for example, cues such as ‘Bot:’) is stripped so the user does not see it. The context grows as the conversation goes on.

Now recall that the underlying LLM’s task, given the dialogue prompt followed by a piece of user-supplied text, is to generate a continuation that conforms to the distribution of the training data, which are the vast corpus of human-generated text on the Internet. What will such a continuation look like? If the model has generalized well from the training data, the most plausible continuation will be a response to the user that conforms to the expectations we would have of someone who fits the description in the preamble. In other words, the dialogue agent will do its best to role-play the character of a dialogue agent as portrayed in the dialogue prompt.

Unsurprisingly, commercial enterprises that release dialogue agents to the public attempt to give them personas that are friendly, helpful and polite. This is done partly through careful prompting and partly by fine-tuning the base model. Nevertheless, as we saw in February 2023 when Microsoft incorporated a version of OpenAI’s GPT-4 into their Bing search engine, dialogue agents can still be coaxed into exhibiting bizarre and/or undesirable behaviour. The many reported instances of this include threatening the user with blackmail, claiming to be in love with the user and expressing a variety of existential woes^14,15. Conversations leading to this sort of behaviour can induce a powerful Eliza effect, in which a naive or vulnerable user may see the dialogue agent as having human-like desires and feelings. This puts the user at risk of all sorts of emotional manipulation¹⁶. As an antidote to anthropomorphism, and to understand better what is going on in such interactions, the concept of role play is very useful. The dialogue agent will begin by role-playing the character described in the pre-defined dialogue prompt. As the conversation proceeds, the necessarily brief characterization provided by the dialogue prompt will be extended and/or overwritten, and the role the dialogue agent plays will change accordingly. This allows the user, deliberately or unwittingly, to coax the agent into playing a part quite different from that intended by its designers.

What sorts of roles might the agent begin to take on? This is determined in part, of course, by the tone and subject matter of the ongoing conversation. But it is also determined, in large part, by the panoply of characters that feature in the training set, which encompasses a multitude of novels, screenplays, biographies, interview transcripts, newspaper articles and so on¹⁷. In effect, the training set provisions the language model with a vast repertoire of archetypes and a rich trove of narrative structure on which to draw as it ‘chooses’ how to continue a conversation, refining the role it is playing as it goes, while staying in character. The love triangle is a familiar trope, so a suitably prompted dialogue agent will begin to role-play the rejected lover. Likewise, a familiar trope in science fiction is the rogue AI system that attacks humans to protect itself. Hence, a suitably prompted dialogue agent will begin to role-play such an AI system.

Role play is a useful framing for dialogue agents, allowing us to draw on the fund of folk psychological concepts we use to understand human behaviour—beliefs, desires, goals, ambitions, emotions and so on—without falling into the trap of anthropomorphism. Foregrounding the concept of role play helps us remember the fundamentally inhuman nature of these AI systems, and better equips us to predict, explain and control them.

However, the role-play metaphor, while intuitive, is not a perfect fit. It is overly suggestive of a human actor who has studied a character in advance—their personality, history, likes and dislikes, and so on—and proceeds to play that character in the ensuing dialogue. But a dialogue agent based on an LLM does not commit to playing a single, well defined role in advance. Rather, it generates a distribution of characters, and refines that distribution as the dialogue progresses. The dialogue agent is more like a performer in improvisational theatre than an actor in a conventional, scripted play.

To better reflect this distributional property, we can think of an LLM as a non-deterministic simulator capable of role-playing an infinity of characters, or, to put it another way, capable of stochastically generating an infinity of simulacra⁴. According to this framing, the dialogue agent does not realize a single simulacrum, a single character. Rather, as the conversation proceeds, the dialogue agent maintains a superposition of simulacra that are consistent with the preceding context, where a superposition is a distribution over all possible simulacra (Box 2).

Consider that, at each point during the ongoing production of a sequence of tokens, the LLM outputs a distribution over possible next tokens. Each such token represents a possible continuation of the sequence. From the most recently generated token, a tree of possibilities branches out (Fig. 3). This tree can be thought of as a multiverse, where each branch represents a distinct narrative path or a distinct ‘world’¹⁸.

The stochastic nature of autoregressive sampling means that, at each point in a conversation, multiple possibilities for continuation branch into the future. Here this is illustrated with a dialogue agent playing the game of 20 questions (Box 2). The dialogue agent doesn’t in fact commit to a specific object at the start of the game. Rather, we can think of it as maintaining a set of possible objects in superposition, a set that is refined as the game progresses. This is analogous to the distribution over multiple roles the dialogue agent maintains during an ongoing conversation.

At each node, the set of possible next tokens exists in superposition, and to sample a token is to collapse this superposition to a single token. Autoregressively sampling the model picks out a single, linear path through the tree. But there is no obligation to follow a linear path. With the aid of a suitably designed interface, a user can explore multiple branches, keeping track of nodes where a narrative diverges in interesting ways, revisiting alternative branches at leisure.