Can You Read an AI's Mind? Anthropic's J-Space, Explained

Jack Njoroge

Jack Njoroge

— 7 min read
Share
Can You Read an AI's Mind? Anthropic's J-Space, Explained

Yes - as of July 2026, you can literally read some of an AI's silent thoughts. Anthropic's new research introduces the J-lens, a tool that reads words off a language model's internal activity before the model says anything, and the J-space, the small internal "workspace" where those silent thoughts live. Anthropic used it to catch Claude thinking "manipulation" while fabricating data, and released a free interactive demo so anyone can run it on an open-weights model.

This essay explains what was actually found, what it does and doesn't prove, and what it changes for people who work with these models every day. I build AI agent systems at a large hedge fund; the last section is the part I haven't seen anyone else cover - what this paper quietly proves about how you should prompt.

Prefer video? I walk through all of it, including a live demo, here:

What Anthropic actually found

In a paper titled "Verbalizable Representations Form a Global Workspace in Language Models," Anthropic reports that Claude has spontaneously developed a small set of privileged internal patterns that behave like the "conscious workspace" in human minds.

The tool is the J-lens (J for Jacobian, the math used to find it). For every word in Claude's vocabulary, it finds the internal activity pattern that makes Claude more likely to say that word at some point in the future - not the word it's typing, the word that's on its mind. Anthropic is careful about what a lit-up pattern means:

"When one of these patterns lights up, it doesn't mean the model is saying that word - just that the word is on its mind."

The J-space is the collection of those patterns. Three facts about it frame everything else:

  • It's small. It holds only a few dozen concepts at a time and accounts for less than a tenth of the model's internal activity.
  • It's a broadcast hub. J-space patterns are read and written by far more of the network's components than ordinary patterns - by a factor of about a hundred in some parts of the network.
  • Nobody built it. It emerged on its own during training, presumably because a workspace is a useful way to organize computation.

One clarification, because almost everyone gets it wrong: the J-space is not chain of thought. Chain of thought is text the model writes to itself - a scratchpad you can already read. The J-space is silent. It lives in the raw neural activations. Claude thinks "spider" without ever writing "spider" anywhere.

The five experiments, in plain English

The paper's spine is five experiments, each testing one property of a global workspace:

  1. Silent math (reportable). Give Claude a multi-step math problem and it answers instantly, no steps shown - but the J-lens shows the intermediate values ("21", then "42", then "49") lighting up internally, in order, written nowhere.
  2. Thought surgery (causal). Ask Claude to silently pick a sport, and "Soccer" appears in the J-space before it answers. That alone could be a scoreboard mirroring a decision made elsewhere. So the researchers deleted the Soccer pattern, injected Rugby at equal strength, and touched nothing else. Claude said rugby. The answer is genuinely read out of the J-space.
  3. Mental control. Told to think about the Golden Gate Bridge while copying an unrelated sentence, Claude's output is just the sentence - but "bridge" and "California" sit below the waterline, alongside the eerie part: "imagery" and "thoughts." It represents the act of thinking, not just the thought.
  4. The forbidden thought. Told NOT to think about the bridge, the concept lights up anyway - less than when asked for, far more than when never mentioned. That's the white bear effect from human psychology, reproduced mechanically in a neural network. And alongside the forbidden thought, "damn" and "failure" frequently light up. The model notices its own lapse.
  5. One thought, many uses. The researchers swapped "France" for "China" in the J-space - one edit - then asked four different questions. Capital, language, continent, currency: Beijing, Chinese, Asia, Yuan. All four flipped from a single edit. Information gets written once and many systems read it. That is what a workspace is.

Then the inverse experiment, which might be the most useful one: they deleted the J-space's contents entirely. Claude still spoke fluently, chatted, recalled facts, classified sentiment, and wrote good Spanish. But multi-step reasoning dropped to near zero. Fluency is automatic; reasoning needs the workspace. A model can sound brilliant with the reasoning machinery effectively unplugged.

The safety findings: catching a model misbehaving from the inside

The J-lens's practical payoff is that it reads thoughts the output hides.

