{"id":4852,"date":"2026-09-15T14:37:30","date_gmt":"2026-09-15T12:37:30","guid":{"rendered":"https:\/\/inablab.org\/ai-wasnt-supposed-to-understand\/"},"modified":"2026-09-15T16:11:19","modified_gmt":"2026-09-15T14:11:19","slug":"ai-wasnt-supposed-to-understand","status":"publish","type":"post","link":"https:\/\/inablab.org\/en\/ai-wasnt-supposed-to-understand\/","title":{"rendered":"AI Wasn’t Supposed to Understand"},"content":{"rendered":"

[et_pb_section fb_built=”1″ _builder_version=”4.27.7″ _module_preset=”default” custom_padding=”1px||1px|||” da_disable_devices=”off|off|off” global_colors_info=”{}” theme_builder_area=”post_content” da_is_popup=”off” da_exit_intent=”off” da_has_close=”on” da_alt_close=”off” da_dark_close=”off” da_not_modal=”on” da_is_singular=”off” da_with_loader=”off” da_has_shadow=”on”][et_pb_row _builder_version=”4.27.7″ _module_preset=”default” custom_padding=”2px|||||” global_colors_info=”{}” theme_builder_area=”post_content”][et_pb_column type=”4_4″ _builder_version=”4.27.7″ _module_preset=”default” global_colors_info=”{}” theme_builder_area=”post_content”][et_pb_text _builder_version=”4.27.7″ _module_preset=”default” global_colors_info=”{}” theme_builder_area=”post_content”]<\/p>\n

<\/h1>\n

That story is getting harder to defend. Every time researchers look inside these models, they find things that behave like ideas. Abstract concepts, forming on their own, that nobody designed and nobody asked for. The most striking case came out of Anthropic this July (Anthropic, 2026). Deep inside Claude, they found a structure that neuroscientists have been arguing about since the 1980s. <\/span><\/p>\n

That story is getting harder to defend. Every time researchers look inside these models, they find things that behave like ideas. Abstract concepts, forming on their own, that nobody designed and nobody asked for. The most striking case came out of Anthropic this July (Anthropic, 2026). Deep inside Claude, they found a structure that neuroscientists have been arguing about since the 1980s. <\/span><\/p>\n

<\/span><\/p>\n

<\/span><\/p>\n

You are aware of information, not of hardware
<\/b><\/h2>\n

<\/span><\/p>\n

The structure matches Global Workspace Theory, which the psychologist Bernard Baars proposed in 1988, and which the neuroscientists Stanislas Dehaene and Jean-Pierre Changeux later developed into one of the leading theories of consciousness (Dehaene & Changeux, 2011).<\/span><\/p>\n

The core idea is simple. Your brain runs many specialized processes in parallel, and most of their work happens outside your awareness. What you’re aware of, the theory says, is the small amount of information that gets selected and broadcast widely across the brain, where many systems can use it at once. That’s what lets you report an idea, reason about it, and combine it with anything else you know. <\/span><\/p>\n

And the claim is about how information flows, not about the hardware it flows through. Neuroscientists have mapped the workspace onto real brain circuits, but nothing in the theory’s logic requires those circuits. In principle, any system organized the right way could have one. <\/span><\/p>\n

<\/span><\/p>\n

<\/span><\/p>\n

J-space, where Claude holds an idea in mind
<\/b><\/h2>\n

<\/span><\/p>\n

Anthropic found the same setup inside Claude. A small part of the model’s activity, less than 10%, works as exactly that kind of broadcast channel. They named it J-space. In plain terms, J-space is where Claude holds the idea it is currently working with. <\/span><\/p>\n

Three experiments make the case.<\/span><\/p>\n

First, ask Claude about “the animal that spins webs.” While it prepares the answer, the concept “spider” appears in J-space. Not the word, the idea. <\/span><\/p>\n

Second, researchers can reach in mid-thought and swap that pattern for the one meaning “ant.” Claude’s answer changes from eight legs to six. So the pattern is not a trace left behind by thinking. It is what the model is thinking with. <\/span><\/p>\n

