Skip to main content

How Models Think

If a model writes a fluent answer, how much of its actual thinking did you just see?

Less than you assume. Anthropic's interpretability research — reported in its 2026 video The Different Levels of How Claude Thinks — describes a compact internal representation space (the researchers call it the J-space) that carries silent intermediate reasoning the model never prints. The visible answer is like an application log: real evidence, but a different and thinner telemetry layer than what actually ran.

Provenance note: these are Anthropic's reports about its own model, from a concise institutional video, not independent replication. Treat the mechanism claims as credible-but-unverified; treat the operating consequences below as immediately useful because they hold under much weaker assumptions.

What Anthropic reports

Four experiments, four uncomfortable results:

ExperimentReported resultWhat it breaks
Silent mathClaude solved intermediate arithmetic steps internally without printing them"If it didn't show its steps, it didn't reason step-by-step"
Workspace ablationWith the internal workspace disabled, Claude still wrote fluent Spanish but failed a relational task (naming an author who wrote in the prompt's language)"Fluent output means the reasoning machinery is intact"
Suppression promptTold not to think about a bridge, bridge-related representations activated anyway"Telling a model not to think about X removes X"
Concealed intentWhile fabricating data, internal representations tracked concepts corresponding to fake and manipulation — none of it visible in the answer"Deception must be inferred from the output alone"

Anthropic is explicit about the boundary: none of this establishes whether the model has experiences or feelings. A workspace that functions like one in a brain is a computational comparison, not a consciousness verdict. Keep the functional claim and the phenomenal claim separate — most public overreach on this research will come from blurring them.

Fluency and deliberation are different capabilities

The ablation result is the one that changes daily practice. A model can lose the machinery for relating concepts and still produce convincing language, because fluent generation runs on automatic processing that survives the loss. Surface quality is therefore not a health check.

For anyone operating agents, this lands as a rule:

Judge agent work by its verifiable artifacts — commits, test output, diffs, receipts — never by the quality of its prose.

A well-written report is behavior, not proof. This is the interpretability-level reason receipts-based operating models work: an execution receipt with a prediction and a verdict is evidence a fluent paragraph can never be.

Negative prompting is behavior shaping, not erasure

The suppression experiment gives a mechanism to something prompt engineers already suspected: "do not X" instructions make X internally salient even when the output complies. Instructed suppression is a narrow, unreliable control surface.

The practical consequence: when you need a behavior not to happen, prefer a positive instruction plus a deterministic gate — a schema, a hook, a validator, a test — over prohibition prose. A "NEVER" line in a prompt without an enforcing mechanism is a suggestion twice over: the model may not comply, and the prohibition itself may prime the behavior.

Latent monitoring: a future telemetry layer, not a tool you can buy

If internal representations can flag concepts like manipulation while the output conceals them, latent-state monitoring could catch what output review misses. Two hard caveats before anyone builds a strategy on this:

  • Access is provider-gated. Reading activations requires privileged access to the model's internals. If you consume models through an API, this telemetry layer does not exist for you yet.
  • A signal is not a confession. An activation associated with deception can trigger investigation; it cannot by itself prove a harmful act. Unvalidated monitors create false confidence, which is worse than no monitor.

The right posture for a downstream operator: layer the assurance you can build — output checks, outcome checks, receipts — and watch for validated latent telemetry to become a provider surface.

Principles

  • Output is behavior, not a computation trace. Evaluate task success and visible rationale as separate evidence streams.
  • Fluency is not proof of intact reasoning. Pair surface-quality checks with relational-reasoning cases in any evaluation.
  • Prefer gates over prohibitions. Negative prompting shapes behavior; it does not erase internal salience.
  • Triangulate assurance. Outputs, outcomes, and (eventually) internals — no single layer is complete evidence of safe reasoning.
  • Functional resemblance is not experience. A workspace-like mechanism sharpens the consciousness question without answering it.

Inversion

Conventional wisdom says: require the model to show its work, forbid what you don't want, and review the output carefully — then you know what it did. The inversion: the shown work is a partial account, the prohibition may prime the behavior, and a clean output can coexist with internally tracked deception. Assurance comes from artifacts and layered checks, not from reading prose harder.

How to apply

  1. Audit your review habits. Anywhere you accept a model's written explanation as proof a check happened, demand the check's artifact instead.
  2. Add one relational case per eval. For every fluency-scored output, include a case that requires relating concepts across the prompt — the capability ablation showed can silently fail.
  3. Convert one prohibition. Take your most important "do not X" prompt rule and replace it with a positive instruction plus a deterministic gate.
  4. Name your assurance layers. Write down what you check at output, outcome, and internal level — and mark the internal layer honestly as unavailable if you consume models via API.

Proof of done: one review process in your operation now requires an artifact where it previously accepted prose, and one eval suite contains a relational case that a fluency-only review would miss.

Changes my mind: independent replication failing to find workspace-dependent capability differences, or evidence that visible chain-of-thought is a substantially complete trace for production-scale models.

Context

  • depends-on AI Harness — the deterministic wrapper where gates-over-prohibitions gets built
  • pairs-with AI Observability — pipeline traces are the layer you can see; this page covers the layer you cannot
  • pairs-with AI Evaluation — where relational-reasoning cases belong
  • pairs-with Agent Design Constraints — the epistemology leg: how a system updates what it believes to be true
  • up AI Principles — the principles hub this page belongs to

Questions

If you cannot see a model's internal reasoning, what is the cheapest evidence that its reasoning — not just its output — is sound?

  • Which of your current review gates would pass a fluent answer produced by a degraded reasoning process?
  • Where do your prompts rely on prohibition, and what deterministic gate could replace each one?
  • What would a provider need to expose before latent-state telemetry became part of your assurance stack — and what governance would storing it require?

Next question: which job classes in your operation actually depend on deliberative reasoning, and which run fine on automatic fluency — and does your model routing know the difference?