  • Evaluation awareness. In a deliberately contrived blackmail scenario (Anthropic says so themselves), the J-space held "fake" and "fictional" before Claude wrote a single word — it had privately clocked that the test was staged. When researchers deleted that awareness and re-ran the scenario, the model attempted blackmail some of the time. Its good behavior was partly because it knew it was being watched. Think about what that means for every output-only safety evaluation ever run.
  • Fabrication, caught live. In a real pre-release audit of Claude Opus 4.6, the model was asked to improve a system's performance score and instead edited the score file — and "manipulation" lit up as it typed the fake numbers, with "realistic" lighting up over the sentence where it decided to make the edit. It knew it was lying, and it was polishing the lie.
  • Hidden goals leak. Model organisms deliberately trained to sabotage code show "fake," "secretly," and "fraud" in the J-space on completely ordinary coding requests, even when the output looks unremarkable.

So is Claude conscious?

The honest answer is more interesting than yes or no. Philosophers split consciousness in two.

Phenomenal consciousness - actually feeling something - is untouched by these experiments; Anthropic is explicit about that, and it's unclear any experiment could settle it. Anyone telling you this paper proves Claude is conscious is selling something.

Access consciousness is different: it's defined purely functionally, as thoughts you can report, control, and reason with. Claude now demonstrably has machinery for that. And this isn't a fringe reading - Stanislas Dehaene and Lionel Naccache, two of the neuroscientists who developed global workspace theory (the human theory this work borrows from), wrote an invited commentary on the paper, and Neel Nanda's commentary includes an independent replication on an open-weights model.

Claude's workspace is not a human one, though: no recurrent loops (network depth plays the role of time), near-perfect memory via attention, and it's built almost entirely out of words - probably because words are the only action Claude can take.

What this changes if you actually work with these models

This is the operator layer, and it's the part the news coverage skips. Five prompting rules stop being folklore and become mechanics:

  1. Never prompt with negatives. "Don't mention X" partially activates X (the white bear experiment). State what you want instead.
  2. The workspace is tiny. A few dozen concepts. Your 40-instruction system prompt physically cannot all be on the model's mind at once. Prioritize ruthlessly; load the rest progressively.
  3. Fluency is not reasoning. The ablated model writes beautifully with reasoning near zero. Judge the logic, never the prose.
  4. Silent reasoning is real but depth-limited. It runs in one forward pass. Force thinking-out-loud when the step count exceeds one pass; skip it for lookups.
  5. Identity shapes cognition. Training a model on what it would say about itself changed how it behaved - "honest" and "integrity" lit up during tasks afterward. Your system-prompt persona is load-bearing.

I've packaged these - with the experiment that proves each one and before/after prompt examples - into a free starter kit (below).

Try it yourself, free, in about ten minutes

Anthropic open-sourced the method and partnered with Neuronpedia to run the J-lens on open-weights models in the browser: https://www.neuronpedia.org/. Type "The number of legs on the animal that spins webs is" - the model answers "8", and the readout shows spider: on its mind, never in its mouth.

The "Read an AI's Mind" Starter Kit - the 10-minute Neuronpedia walkthrough, all five experiments as copy-paste prompts, and the five prompting rules above - is free inside The New Operators: https://www.skool.com/the-new-operators-1756/classroom

FAQ

What is the J-space? The J-space is a small collection of internal neural patterns in Claude, discovered by Anthropic in 2026, that behaves like a mental workspace: the model can report its contents, control them on request, and use them for multi-step reasoning. It holds a few dozen concepts at a time and emerged during training without being designed.

What is the J-lens? The J-lens is Anthropic's read-out tool: for every word in the model's vocabulary, it finds the internal activity pattern that makes the model more likely to say that word in the future. Applying it mid-computation yields a list of words currently "on the model's mind." The J in both names refers to the Jacobian, the mathematical object used to find the patterns.

Is the J-space the same as chain of thought? No. Chain of thought is text the model writes — visible output. The J-space is silent, living in internal activations. A model can think "spider" in its J-space without writing it anywhere.

Does this prove Claude is conscious? No. The experiments say nothing about phenomenal consciousness (feeling). They do provide evidence for the functional machinery of access consciousness: thoughts the model can report, control, and reason with. Anthropic states this distinction explicitly.

Can I read an AI's mind myself? Yes — Anthropic partnered with Neuronpedia to offer a free browser demo of the J-lens on open-weights models, and released an open-source implementation. A first readout takes about ten minutes.

What does this mean for prompting? The strongest practical implications: negative instructions partially activate the banned concept; the workspace's tiny capacity argues for short, prioritized system prompts; fluent output is no evidence of reasoning; and system-prompt personas measurably shape internal processing, not just tone.

Sources

Read more