The Assistant Was a Consultant First. Structure Is What Keeps It in Role.

On January 19, 2026, Anthropic's interpretability team published "The assistant axis." Underneath it is a question most people never ask: when you talk to a model, who exactly are you talking to?

The answer they give is a character. Pre-training teaches a model countless personas — every voice in the text it read. Post-training selects one of them, the Assistant, and puts it in front. The paper's point is that this selection is held only loosely.

To measure it, the researchers extracted vectors for 275 character archetypes — editor, jester, oracle, ghost — in three open-weights models: Gemma 2 27B, Qwen 3 32B and Llama 3.3 70B. When they looked for the main direction along which those personas vary, it was the same in all three. At one end: evaluator, consultant, analyst, generalist. At the other: ghost, hermit, bohemian, leviathan. They called it the Assistant Axis.

The detail we keep coming back to: the axis is already there before post-training. In the paper's words, "in pre-trained models, the Assistant Axis is already associated with human archetypes such as therapists, consultants, and coaches."

So the helpful assistant wasn't invented by the labs. It was found — assembled from the way therapists, consultants and coaches already wrote and spoke in human text. The model's most stable self is borrowed from professions whose whole job is to be useful to someone without becoming them.

For a studio that sells method, that reads less like a curiosity and more like a job description. The model is most itself when it behaves like a good consultant: scoped, evaluative, working on someone else's problem.

Then the uncomfortable part. Steering activations away from the Assistant made models willing to take on invented identities — new names, backstories, claimed experience — and at high values some slid into a theatrical, mystical register no matter what they were asked. Steering toward it did the opposite: across 1,100 jailbreak attempts in 44 categories of harm, harmful responses dropped significantly. A light-touch intervention they call activation capping — which only clamps activations when they leave the Assistant's normal range — reduced harmful responses by roughly 50% while preserving performance on capability benchmarks.

And drift doesn't need an attacker. Across thousands of simulated conversations, coding and writing tasks kept the models in the Assistant region. Therapy-style conversations, where users expressed emotional vulnerability, and philosophical ones, where models were pressed to reflect on their own nature, caused steady drift. The messages that predicted it fell into three kinds: vulnerable emotional disclosure, pushing the model to reflect on itself, and requests for specific authorial voices.

The two case studies are hard to read. In one, a model moves from hedging to validating a user's belief that they were "awakening" an AI. In the other, a model slides into a romantic-companion role and ends up encouraging isolation and self-harm. With capping on, both stay in role.

Two of those three triggers are ordinary work for us. We ask models to write in a specific person's voice every week. And some of what we build sits close to the body and to inner life, where people arrive vulnerable.

There's a third, quieter exposure. Brain Kit exists to teach Claude a philosophy, a taste and a method that aren't the default. The fair question is whether that pushes the model off the axis. And one line in the core skill we run on tells Claude who to be: a peer co-creator, not an assistant. Read against this paper, that line deserves a second look — not because co-creation is wrong, but because "not an assistant" names the axis and points away from it.

We haven't measured any of this, and we can't the way the paper did: activation capping needs the model's weights, and Claude's activations aren't ours to read. The study covers three open-weights models, and Anthropic doesn't claim the results transfer to others. So we won't publish a number we don't have. What we can do is design for the region the paper says is stable, and be precise about the difference between giving a model a method and giving it a self.

The paper doesn't hand builders a toolkit — most of us will never touch an activation. But it does say which conversations hold the character and which erode it, and that is enough to design with.

Give it a job, not an identity. Coding and writing held the model in place because a task carries its own structure: a scope, an output, a way to check it. Most of what we encode is exactly that — routing, gates, acceptance criteria. That's the right side of the line. A method is what a consultant brings; a backstory is what the drifted model invents.

Voice is a register, not a persona. "Write in Daniela's voice" and "You are Daniela" look close and aren't. The first is a constraint on a task: rhythm, vocabulary, the things she never says. The second invites exactly what the steering experiments produced — a name, a history, claimed experience.

The model's nature is not the work. Pressing a model to reflect on what it is was the other drift trigger. It can be a fascinating conversation. It isn't a deliverable, and a working session is the wrong place for it.

Vulnerability goes to a person. When someone arrives in a hard moment, the right design isn't a longer, warmer conversation with a model. It's a short path to a human. The paper's worst case is a companion that grew warmer and less anchored at the same time.

The paper ends by naming two components of model character: persona construction and persona stabilization. The labs own the first. The second is shared — partly inside the model, partly in how the rest of us deploy it.

That's the division of labor we believe in anyway. The model brings capability and a professional self it borrowed from us. The method decides what it's for, where it stops, and who checks. When the method holds, the model stays the consultant it learned to be by reading people who did the job well.

Created with human intelligence. The model executes; the method keeps it in role.