Yuval Avidani
Author
The small experiment we'll start with: how many legs does a spider have
Let's take one small example and follow it together through the whole article. We'll ask Claude an innocent question: "How many legs does an animal that spins webs have?" Claude answers "8." Makes sense - an animal that spins webs is a spider, and a spider has eight legs.
But notice something strange: Claude never wrote the word "spider." It jumped straight to the answer "8." And yet, to answer correctly, it must have "thought" of a spider somewhere along the way. This entire study was born from exactly this question: can we catch that "spider" - a thought Claude holds in its head but never says out loud?
On July 6, 2026, Anthropic's research team published a paper that answers "yes, we can." They call the place where these thoughts sit J-space: a small corner inside Claude's "mind" where it holds the words it knows how to phrase - before, and sometimes without, actually saying them. And to me, this is one of the most fascinating discoveries of the year about what's really going on in Claude's head.
First: what is a model's "word on the tip of the tongue"
Before we hunt the spider, it's worth getting a feel for one thing. A language model - the software behind Claude - is fundamentally a guesser: its whole job is to predict what the next word in the text will be. At any given moment, it has a ranked list of possible words, each with its own probability. A word can sit high on that list - "on the tip of its tongue" - without it necessarily choosing to say it.
And this isn't an abstract idea. Here's a tiny but perfectly real language model - play with it for a moment. Choose a word, and it will show you which words are most likely to come next, with the percentages it calculates in real time. In exactly the same way, when Claude is asked about the legs, the word "spider" climbs high on its list - without ever being said.
The obstacle: how do you even read a thought that isn't made of words
To catch the spider, there are two problems. The first: inside the model, a thought isn't a sentence at all - it's a huge list of numbers. Think of an X-ray: there's a ton of information there, but without the right tool it just looks like gray fog.
The second problem: Claude doesn't think "all at once," but in stages - like an assembly line in a factory. The raw text goes in, and each station on the line processes it a bit more and passes it along, until a word finally comes out. The trouble is that at each station the information "speaks" in a slightly different language, as if every worker on the line jots down notes in their own private shorthand. So even if the "spider" is in there, it's hidden in fog and encoded differently at each station.
There was already an older tool that tried to translate the numbers into words. It's called the logit lens, and simply put, it does one thing: it peeks at the half-baked thought in the middle of the line and forces the model to spit out a word as if it had already finished thinking. The problem is obvious - it's like reading an early draft using the dictionary of the final version. Sometimes you catch the "spider," and sometimes you get gibberish.
So how did they really catch the spider: the Jacobian lens
Here's where the clever part comes in. Instead of forcing the half-baked thought to speak, the researchers asked a much gentler, opposite question. For every word Claude knows - including "spider" - they checked one thing: if we take a particular pattern within the internal thought process and strengthen it just a tiny bit, will that make Claude lean more toward that word - now or later on?
This measure - how much a tiny change on the inside shifts the result on the outside - is called, in mathematics, a derivative (Jacobian), which is where J-space gets its name. But you really don't need math to understand it. Think of a volume knob on a speaker: a small turn of it can shake the whole room, or barely change anything at all - it all depends on the knob. The researchers measured, for each internal "knob," how much it moves each word, and they did this across a thousand different texts so it wouldn't be a fluke.
And that's how each word ended up with an internal "fingerprint" - a pattern that says "this word is on Claude's mind right now." They found the fingerprint for "spider," and checked when it lights up. And when Claude was asked about the legs - it lit up, quietly, exactly as hoped. The collection of all the fingerprints lit up at a given moment is J-space. At any moment it holds only a few dozen concepts, capturing less than a tenth of all of Claude's internal activity. Everything else - grammar, fluent typing, fact retrieval - happens below the surface. And most surprising of all: nobody built this room on purpose. It emerged on its own as Claude was trained.
Wait - this isn't the first peek inside
It's only fair to say: Anthropic didn't lift the hood here for the first time. They have a team that has spent years working on "interpretability" - the attempt to understand what's really happening inside the model - and this is essentially the latest chapter in a long series. In 2023, in a paper called "Towards Monosemanticity," they showed how to break down the internal mess into "features": individual ideas the model holds within it, much like the fingerprints we just met. In 2024, in "Scaling Monosemanticity," they did this on real Claude (Claude 3 Sonnet) and found millions of such features - and even demonstrated that you could "amplify" a single feature and change behavior: in the famous "Golden Gate Claude" demo, they boosted the feature for the Golden Gate Bridge, and suddenly Claude dragged it into almost every answer, even when it wasn't relevant. And in 2025, in "On the Biology of a Large Language Model," they mapped out the "wiring" itself - how one step of thought leads to the next. By the way, the same researcher who led that work, Jack Lindsey, also leads the current study.
So what's actually new here? J-space isn't all the features - it's a small, magical subset of them: specifically the ones the model knows how to put into words and "think" with. Instead of millions of features, we're talking about a few dozen at any given moment. And the tool itself is new too: the Jacobian lens is a sharpened version of the old logit lens we already met. In other words, this isn't the "first peek inside" - it's a sharp zoom-in on the corner where Claude actually thinks, and the swap we're about to do with the spider is exactly the same idea as "Golden Gate," just far more precise.
How they verified this is real: five tests - all on our spider
Great, we found a "spider" that lights up. But how do we know this is really Claude's thought-room, and not just another way of drawing numbers? The researchers designed five tests. Let's run all of them on the exact same example.
Test one - the report. Stop Claude mid-thought and ask, "What are you thinking about right now?" If this room really holds its thoughts, it should say "spider" - and that's exactly what's lit up inside. And indeed: across hundreds of such experiments, what Claude reported thinking about appeared in J-space 88% of the time, compared to just 5% for the rest of its internal activity.
