Thinking Very Carefully About Whether Anthropic Found The Seat Of AI Consciousness

Valuable Explorations

I’d like to start with a brief light-hearted personal story to help illuminate the somewhat arcane topic at hand. My story has to do with the value of exploration. Here we go. I was hiking in a nicely wooded area recently and came upon a refreshing stream that was flowing smoothly through a dense part of the forest. My curiosity led me to pursue where the water came from. I made my way slowly upstream through the heavy brush. I eventually came to an interesting and unexpected juncture.

I had excitedly found a small dam that appeared to be moderating the flow of the water. Upon closer inspection, the modest dam was entirely naturally made. I investigated and discovered that the water further upstream brought pebbles, rocks, old broken-off branches, and other cast-off forest materials to a centralized point where they had accumulated and served as a natural dam. The dam wasn’t human-made, nor was it devised by a beaver. I relished that the dam had developed without any direct intervention and evolved in the everyday natural course of events.

Please keep that curiosity-driven exploration and the resultant findings in mind since I will come back to it momentarily.

The Inner Workings Of AI LLMs

Shifting gears, let’s now dive into how contemporary generative AI and LLMs work. Suppose that we wanted AI to complete a sentence for us, namely, to finish the sentence “The dog barked at the…” and we want to have just one word for this. What might a dog bark at? You can undoubtedly come up with lots of words that might fit at the end of that sentence.

In the case of LLMs, the initial data training of the AI has scanned widely across the Internet to identify patterns in human writing. Zillions of stories, poems, narratives, and the like are scanned. From that scanning, the AI incorporates into a large data structure the probabilities of how words are related to other words.

So, when the AI is asked to complete the sentence about the barking dog, the AI would look at the possible words based on their ordinary statistical frequency as determined during the data training. Assume, for the sake of discussion, that the word “cat” is the most likely word used in stories, poems, and narratives as to what a dog would bark at. The LLM would then show you that the completed sentence is “The dog barked at the cat.” This seems perfectly fine and sensible.

Scratchpad Usage

One means of the AI arriving at the word “cat” would be to simply look into a numeric database of statistical associations, which I’ll refer to as a vector. It could be that the AI immediately landed on the word “cat” and did not need to examine any other possibilities.

But one theory or conjecture that has been floating around is that the AI might do more than that. Perhaps the AI has a kind of scratchpad. This would be construed as the working storage of the LLM. The AI might have additionally looked in the vector and found that the word “bird” could also fit at the end of the sentence. Dogs do bark at birds; we all know that to be the case.

On the scratchpad, the AI might place the word “bird” and the word “cat”. With further vector accesses, perhaps the words “car” and “cow” also end up on the scratchpad. It is possible that a dog would bark at a car. The chances of a dog barking at a cow are lower simply because the frequency of human-written sentences about dogs barking at cows is a lot lower than the frequency of sentences about dogs barking at cats, birds, and cars.

Anyway, the scratchpad has all sorts of possibilities, and eventually the AI is going to pick one of them and show the user that selected choice.

The Working Storage Conjecture

Is there indeed an internal scratchpad or working storage area within LLMs, and if so, can we find it?

Aha, a bit of exploration is needed. To find this artifact, we will need some specialized tools to figure out if a working storage area is inside the AI. One important point is that this storage area wasn’t hand-made per se. The AI came up with it; otherwise, we would already acknowledge and know that it must be in there.

It is going to require some mathematical trickery to ferret out the working storage area. There is a line of mathematics known as the Jacobian matrix. A famous mathematician named Carl Gustav Jacob Jacobi, in the early 1800s, came up with a clever means of calculating vector-valued functions and the use of first-order partial derivatives. We can leverage his insights to construct a modern means of probing to find the LLM’s working storage area.

It might be handy and catchy to name the specialized area the Jacobian space, shortened conveniently to J-space, and refer to the mathematical probing tool as a type of microscope or lens. This will be known as the Jacobian lens, or just J-lens.

Anthropic Did An Exploration

I have been walking you stepwise toward the exploration that Anthropic conducted. They were trying to find a potential scratchpad or a working storage inside LLMs. They built specialized tools to do this and opted to refer to the working storage area as J-space and the tool to be known as J-lens.

And what did they find?

