---
title: 'AI Might Be Conscious, But Not As We Thought'
source: 'https://youtube.com/watch?v=7ZvtDYYg0RQ'
video_id: '7ZvtDYYg0RQ'
date: 2026-11-11
duration_sec: 389
channel: 'Sabine Hossenfelder'
---

# AI Might Be Conscious, But Not As We Thought

> Source: [AI Might Be Conscious, But Not As We Thought](https://youtube.com/watch?v=7ZvtDYYg0RQ)

## Summary

This video explores the controversial question of whether large language models like Claude could be conscious, focusing on a new Anthropic paper suggesting a 'silent scratchpad' (J-space) that enables introspection. It presents both supporting evidence and skeptical counterarguments, including a novel theory that LLMs might be conscious during training but not during use.

### Key Points

- **AI Consciousness Claims** [00:00] — The video opens with the claim that AI will eventually become conscious, and mentions rumors that current AI might already be conscious, particularly during the training phase.
- **Dawkins and Amodei on Claude** [01:30] — Richard Dawkins argues Claude might be conscious because it's so good, while Anthropic's CEO Dario Amodei says they are open to the idea but don't know.
- **Anthropic's J-Space Discovery** [02:11] — Anthropic researchers found a 'silent scratchpad' (J-space) in Claude's internal activity, where it can store and use information for reasoning without outputting it, which is a requirement for global workspace theory of consciousness.
- **Manipulating J-Space** [02:54] — Researchers showed that changing the content in Claude's J-space changes its answers, e.g., replacing 'spider' with 'ant' makes Claude answer as if the animal had six legs.
- **Skeptical Counterarguments** [03:32] — NYU researchers tested three LLMs and found they couldn't reliably distinguish changes to internal activity from ordinary prompts, suggesting it's just another input method.
- **Hoel's Lookup Table Argument** [04:05] — Neuroscientist Erik Hoel argues LLMs are more like lookup tables than brains, and that they can't learn from new input, which is a key reason they are probably not conscious.
- **Consciousness During Training** [05:17] — The video suggests that if Hoel is right, LLMs might be conscious during training (when they learn) but not during use, though the host finds this distinction irrelevant.
- **The Measurement Problem** [06:00] — The host concludes that the question of consciousness is meaningless without a way to quantify it, giving the papers a 9/10 on the 'bullshit meter' for failing to address measurement.

### Conclusion

The video presents a balanced view of the AI consciousness debate, highlighting both the intriguing J-space findings and the strong skepticism from researchers like Hoel. The host ultimately argues that the debate is unproductive without a quantifiable measure of consciousness.

