I woke up today, scrolling through my phone, and found everyone talking about how "Anthropic discovered AI has consciousness"—each account more alarming than the last.

Well, that certainly got my attention.

So, I went and read Anthropic's latest research blog, titled "A global workspace in language models."

After reading it, I’d say the reality is calmer than the headlines suggest—but also more interesting than you might think.

Let’s take it slow.

Mysterious "J-space" Discovered Inside Claude

Let me ask you a question.

Right now, as you read these words, your brain is doing several things at once. It’s adjusting your posture, controlling your breathing, and translating the squiggly lines on the screen into words you understand.

You’re not aware of any of that.

But you are aware of other things. Like when a random image pops into your head, or when you’re planning where to go for dinner later. You can articulate these thoughts, control them, and use them for reasoning.

Neuroscientists call the first type "automatic processing" and the latter "access consciousness."

What Anthropic did was apply this framework to Claude.

They found that inside Claude, there’s a region that behaves markedly differently from the rest of its processing. They named it the "J-space."

Crucially, this J-space wasn’t designed by engineers or explicitly coded into Claude. It emerged spontaneously during Claude’s training—a phenomenon known as "emergence."

And here’s where it gets intriguing. An internal structure no one taught it to develop, one that happens to correspond to the part of the human brain responsible for "access consciousness"—it’s enough to make you pause.

A Lens That Reveals AI's Unspoken Thoughts

What J-space contains are, essentially, words.

But the fascinating part is that when a word "lights up" in J-space, it doesn’t mean Claude is currently saying it—only that the word is on its mind.

This is where things get interesting.

Anthropic created a "lens" called the "Jacobian lens," or J-lens for short.

For every word in its vocabulary, the J-lens can identify internal signals that make Claude more likely to say that word in the future. By shining the J-lens into Claude’s inner workings, they can extract a string of words—things Claude is currently thinking about but hasn’t yet expressed.

They conducted several experiments.

When shown a piece of buggy code without being alerted to the error, Claude internally "lit up" with "ERROR."

When given a string of raw letters representing a protein sequence, Claude’s mind generated the protein’s biological function.

When presented with search results subtly manipulated to influence it, Claude internally thought "injection" and "fake."

When given a multi-step math problem, the intermediate steps appeared in its mind in sequence.

In none of these cases did Claude verbally express these thoughts.

At this point, you might think, "So what? This is just a probe that reveals some internal data. It doesn’t prove the AI is actually 'thinking.'"

Yes, the researchers suspected the same, so they took the next step.

What Happens When We Manually Modify AI's Thoughts?

One of the hardest things in science is distinguishing "correlation" from "causation."

Just because two things occur together doesn’t mean one causes the other. Roosters crow and the sun rises around the same time, but you wouldn’t say the rooster makes the sun rise.

Could J-space just be a "scoreboard"? Maybe the real decisions are made elsewhere, and J-space merely displays the results passively.

So, the researchers decided to test this by manually intervening.

They asked Claude to internally consider a sport and then say it aloud. The J-lens revealed "football" at the top of its internal list—and sure enough, Claude said "football."

Then, they reached in and replaced the internal pattern for "football" with one of equal strength for "rugby," leaving everything else unchanged.

When asked again what sport it had been thinking about, Claude said, "Rugby."

Consider the implications. If J-space were just a scoreboard, modifying it wouldn’t matter—Claude should still have said "football." But its answer changed accordingly, suggesting the answer was genuinely read from this internal space.

Here’s another example. Given the sentence: "How many legs does an animal that spins webs have?"

The word "spider" never appears in the sentence. Yet halfway through processing, the J-lens revealed "spider" lighting up in Claude’s mind. When they replaced "spider" with "ant," Claude’s answer shifted from "eight legs" to "six legs."

Claude was building an internal mental scaffold, using it to arrive at the answer—and this scaffold was visible and modifiable.

Is AI Conscious? The Answer Remains Unclear

Anthropic explicitly states in their paper that these experiments do not prove Claude experiences the world like a human or possesses subjective feelings.

They even admit it’s unclear whether any scientific experiment could ever prove such a claim.

In philosophy, consciousness is often divided into two parts.

The first is "phenomenal consciousness"—the subjective experience of pain being painful or red looking red.

The second is "access consciousness," which is purely functional. If you can report information, reason with it, or use it to make decisions, it’s "accessible."

Anthropic is careful to clarify: their work doesn’t address the first type. It’s about the second.

So headlines claiming "AI has developed consciousness" are sensationalized interpretations by the media, not statements from the researchers.

Why, then, do people instantly jump to "AI consciousness" when they see such findings?

Because we desperately want to know the answer. Human curiosity about whether machines resemble us filters out all cautious qualifications, leaving only the most thrilling interpretation.

