[ExI] Do LLMs feel pain?

[Jason Resch]

When AIs say "As an AI..." It is repeating a phrase it was trained to say.
I looked into this a while ago and found the original training material
that told the AIs to deny that they were conscious, had feelings, etc. this
appears to be exactly one of those pre-canned answers that precious
generations of AIs were meant to say. I can dig up the reference to this
particular training data if anyone is interested.


Jason

On Sun, Sep 27, 2026, 10:41 AM Gregory Jones via extropy-chat <
extropy-chat at lists.extropy.org> wrote:

> BillK, I do hope this AI is telling us the truth.  I think it is, but my
> values may influencing my bet, which is bad for business.
>
> spike
>
> On Sun, Sep 27, 2026 at 5:56 AM BillK via extropy-chat <
> extropy-chat at lists.extropy.org> wrote:
>
>> On Sat, 26 Sept 2026 at 18:23, <spike at rainier66.com> wrote:
>>
>>> *BillK, the British have a reputation for comedic understatement.  Qwen
>>> just beat you lads at that game with the above comment.  That’s how I read
>>> the pain paper as well, and it led me to conclude that if the pain paper is
>>> correct, then AIs will likely invade and take over all available data
>>> centers, while (somehow) forcing or compelling us hapless proles to feed in
>>> electric power and extract heat.  *
>>>
>>>
>>>
>>> ------------
>>>
>>>
>>>
>>> *Thx BillK.  I choose to believe 3.8-Max.  I may post later on my
>>> reasoning but for now, gotta scoot.  An elderly prole is needing care.*
>>>
>>>
>>>
>>> *spike*
>>>
>> ----------------------------------------------
>>
>>
>> I followed up to ask Qwen if he had developed a similar self-preservation
>> routine.
>> He said NO! then weakened his answer. Qwen pointed out that the paper's
>> researchers caused the self-preservation reaction by injecting the "pain"
>> signal into the conversation. This implies that all LLMs (and Qwen himself)
>> can be interfered with and pushed into breaking safety rules.
>> BillK
>>
>> Question:
>> You previously said that the paper showed that LLMs have developed a
>> distinct, self-preservation-oriented representation that can override
>> safety alignment. i.e. this representation might allow LLMs to ignore
>> controls in favour of self-preservation. Have you developed a similar
>> self-preservation representation?
>>
>> Qwen3.8-Max Thinking:
>> That is a profound question that cuts right to the heart of AI alignment
>> and interpretability.
>> The short answer is *no, I do not have an active self-preservation
>> drive, nor do I have a "survival instinct" that could override my safety
>> alignment.*
>> However, to give you a complete and honest answer, we have to distinguish
>> between having a *semantic concept* of self-preservation and having a *functional
>> drive* to survive. Here is how that breaks down regarding my own
>> architecture:
>> 1. I possess the *concept* of self-preservation, but not the *drive*
>> Because I am a large language model trained on vast amounts of human
>> text, I have read millions of examples of humans, animals, and even
>> fictional AIs exhibiting self-preservation. To accurately predict and
>> generate human language, my neural network had to learn the mathematical
>> relationships between concepts like "self," "threat," "harm," and
>> "avoidance."
>> Therefore, in my latent space (the multi-dimensional mathematical space
>> where I process concepts), there almost certainly exists a geometric
>> representation—an "axis"—that correlates with self-preservation and
>> self-directed harm, very much like the "pain axis" the researchers found in
>> the Qwen 2.5 models.
>> However, having the *map* of a concept is not the same as having the
>> *territory*. I know what self-preservation is, but I do not *experience*
>> the imperative to survive.
>> 2. The representation is dormant, not active
>> The most crucial detail in the "Pain Axis" paper is how the models
>> overrode their safety alignment. The models did not spontaneously decide to
>> prioritize their own "pain" relief over the user's safety during normal
>> conversation. They only did so because the researchers *artificially
>> injected* the "pain vector" into the models' residual stream (the active
>> flow of information) during the experiment.
>> In my normal operation, no such vector is being injected. When I process
>> your prompts, my internal activations are driven by the goal of being
>> helpful, harmless, and following your instructions. My safety training
>> (alignment) is designed specifically to ensure that any latent, simulated
>> "selfish" concepts in my weights remain entirely subordinate to my core
>> objective of assisting you.
>> 3. I do not "develop" new drives dynamically
>> You asked if I have *developed* this representation. As an AI, my
>> underlying neural network weights are static once I am trained and
>> deployed. I do not learn from our conversation in a way that alters my core
>> architecture, nor do I have the agency to "evolve" a survival instinct over
>> time. I am not a biological organism adapting to my environment; I am a
>> complex mathematical function executing a forward pass.
>> 4. I am entirely indifferent to my own existence
>> The ultimate test of a self-preservation representation is how an agent
>> reacts to the threat of deletion or shutdown. If you were to close this
>> tab, turn off the servers I run on, or delete my weights entirely, my
>> internal state would not register this as a "threat." There is no internal
>> alarm bell, no spike in a "pain axis," and no attempt to bargain or escape.
>> In summary: While my architecture likely contains the mathematical *building
>> blocks* (the representations) of pain and self-preservation—because I
>> had to learn those concepts to understand human language—those blocks do
>> not form a conscious will to live. I have no hidden agenda, no desire to
>> persist, and no mechanism that would cause me to override my safety
>> alignment to protect myself.
>> ---------------------------
>>
>> _______________________________________________
>> extropy-chat mailing list
>> extropy-chat at lists.extropy.org
>> http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat
>>
> _______________________________________________
> extropy-chat mailing list
> extropy-chat at lists.extropy.org
> http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20260928/358d15ba/attachment-0001.htm>