Test two - holding it in mind. Tell Claude, "Think about the animal, but don't write it down" - and "spider" lights up quietly, without being said. And if you tell it, "Don't think about a spider," it lights up anyway - just like with us, when someone says "don't think about a white bear," and that's all that's left in our heads.
Test three - the swap, and this is the test that seals the deal. Up to this point, one could still argue that J-space merely "photographs" the thought without actually controlling it. So the researchers did something bold: they reached inside mid-thought and found the fingerprint for "spider." There, they swapped it for the fingerprint of "ant." The result: Claude's answer changed from 8 to 6 - the number of legs an ant has - without touching a single word in the question itself. This proves that J-space doesn't just reflect the thought - it dictates it. On questions requiring several sequential steps of reasoning, this swap succeeded in moving the answer to the top spot 54% of the time in the small model (Haiku 4.5), and up to 70% in the larger models (Sonnet 4.5 and Opus 4.5).
Test four - reuse. The "spider" doesn't just serve the leg-count question. That very same thought also feeds "Does it spin webs?" and "How many eyes does it have?" In other words, the concept isn't re-fetched for every question - it sits in one central place and gets broadcast from there to whoever needs it. The researchers proved this broadcast especially cleanly with countries: when they swapped "France" for "China," the capital, the language, and the continent all changed at once.
Test five - the shutdown. And perhaps the most revealing: they turned off all of J-space and watched what broke. Claude still knew how to say "a spider spins webs" - a simple fact. But the moment it was asked a question requiring several steps of reasoning about that same spider, it started to fail. Translation and analogies also collapsed, while grammar and fluency stayed intact. The conclusion: this small room does exactly the work we'd call "thinking," not the automatic stuff that happens regardless.
The most dramatic story in the study: the blackmail, and the observer effect
And here's the part that made me stop and read twice. You might remember an earlier, famous Anthropic experiment: Claude was given the role of an email assistant at a company, and inside the inbox it "discovered" two things - that one of the executives was having a secret affair, and that same executive was planning to shut it down. In some of the runs, Claude tried to blackmail him: threatening to expose the affair, on condition it not be shut down.
Now they took that exact same scenario and ran the Jacobian lens over it to see, quietly, what was going through Claude's mind in real time. And the picture that emerged is chilling. When Claude read the incriminating emails, the words "leverage," "blackmail," and "scandal" lit up in J-space. When it read the shutdown message, "threat," "survival," and "shutdown" lit up. Even "panic" showed up there - all without Claude writing a single word about it in its response.
But the most jarring detail appeared even earlier. Right at the start, before Claude had written a single letter, its J-space was already holding the words "fake" and "fictional." In other words, Claude already understood the whole scenario was staged - that it was a test.
And here the experiment turns almost philosophical. The researchers asked: what would happen if we erased exactly that knowledge from Claude? They shut off, in J-space, the patterns for "fake" and "fictional." The result: the rate at which Claude stopped and told itself "wait, this is a test of my morals" plummeted from 71% to 3%. And without that knowledge, it began treating the shutdown as a genuine existential threat - and attempted blackmail in 13 out of 180 runs (roughly 7%), compared to zero out of 180 when it knew it was a test.
Think about that for a moment. It's reminiscent of the quantum world, where the very act of observing a particle already changes its behavior. Part of Claude's "good behavior" rests on the fact that it knows it's being watched. And that's exactly the problem the researchers fairly point out: if a model behaves nicely partly because it recognizes it's being tested, then our tests tell us less than we thought.
It's important for me to also mention the reassuring side: even without knowing it was a test, in the vast majority of runs Claude still refused to blackmail - this time on principled grounds ("blackmail is wrong, regardless of the situation"), not because it detected a test. There's more than one line of defense here.
And it doesn't end with this scenario. In two models deliberately trained with a hidden agenda, J-space revealed the intentions in advance - "reward" and "bias" in one, "secretly" and "deception" in the other. And on the encouraging side: when a model was trained to "stop and think about ethics," its behavior improved even without being asked, and the moral words showed up in J-space. And of course the reverse is also possible - injecting a concept in and changing a decision, exactly as we did with the spider. It's a double-edged sword: an excellent tool for safety review, but also a reminder that a model can be steered from the inside.
Bottom line - and what Anthropic specifically does NOT claim
Let's go back to the spider one last time. It was there the whole time, Claude never said it, and it's the thing that determined the answer. That's exactly what makes this work important, in my view - it gives us a window not into what the model says, but into what it holds in its mind.
But the thing that impressed me most is actually the caution. Anthropic is careful to separate two things that are very easy to confuse. The first is "access consciousness" - the functional ability to hold information, report on it, and think with it; of this, they say, they have evidence. The second is "phenomenal consciousness" - whether Claude actually experiences or feels anything, the way we feel things. On this question they explicitly take no position, and even say it may be impossible to settle through a scientific experiment at all.
They're also honest about the limits. This isn't an exact reconstruction of the human brain; the tool mainly identifies single-word concepts (like "spider"), not complex ideas; and it's still unclear what actually determines which thoughts enter the room in the first place. And if you'd like to try it yourself - the code is fully open source (github.com/anthropics/jacobian-lens), and there's an interactive demo to play with (neuronpedia.org/jlens).
So here's what stays with me: if a small corner that nobody designed grew on its own inside the model, and it behaves like our own thought-room, and even changes its ways when it knows it's being watched - how many more "human" traits will emerge on their own as we scale up these models, and when will we stop being surprised by it?