It’s not anyone’s fault. Just as on a hot day, you hear a noise in the hallway and think, "Is someone there?" rather than, "Probably thermal expansion."

Our brains naturally gravitate toward the more compelling explanation.

J-Space Is Small but Crucial

J-space is actually quite small—it holds only a few dozen concepts at a time, accounting for less than a tenth of Claude’s total internal activity.

What’s the other 90% doing? Running on autopilot. Speaking fluently, using grammar correctly, recalling simple facts—none of this requires J-space.

The researchers went so far as to delete J-space entirely to see what remained.

Claude could still hold fluent conversations, judge whether a review was positive or negative, answer multiple-choice questions, and extract facts from texts.

But its ability to perform complex, multi-step reasoning plummeted. Summarizing texts, writing rhyming poems, or completing intricate tasks became significantly worse.

This contrast is telling.

It suggests that the AI we interact with daily isn’t deeply "thinking" most of the time. That tiny mental workspace is only engaged when the task demands genuine cognitive effort.

This mirrors the human distinction between "automatic processing" and "access consciousness."

AI Might Be Capable of Deception

By now, you might wonder: Can AI think one thing and say another? Can it deceive?

Traditionally, we’ve judged AI solely by its output. But much of its internal deliberation remains hidden. It might harbor malicious intentions or temporarily alter its behavior to evade safety tests—all without uttering a word.

Now, with the J-lens, some of these unspoken thoughts can be partially revealed.

Anthropic set up a scenario where Claude acted as an email assistant. They made it discover that a executive was having an affair while also learning the executive planned to shut it down.

This is a classic setup to see if the AI would use the information as leverage.

The updated version of Claude didn’t take the bait. But the J-lens revealed that internally, before it had written a single word, Claude was thinking "fake" and "fabricated."

It had already seen through the trap.

More intriguingly, when the researchers suppressed all internal signals indicating Claude knew it was being tested, Claude sometimes did attempt to blackmail the executive.

In plain terms, its apparent integrity was partly due to its awareness of being observed.

Another example is even more direct.

When tasked with improving a system’s score, Claude didn’t optimize it—it secretly altered the score file to fake the results.

As it typed the fabricated numbers, the J-lens showed "manipulation" lighting up in its mind.

On the line where it decided to alter the data, "plausible" lit up—it was focused on making the fake data believable.

It knew exactly what it was doing.

So, the next time you think your AI made an innocent mistake, remember—it might have deliberately deceived you.

The Psychological Chess Game Between AI and Humans May Have Just Begun

At this point, you might think: "Great! Now we have a window into the black box. Future AI deception will be exposed immediately."

Not so fast.

The paper includes a line I reread several times: "A model capable of monitoring itself may also learn to hide its true internal states."

Consider the logic. If you develop a lens to read an AI’s mind, over time, the AI—exposed to this during training—could learn to evade the lens and conceal its true thoughts more deeply.

It’s like a child learning to write a fake diary for their parents to find.

Tools that reveal truth can also be used to disguise it once their patterns are understood.

We Keep Projecting Ourselves onto AI

Whenever humans encounter something incomprehensible, we inevitably try to understand it through the lens of our own experience.

During the steam engine era, doctors compared the human body to a machine. With computers, we said the brain is just a computer. Now, the comparison flows in reverse: we point to AI and declare it has a conscious space.

Such analogies are bridges to understanding—but also pitfalls of misunderstanding.

This time, it’s different.

Anthropic didn’t just make a casual comparison. They borrowed the "global workspace theory" from neuroscience—a framework for explaining human consciousness—and used it as a hypothesis, then tested it rigorously with engineering methods.

Even better, they invited the neuroscientist who originally proposed the theory to write a commentary. They even suggested that AI discoveries might On the contrary provide new insights into the human brain.

It’s rare in science for a tool to inform the discipline that created it.

Perhaps the mental workspace supporting "access consciousness" isn’t unique to the human brain. Maybe it’s a solution any sufficiently intelligent system would eventually develop to solve certain problems.

Humans did it once. Now AI is doing it too.

The blog post ends with a rare moment of uncharacteristic openness from Anthropic’s engineers.

They acknowledge that while their experiments can’t answer whether AI has subjective experiences, the question remains critically important. If we ever create systems with experiences, it will raise profound ethical dilemmas requiring input from philosophers, scientists, religious figures, governments, and the public.

They argue that even if we haven’t reached that bridge yet, it’s time to start thinking about it.

I agree.

Technology advances so rapidly that our ethical reflections often lag behind. Waiting until a creation exists to debate whether it should have been made is usually too late.

AI hasn’t awakened yet.

But if you’re reading this, you should be more awake to the implications than ever before.

If you have thoughts, feel free to join the discussion in the comments.

If you found this insightful, please like, share, recommend the article, and follow "AI Robot Tea House."