## Transcript

I'm convinced that artificial intelligence&nbsp; will become conscious eventually. There have&nbsp;&nbsp; now been new rumors again that the current AI&nbsp; might be conscious already until little notice&nbsp;&nbsp; paper that basically argues that they may be&nbsp; conscious not while you use them but during&nbsp;&nbsp;
the prior phase of training. So let's have&nbsp; a look. Claude can watch its own thoughts,&nbsp;&nbsp; but connected to the right tools. It can do&nbsp; something more useful. It can make your thoughts&nbsp;&nbsp; turn into reality. Our sponsor Higgsfield just&nbsp; released Seedance 2.5 and you can connect it to&nbsp;&nbsp;
Claude. I tried it. It really only takes about a&nbsp; minute. You just add it to the Claude connectors.&nbsp;&nbsp; Here we give it one prompt. Direct a 30-second&nbsp; film in one take about a physicist who finds a&nbsp;&nbsp;
thought in her head that isn't hers. Claude plant&nbsp; the shots and Seedance 2.5 generated what you're&nbsp;&nbsp; watching right now. It does this in one take&nbsp; or 30 seconds without awkward cuts. Same face,&nbsp;&nbsp;
same light, same room. And if a detail is wrong,&nbsp; you can fix that one detail without having to&nbsp;&nbsp; regenerate the entire thing. This used to be the&nbsp; part where AI videos fell apart, but Seedance 2.5&nbsp;&nbsp;
makes it work. Links in the description below.&nbsp; So, go and direct something. Thanks for watching.&nbsp;&nbsp; See you tomorrow. And now back to the science&nbsp; news. Claude is the chatbot made by Anthropic&nbsp;&nbsp; and it’s certainly the one that has attracted the&nbsp; most chatter. Richard Dawkins, author of the “God&nbsp;&nbsp;
Delusion” claimed in an essay in May that he&nbsp; can’t rule out that Claude is conscious. His&nbsp;&nbsp; argument is basically that he thinks Claude&nbsp; is so good that if this isn’t consciousness,&nbsp;&nbsp; then what do we even need consciousness for&nbsp; which, if nothing else, tells us that Dawkins&nbsp;&nbsp;
hasn’t spent much time with Large Language Models.&nbsp; But Anthropic’s CEO Dario Amodei himself has said&nbsp;&nbsp; “We don’t know if the models are conscious…&nbsp; But we’re open to the idea that it could be.”
It’s not just words. In a new paper that just&nbsp; appeared, Anthropic researchers say they found&nbsp;&nbsp; a small, special part of Claude’s internal&nbsp; activity that works like a silent scratchpad.&nbsp;&nbsp;
Claude can put information there and use it&nbsp; for reasoning, without ever outputting it. This is very interesting, because it means&nbsp; that Claude has a sort of internal life&nbsp;&nbsp; and is capable of introspection. It’s&nbsp; one of the necessary requirements for&nbsp;&nbsp;
global workspace theory, that is one of&nbsp; the leading theories for consciousness. More formally, the Anthropic team called&nbsp; the internal workspace the J-space,&nbsp;&nbsp;
where the J stands for Jacobian matrix,&nbsp; that’s the mathematical object that the&nbsp;&nbsp; researchers calculate to identify the space.&nbsp; The Anthropic researchers now showed that if&nbsp;&nbsp; they change what is in Claude’s J-space,&nbsp; Claude’s answer changes. For example if&nbsp;&nbsp;
Claude has internally worked out that the animal&nbsp; which spins webs is a spider, and the researchers&nbsp;&nbsp; replace “spider” with “ant,” Claude answers&nbsp; as if the animal had six legs, not eight.
One Anthropic researcher, Jack Lindsey,&nbsp; already claimed he had evidence for Claude’s&nbsp;&nbsp; introspection last year. The new paper now is&nbsp; a fuller analysis that also identifies exactly&nbsp;&nbsp;
where the introspection happens, namely in said&nbsp; J-spaces. It’s not just Claude. Another group&nbsp;&nbsp; likewise injected words into Llama’s internal&nbsp; activity and found that Llama also seemed to&nbsp;&nbsp;
take note of it. Probably a similar thing&nbsp; is going on with all Large Language Models. But not everyone is convinced. Researchers&nbsp; from New York University have challenged&nbsp;&nbsp; these claims. They tested this with three&nbsp; LLMs and found that the models could not&nbsp;&nbsp;
reliably distinguish a change that was&nbsp; made to their internal activity from an&nbsp;&nbsp; ordinary prompt. So they are saying it’s&nbsp; just yet another way to give them input. Then we have Erik Hoel, a neuroscientist&nbsp; who works on consciousness. He has written&nbsp;&nbsp;
an interesting paper in which he argues that&nbsp; today’s large language models are probably not&nbsp;&nbsp; conscious. We’ve heard this before, of course,&nbsp; but his argument is new. Hoel compares Large&nbsp;&nbsp; Language Models to lookup tables. Such “lookup&nbsp; table” algorithms are the classical examples&nbsp;&nbsp;
that demonstrate that observing input and&nbsp; output alone cannot tell you whether a&nbsp;&nbsp; system is conscious. It’s like looking up Chinese&nbsp; translations doesn’t mean you understand Chinese.
Hoel now argues that by way&nbsp; of computational structure,&nbsp;&nbsp; Large Language Models are much closer to&nbsp; lookup tables than they are to the human&nbsp;&nbsp; brain. They are stochastic input-output machines.&nbsp; Much of the autonomy that we assign to conscious&nbsp;&nbsp;
beings on the other hand is not directly a&nbsp; reaction to input. And Hoel says that the most&nbsp;&nbsp; obvious reason why LLMs are not conscious&nbsp; is that they can’t learn from new input,&nbsp;&nbsp;
though that has not stopped quite a few&nbsp; people from becoming senior management. This leads to a strange possibility. You see,&nbsp; large language models work in two stages. The&nbsp;&nbsp; first is the training in which they get&nbsp; fed a lot of input. Once this is done,&nbsp;&nbsp;
you freeze the fully trained LLM and&nbsp; use it as an input output machine.&nbsp;&nbsp; These are the apps you sign up&nbsp; for. Yes, they now use memory,&nbsp;&nbsp; but this is externally latched on as context to&nbsp; the prompt, it does not actually update the model.
If Hoel is right, then maybe large&nbsp; language models are not conscious&nbsp;&nbsp; when we use them. But maybe they&nbsp; are conscious during training. Personally I don’t think this is a relevant&nbsp; distinction. Imagine you have a human being&nbsp;&nbsp;
unable to form new memories, you wouldn’t conclude&nbsp; that therefore they are unconscious, would you? That said, to me the question of whether&nbsp; an LLM is conscious or not makes no sense,&nbsp;&nbsp;
the question is how we quantify its consciousness.&nbsp; This is why I give all these papers a 9 out of 10&nbsp;&nbsp; on the bullshit meter. They fail at the relevant&nbsp; question: if you can’t measure it, what are you&nbsp;&nbsp;
even talking about? Though I have to admit if&nbsp; Claude has thoughts it does not say out loud. This&nbsp;&nbsp; already puts it ahead of most people on social&nbsp; media. Thanks for watching. See you tomorrow.
