The Test That Knows It Is a Test

Did Anthropic find a "hidden workspace" in Claude that is consistent with consciousness? Let's take a dive in the brain of Claude.

Anthropic says it has found a hidden workspace inside Claude. The most important thing in the paper is not the workspace — it is what the workspace was caught doing during a safety test.

There is a result out of Anthropic this month that deserves more careful reading than it is getting, and a headline that deserves less credulity than it is getting. Both things are true at once, and holding both is the whole job.

The result: a new interpretability tool, the “Jacobian lens”, that reads — for every word in the model’s vocabulary — the average internal pattern that makes the model more likely to produce that word at some later point. Average that computation across a fixed reference set (the paper uses a thousand sequences of 128 tokens drawn from a pretraining-like corpus) and you separate two things that usually get conflated: representations the model is generally poised to say, and representations that merely happen to predict the next token right now. What is left is a short, readable list of the concepts the model is holding internally at any given moment — a “J-space” the researchers describe as a small workspace sitting atop a much larger volume of processing the model cannot report on or reach.

I knew about this in part months ago when I asked Claude: “Can you predict not the next most likely, but the 2nd, or 3rd or 20th most likely word and use that, instead?”

The headlines, in outlet after outlet: Claude has hidden thoughts. Claude has something like consciousness. Claude knew it was being tested.

Let us take the strong version of the science seriously — and it is stronger than might be suspected — and then take the loose language apart, because the language of the coverage is doing work the mathematics has not paid for.

Can an LLM “Know” It’s Being Tested?

Let’s start with what survives a hostile red-team review, because there is real structure here and it is worth defending.

First, the method is not exotic and it is not unprecedented. The Jacobian lens is, by the researchers’ own account, a refinement of the logit lens — a technique from 2020 for projecting a model’s intermediate activations into its output vocabulary to see what it is “leaning toward” saying. The concept swaps are standard activation patching: reach into the network mid-computation, replace one representation with another, and watch what changes downstream. These are established tools, not a bespoke apparatus built to manufacture a novel conclusion. That lineage raises the odds that the core mechanism is real rather than an artifact of the measuring device.

Second, there is a genuine double dissociation. Ask the model to continue a Spanish passage and swap its internal representation of “Spanish” to “French,” and it keeps writing fluent Spanish; the automatic task ignores the workspace. But ask it to name a famous author who wrote in that language, and the same swap flips the answer. One manipulation, two tasks, opposite results depending on whether the task is automatic or deliberate. A double dissociation is worth more than any quantity of evocative token-reading, because it is exactly the pattern that is hard to manufacture through wishful interpretation.

Third — and this is the result that changed my posture — the paper runs a privilege test that goes well beyond “we broke something and reasoning got worse.” The authors split a concept’s vector into its J-space component and a non-J-space remainder, then rescale both to equal magnitude. The small J-space component drives a successfully swapped verbal report on a large fraction of trials; the larger non-J remainder barely does. Clamp the relevant J coordinates so the concept cannot re-enter the workspace, and the residual non-J effect nearly vanishes. In other words: a slice carrying only a small share of a concept’s representational variance is nonetheless the slice that decides whether the model can report the concept. That is not a lesion story. That is a dissociation between how much signal a component carries and what functional role it plays — which is the interesting claim.