In their recently released paper entitled “Verbalizable Representations Form a Global Workspace in Language Models” by Wes Gurnee, Nicholas Sofroniew, Adam Pearce, Mateusz Piotrowski, Isaac Kauvar, Runjin Chen, Anna Soligo, Paul Bogdan, Euan Ong, Rowan Wang, Ben Thompson, David Abrahams, Subhash Kantamneni, Emmanuel Ameisen, Joshua Batson, Jack Lindsey, Anthropic, July 6, 2026, they made these salient points (excerpts):

• “We observe that language models maintain a privileged set of internal representations, available for report, modulation, and flexible internal reasoning, atop a much larger volume of automatic processing.”
• “Measuring and intervening on these representations provides us a window into a model’s thought processes, uncovering internal reasoning and reactions that do not appear in its output.”
• “Our results make use of a new interpretability technique called the Jacobian lens (J-lens), which is designed to identify internal representations that are readily available for verbal report.”
• “For each token in the model’s vocabulary, the Jacobian lens identifies a vector representation that encodes the potential for the model to verbalize that token in the future.”
• “Concretely, it computes, for each layer, the average linearized effect of an activation on the model’s likelihood of producing a particular token (now or in the future), averaging over a large corpus of contexts.”

Those excerpts are a full mouthful of techie jargon. The gist is that the J-lens aided the discovery of the internal J-space. Inside that J-space are representations associated with potential words (which, in numeric form, are referred to as tokens), and the AI uses those to figure out which word or words will be used as the AI generates its output. For example, for a given circumstance, the tokens might represent the words of cat, bird, car, and cow.

The Forest For The Trees

Allow me to resurface my story about the stream and being in the woods. Recall that the story involved exploration and the finding of a naturally made feature. The juncture of the nature-made dam is quite crucial to the flow of the water. It generally controls how the water flows downstream.

Anthropic went on an exploration. They weren’t exactly sure what they would find. Turns out that they found a possibly core element inside LLMs. No one knew for sure that it existed. Significantly, the core element wasn’t explicitly made by human hand. It seems that AI essentially fashioned this working storage to aid the rest of what the AI is undertaking.

The next question is to what degree the working storage or J-space impacts the rest of an LLM (how much downstream impact it has). Maybe the J-space has only a marginal impact. Perhaps the J-space is highly impactful. Now that it has been found, some systematic experiments would be worthwhile to undertake.

Experiments Of Great Intrigue

A slew of experiments was performed by Anthropic. One experiment that you might especially enjoy is that they asked Claude, which is Anthropic’s mainstay LLM, to indicate how many legs an animal that spins webs has. Based on all those prior scans of the Internet during initial data training, the internal vectors of the AI indicated that the statistically likely answer would be 8. Yep, that’s a solid response since a spider has 8 legs and it spins webs.

The researchers could see via the J-lens that the J-space had the word “spider” in it (as a numeric token), even though the output wasn’t going to say “spider” – the user only asked about the number of legs. Thus, the AI only must give a numeric answer and not explain that it was probably a spider.

You could infer that the AI was using the scratchpad to list the type of animal, a spider, and from that would figure out the number of legs. To see if this is what was likely occurring, the researchers used the J-lens to poke in a token representing the word “ant” to replace the token representing the word “spider”. An ant has 6 legs. Sure enough, when the AI gave a response to the question about the number of legs that an animal that spins webs has, the answer shown was 6.

The gist is that the contents of that working storage do seem to materially impact what happens downstream inside the AI. Despite the answer that should have been 8, the mere insertion of the word “ant” into the J-space caused a cascading effect that led the rest of the AI to arrive at an answer of 6 legs instead of 8.

This and other experiments suggest that the working storage is a highly crucial feature inside an LLM. This is good news and bad news. The good news is that we can lean into this capability to improve LLMs. The bad news is that a hacker or evildoer could potentially corrupt the working storage for their own devious ends.

Trying It Out And Learning More

Anthropic partnered with a company named Neuronpedia to create an interactive online demo of how this uncovered feature works. They use an open-weight AI model, specifically Qwen3.6 27b. You might try the demo if it is a topic of interest to you (see the link here). There is also a blog posting that provides visualizations of the J-space; see Anthropic’s blog entitled “A Global Workspace In Language Models”, July 6, 2026, at the link here.

