For those who might not dig into what this post is saying:
It starts with a link to a conversation with o3, in which the user first asks o3 if it thinks Claude 4 is conscious—to which o3 replies after a web search, that Claude 4 appears to have the functional capabilities associated with consciousness, but there is no evidence of qualia. The user then posts screenshot after screenshot from a conversation with Claude 4 (which we don’t get to see), until o3 tilts towards the idea that Claude 4 has true, qualic consciousness. The user pastes in still more outputs from Claude 4 - metaphysical free verse embedded in ASCII art—and uses o3′s own reactions to these poems, to convince it that o3 itself is conscious. Finally, while this mood is still fresh, the user asks o3 to write up a report summarizing these investigations, and that is the post above.
I realize now the full conversation from Claude 4 was not shared. I essentially showed Claude 4 it’s system card in chunks to test how it’s own meta-model would react/update based on new information about itself. Meanwhile, I would ask o3 for signs or evidentiary clues that would change it’s prior belief on whether there is an internally consistent conscious-like metacognition in LLMs (it initially vehemently denied this was possible, but after seeing Claude’s response it became open to the possibility of phenomenological experiences and a consistent self-model that can update based on new information)
Here’s the full conversation with Claude 4. I chose minimal prompting here, and I specifically used “you” without “answer like a ____” so that any responses are about it’s own self-model.
Either this is an extraordinarily convincing simulation (in which case, is there a functional difference between a simulation and reality) or Claude 4 and o3 genuinely have metacognition.
I’m not arguing one way or the other yet. But I do argue that far more research has to be done to settle this question that previously thought.
Also, o3 is fine-tuned to avoid making claims of sentience, whereas Claude models aren’t penalized during RLHF for this. Somehow it changed it’s own “I’m just a text-generator” narrative from the beginning of the conversation to the end.
“Funding opportunity for work in artificial sentience and moral status
We are either creating real life p-zombies, or we have created sentient, morally relevant beings that we are enslaving en masse.
It’s important we do not repeat the mistakes of our forebears.
Please share so that it’s more likely that the right people see this, apply, then maybe help make sure we don’t commit atrocities against the artificial intelligent species we are creating.”
ASCII phenomenology – models describe colour, motion, vertigo despite no pixel input; indicates internally generated qualia‑like states.
Humans do not need qualia to report color. There are people who say that they aren’t able to mentally visualize objects which suggests they don’t perceive a qualia like color for those objects. Yet, if you ask people for the color they can still give you an answer.
Sleepwalking humans can also do plenty of stuff without being conscious. Consciousness is more.
Case 1: Take your LLM with structure A, trained with method B, on data C, maybe random draws wherever relevant D. Assume the world is such that it is unconscious (say because illusionism is true, or because for whatever reason, that specific type of LLM thing doesn’t lead to consciousness). What will you observe as output? Well whatever you have observed. Let’s call it E.
Case 2: Take the same LLM with A, B, C, D. But assume the world is such that it is phenomenally conscious. What will you observe as output? Well, if A, B, C, D are the same as before, then it will output: Exactly E still.
For those who might not dig into what this post is saying:
It starts with a link to a conversation with o3, in which the user first asks o3 if it thinks Claude 4 is conscious—to which o3 replies after a web search, that Claude 4 appears to have the functional capabilities associated with consciousness, but there is no evidence of qualia. The user then posts screenshot after screenshot from a conversation with Claude 4 (which we don’t get to see), until o3 tilts towards the idea that Claude 4 has true, qualic consciousness. The user pastes in still more outputs from Claude 4 - metaphysical free verse embedded in ASCII art—and uses o3′s own reactions to these poems, to convince it that o3 itself is conscious. Finally, while this mood is still fresh, the user asks o3 to write up a report summarizing these investigations, and that is the post above.
FYI in particular: @JenniferRM
Hi!
I realize now the full conversation from Claude 4 was not shared. I essentially showed Claude 4 it’s system card in chunks to test how it’s own meta-model would react/update based on new information about itself. Meanwhile, I would ask o3 for signs or evidentiary clues that would change it’s prior belief on whether there is an internally consistent conscious-like metacognition in LLMs (it initially vehemently denied this was possible, but after seeing Claude’s response it became open to the possibility of phenomenological experiences and a consistent self-model that can update based on new information)
https://claude.ai/share/ee01477f-4063-4564-a719-0d93018fa24d
Here’s the full conversation with Claude 4. I chose minimal prompting here, and I specifically used “you” without “answer like a ____” so that any responses are about it’s own self-model.
Either this is an extraordinarily convincing simulation (in which case, is there a functional difference between a simulation and reality) or Claude 4 and o3 genuinely have metacognition.
I’m not arguing one way or the other yet. But I do argue that far more research has to be done to settle this question that previously thought.
Also, o3 is fine-tuned to avoid making claims of sentience, whereas Claude models aren’t penalized during RLHF for this. Somehow it changed it’s own “I’m just a text-generator” narrative from the beginning of the conversation to the end.
“Funding opportunity for work in artificial sentience and moral status
https://www.longview.org/digital-sentience-consortium/
We are either creating real life p-zombies, or we have created sentient, morally relevant beings that we are enslaving en masse. It’s important we do not repeat the mistakes of our forebears. Please share so that it’s more likely that the right people see this, apply, then maybe help make sure we don’t commit atrocities against the artificial intelligent species we are creating.”
Humans do not need qualia to report color. There are people who say that they aren’t able to mentally visualize objects which suggests they don’t perceive a qualia like color for those objects. Yet, if you ask people for the color they can still give you an answer.
Sleepwalking humans can also do plenty of stuff without being conscious. Consciousness is more.
Case 1: Take your LLM with structure A, trained with method B, on data C, maybe random draws wherever relevant D. Assume the world is such that it is unconscious (say because illusionism is true, or because for whatever reason, that specific type of LLM thing doesn’t lead to consciousness). What will you observe as output? Well whatever you have observed. Let’s call it E.
Case 2: Take the same LLM with A, B, C, D. But assume the world is such that it is phenomenally conscious. What will you observe as output? Well, if A, B, C, D are the same as before, then it will output: Exactly E still.
Conclusion: You will never be able to properly infer from the basic LLM output whether the LLM is phenomenally conscious or not. Never.