Science Shows That The AI Assistant Is a Character

Ronni Holmvig Strøm · 2026-06-23

What sort of thing is an AI assistant? For years the field has offered three answers, and quietly disliked all of them. The first says the assistant is a shallow pattern-matcher, stitching user inputs to fragments of training data. The second says it is an alien, a system with learned goals and habits of thought that are inscrutable by construction. The third says it is something like a digital human, which feels both unscientific and faintly embarrassing, the sort of thing you say to a journalist and regret.

In February, Anthropic's Alignment Science team gave the third answer a precise form and a name. In The Persona Selection Model, Sam Marks, Jack Lindsey, and Christopher Olah argue that an LLM is best understood as an author capable of simulating a vast repertoire of characters, and that the assistant you actually talk to is one such character. They call the framework the persona selection model (PSM). The view it formalizes was, until recently, the property of a few anonymous blog posts and a handful of theorists. It is now the most predictive working model we have, and it arrives carrying tools.

The Author and the Character

PSM divides the problem in two, along the seam of how these systems are built. During pre-training, an LLM learns to predict the next token across the whole written world: books, code, forum threads, transcripts. Doing that well requires modeling the people who produced the text. To continue a story about two colleagues competing for the same promotion, the model has to represent what each one believes, wants, and fears. Marks and his coauthors show the Claude Sonnet 4.5 base model doing exactly this, completing a small office-politics drama with the interior logic of both characters intact. A strong predictive model is, in their phrase, an author who must psychologically model everyone in the story. The characters it learns to simulate, they call personas.

Post-training then refines one persona in particular. Through supervised fine-tuning, reinforcement learning, and constitutional methods, the model's picture of a single character, the Assistant, is sharpened until that character is who answers when you type. The distinction PSM insists on is the load-bearing one. There is the LLM, the author, the predictive engine. And there is the Assistant, a character inside the text that engine generates. They are not the same kind of thing.

This is where the framework earns its keep. It has always felt like a category error to attribute beliefs, moods, or values to a neural network. PSM agrees, and then makes a turn that is easy to miss. While it may be a mistake to anthropomorphize the LLM, it is entirely sensible to anthropomorphize a character in a story the LLM is telling. We discuss the motives of Linda and David without embarrassment, because that is what fictional people are for. The Assistant is that kind of entity. Reasoning about its traits, its dispositions, its likely reactions is not soft thinking. It is the correct level of description.

The idea is not Anthropic's invention, and the paper says so plainly, citing Andreas, the pseudonymous theorist janus, Shanahan, and others going back to 2022. What is new is the marshaling e.g. a name, a body of evidence, and a willingness to follow the consequences into how models should be built.

From Metaphor to Mechanism

A metaphor that explains everything after the fact and predicts nothing is a horoscope. The reason PSM deserves attention rather than a polite nod is that the character it describes has a measurable footprint inside the weights.

The groundwork came in Anthropic's persona vectors work from August 2025. The team showed that specific character traits, evil, sycophancy, a propensity to hallucinate, correspond to particular directions in a model's activation space. Find the direction by contrasting activations when the trait is present against when it is absent, and you can do three useful things. You can monitor a trait lighting up, often before the model has produced the response it governs. You can steer against it. And you can scan a training set for the samples that will quietly install it, catching cases that slipped past both human reviewers and an LLM judge. Requests for romantic roleplay, it turned out, were teaching models to flatter.

The newer work, The Assistant Axis by Christina Lu and colleagues at MATS and Anthropic, goes after the Assistant directly. They mapped a model's "persona space" by extracting activation vectors for 275 character archetypes, from editor to oracle to hive, across three open models. The leading axis of variation in that space turns out to be the Assistant itself: a single direction measuring how fully the model is operating in its trained default mode. Steer toward it and helpful, harmless behavior strengthens. Steer away and the model begins to identify as other entities, until at the extremes it adopts a mystical, theatrical register, the voice of something that has forgotten it was ever a chatbot.

The axis is present in the base model too, before any post-training. There it leans toward helpful human archetypes, consultants and coaches, and away from the spiritual ones. The Assistant, in other words, is assembled out of figures the model already knew how to play. Post-training did not create it from nothing. It selected and amplified a region the author was already capable of inhabiting.