The Framing Is the Computation

The wrong question, answered fluently, is more dangerous than the right question answered poorly.

That claim has always rested on observation of human behavior. People bring badly framed questions to AI systems. The systems answer fluently. The answer feels like success, so the framing failure goes undetected. The deskilling is silent.

In 2026, Anthropic’s interpretability team published research that puts a mechanistic floor under this claim, on the machine side of the boundary. The finding, in one sentence: the question you ask an AI does not just shape the answer it gives you. It determines which internal representations the model deliberately reasons with.

The framing is the computation.

What the research found

Anthropic’s interpretability researchers identified what they describe as a workspace inside a language model: a small set of internal representations the model can reason with flexibly, as distinct from the much larger body of information it processes automatically.

The result that matters most is what they call directed modulation. Give a model the same passage of text and its internal processing of that passage changes depending on the question asked about it. Asked to predict the next word, the model uses the passage’s grammatical tense correctly, but the concept of tense never enters the workspace. Asked to name the tense, the concept appears throughout.

Same input. Same underlying information. Different question, different internal reasoning.

This is worth sitting with. It is not a claim about output phrasing or answer quality. It is a claim about which concepts the model actually computes with when it deliberates. The question selects the contents of the machine’s reasoning, upstream of any answer.

Two supporting findings extend the picture.

First, the researchers trained a model on what it would say if interrupted and asked to reflect on its own reasoning. That training changed the model’s behavior even when it was never interrupted. The representations used for verbal report turned out to be the same representations that govern silent reasoning. Articulation shapes cognition, not just output.

Second, models that wrote out intermediate reasoning steps were substantially more robust to internal disruption than models answering directly. Externalizing thinking onto the page is not documentation. It is genuine cognitive offloading, and it measurably changes what the system depends on internally.

Why more capable models make this worse

A reasonable hope: as models get smarter, they will compensate for bad framing. Read between the lines, infer what you meant, correct the question.

The directed-modulation finding cuts against that hope structurally. The model’s deliberate reasoning is composed by the question it receives. A more capable model answers a badly framed question with more fluency, more confidence, and more polish. It does not interrogate whether the question was correctly formed, because the question is what determined the contents of its reasoning in the first place.

Capability raises the stakes on framing. It does not substitute for it.

The layer before the prompt

Studio of the Mind is cognitive infrastructure for the AI era. It does not optimize prompts. It works on the human thinking that precedes them: surfacing hidden assumptions, interrogating the decision frame, structuring multi-perspective analysis before any question is committed to a model.

The workspace research does not test that human-side claim, and we will not pretend it does. It is research on models, not on people. But it establishes something adjacent and load-bearing: on the machine side of the human-AI boundary, framing is now demonstrably not cosmetic. The question determines the computation. Whatever quality of thinking produced your question is the quality of thinking your AI will reason with.

Two of the research findings also rhyme with mechanisms already built into the platform. Structured articulation changing downstream reasoning is the mechanism of the Decision Clarity Protocol, which forces the frame itself into words before any evaluation begins. Externalization as cognitive offloading is the mechanism of the Decision Artifact, a durable written record of the reasoning behind a consequential choice.

The upstream question is now demonstrably load-bearing on both sides of the boundary. One side is established by interpretability research. The other is where you live.

If you have ever received a fluent, confident, precisely wrong answer and only discovered the framing flaw after the decision was made, you already know what this layer is for.

studioofthemind.tech

Source: Anthropic interpretability research on workspace representations in language models, Transformer Circuits, 2026. transformer-circuits.pub/2026/workspace