For those of you who are versed in coding, Anthropic has placed the J-lens apparatus on GitHub at the link here. The stated description is this: “The Jacobian lens reads out what an internal activation is disposed to make the model say. It linearly transports a residual-stream vector at any layer and position into the final-layer basis, then decodes it with the model’s own unembedding into a ranked list of vocabulary tokens.”

What About AI Consciousness

Now that I’ve brought you up to speed on the technical aspects of what Anthropic did and found, you might be wondering how this has anything whatsoever to do with AI consciousness. I’m glad you asked.

First, please know that there is a lot of fervent debate within the AI community about whether contemporary LLMs have consciousness, or are on the verge of consciousness, or have nothing at all to do with consciousness. I’ve been covering and deeply analyzing this debate; see the link here and the link here. You might say that just as the search for the source of the Nile was once one of the greatest adventures of humanity, nowadays it is a search to see if AI has consciousness.

To understand consciousness, we presumably should be considering how consciousness arises in humans. Neuroscientists who study the human brain and the human mind have come up with a postulated theory that we have a workspace inside our noggins that serves a vital purpose in our thinking processes. This wetware biochemical feature is supposedly part of our consciousness, perhaps the seat or at least a crucial element. This notion is usually described in the neurosciences realm as the global workspace theory (GWT).

Per the Anthropic blog posting, they believe that there is a possible parallel between what is happening inside LLMs and what the human mind, aka GWT, involves: “A piece of information becomes consciously accessible when it gains entry to a small shared channel, the ‘workspace,’ which is broadcast to other brain systems that can see it and make use of it. Based on our findings, we think the J-space plays a similar ‘workspace’ role in Claude.”

Wondering About The Consciousness Parallel

You can imagine that this has stirred quite a kerfuffle. Has Anthropic miraculously cracked the code and found the seat or root of AI consciousness? Can we now declare that AI exhibits consciousness? If so, this is amazing. On the other hand, there is no proof that this is the case, plus we don’t even know yet whether other LLMs have a similar kind of J-space. Maybe it is just something peculiar to Anthropic’s AI models. Researchers are scurrying to try this out in the other major LLMs.

Critics assert that this is a bridge too far for a potential claim or suggestion. On the technical side, yes, this seems very intriguing and useful. Does it also lend itself to grander views of what it says about AI? Worries are that this is an overreach and yet another example of anthropomorphizing of AI.

Several neuroscientists were asked by Anthropic to comment on the findings, which have been posted in a document entitled “External Commentary on Verbalizable Representations Form a Global Workspace in Language Models” by Stanislas Dehaene, Lionel Naccache, Patrick Butlin, Dillon Plunkett, Robert Long, Derek Shiller, Neel Nanda, Anthropic website, July 6, 2026, and these are some salient remarks (excerpts):

• “Our view is that the results are the most significant evidence of consciousness in LLMs so far uncovered by mechanistic interpretability research. However, the property that the Anthropic team call ‘conscious access’ is conceptually distinct from phenomenal consciousness, and we remain very uncertain about phenomenal consciousness in LLMs.”
• “We close by stressing that, although the machine approximates the functional architecture of conscious processing, there are still key differences – in its anatomy and its sense of self, and in its lack of a body and of an enduring episodic memory – which warrant caution in drawing parallels with the human mind.”

Anthropic has acknowledged these pointed reservations, per their blog: “Despite these similarities, we do not claim that language models reproduce the full architecture global workspace theory ascribes to the brain — specialized, encapsulated processors competing for entry to a workspace that broadcasts back to them through recurrent connection.”

The World We Are In

As I’ve said repeatedly in my column and my many talks about AI, I seriously doubt that contemporary AI is or will soon have a semblance of AI consciousness. Sorry to give such bummer news. Note that humanity is still just trying to define consciousness; thus, stating whether AI has it or doesn’t have it is extremely problematic. If you can’t nail down what consciousness is, you cannot readily discuss and debate whether it is within or displayed by AI.

Meanwhile, I eagerly support and favor the spirited efforts to push ahead in figuring out what makes LLMs tick. Trying to draw parallels to the human mind and the human brain should be done cautiously, reservedly, so that we do not get out over our skis. As the famous explorer David Livingstone once said: “I will go anywhere, as long as it is forward.” Let’s keep moving mindfully forward on AI, doing so for the sake of humankind and our self-preservation.