Third, remove J-space entirely and Claude can still chat and recall facts, but every ability that depends on holding an idea in mind while working with it collapses. Summaries, step-by-step reasoning, poetry. All gone. <\/span><\/p>\n

Holding an idea in mind was supposed to be exactly what these systems lack. In this one respect, at least, what happens inside Claude looks less like autocomplete and more like what a human mind does. <\/span><\/p>\n

<\/span><\/p>\n

<\/span><\/p>\n

The interesting part was never the ingredients<\/b><\/h2>\n

<\/span><\/p>\n

That’s a strange thing for a next-word predictor to build on its own. It also weakens the most common dismissal of these models. “It’s just predicting tokens” is an argument about ingredients. Silicon instead of cells. Statistics instead of thought. <\/span><\/p>\n

But if Global Workspace Theory is right, the interesting part of a mind was never the ingredients. It was the organization. Claude has the wrong ingredients on every count, and the organization appeared anyway. <\/span><\/p>\n

Heard with that in mind, “it’s just predicting tokens” sounds a lot like “your brain is just chemicals.” True, and beside the point. <\/span><\/p>\n

None of this means Claude is conscious, and I’m not claiming it is. Consciousness is not where the news is. <\/span>The news is that the model builds abstract representations on its own and organizes them the way a leading theory of consciousness says the human brain does.<\/b> That’s the whole claim, and it’s already remarkable.<\/span><\/p>\n

<\/span><\/p>\n

<\/span><\/p>\n

An experiment you cannot run on a brain<\/b><\/h2>\n

<\/span><\/p>\n

Still, there’s something you can do with Claude’s workspace that you can’t do with a brain. Delete it and watch what breaks. <\/span><\/p>\n

In humans, the evidence for the theory is mostly indirect. Brain scans show correlations, and the closest things to causal tests come from accidents, split-brain patients, comas, anesthesia. You can’t switch off a person’s workspace cleanly, reversibly, and repeatably, and check exactly which abilities disappear. <\/span><\/p>\n

In Claude you can. Remove it, rewire it, rerun the experiment a thousand times before lunch. No theory of consciousness has ever had a testing ground like this. <\/span><\/p>\n

That testing ground may end up mattering for humans more than for machines. A large part of what goes wrong in conditions like ADHD and dementia involves precisely the ability J-space supports, keeping an idea active while you work with it. <\/span><\/p>\n

Attention slips. Threads of reasoning fall apart mid-step. Today we study those failures from the outside, through symptoms and scans. A model with a workspace gives us a mechanism we can damage with precision, in a hundred controlled ways, and each way of breaking it predicts a specific pattern of failure we can then look for in people. It won’t replace neuroscience. <\/span><\/p>\n

It hands neuroscience something it has rarely had, a working mechanical model of the thing that’s failing. <\/span><\/p>\n

This is the same bet behind the mindset model we’re developing at INAB, a simulation of how thoughts form and compete for attention. We build the mechanism precisely so we can break it on purpose. Weaken the grip of ideas the way ADHD seems to, degrade the machinery the way dementia does, and watch how thinking changes. <\/span><\/p>\n

Anthropic’s finding is encouraging for that whole program. It says that when a system learns to handle ideas, it builds machinery worth studying. And worth breaking. <\/span><\/p>\n

<\/span><\/p>\n

<\/span><\/p>\n

Where is all this heading?<\/b><\/h2>\n

<\/span><\/p>\n

Nobody knows. Some researchers believe these models are close to their ceiling and the surprises are mostly behind us. Others believe we’ve barely started to read what’s inside them. <\/span><\/p>\n

Will they ever reach something like consciousness? It still seems unlikely to me. But “unlikely” is also what we said about them writing code, passing medical exams, and building a global workspace on their own. <\/span><\/p>\n

Year after year, they keep surprising us.<\/span><\/p>\n

I’ve stopped betting on the line holding.<\/span><\/p>\n

<\/strong><\/span><\/p>\n

<\/strong><\/p>\n

<\/strong><\/p>\n

References:<\/strong><\/strong><\/span><\/p>\n