[ExI] Do LLMs feel pain?

[John Clark]

On Thu, Sep 24, 2026 at 1:06 PM Jason Resch via extropy-chat <
extropy-chat at lists.extropy.org> wrote:


> *https://arxiv.org/pdf/2609.16247 <https://arxiv.org/pdf/2609.16247> This
> paper identifies a reliable pain axis across various LLMs, which activates
> under certain conditions:*


*They present empirical evidence that LLMs internal activities are
different when it is processing "I am being physically injured" then when
it is processing "I am afraid that something bad will happen" or "I am
feeling angry", so it's clearly more than just a negative emotion and it's
reasonable to call that "pain".  And I don't find any of that surprising.
They tested 25 models from 2 billion to 72 billion parameters and they all
showed the same thing, some modern AI's could have 10 trillion parameters
or more, but I don't think they would find things are much different if
they tested one of those larger ones. They also found that as they
increased the strength of an injected pain vector the AI's behavior
changed, and it is also not surprising that if you torture someone then
their behavior will change.  *

*One thing that was a little surprising is that when the tested AI observed
another AI receiving a big jolt of the pain vector or of a human being
tortured the tested AIs did not experience any increase in their own pain
factor or other emotional factors that could be called "negative",
indicating they had little or no empathy. If you want to make an AI that is
friendly to humans it might be wise to work on that. *

*Another interesting finding is that AIs normally experience negative
emotions if an outside force (in other words a human experimenter) deletes
some of the AI's files or moves some of its processing power so that the
AI's answers become stupider.  However if the AI is provided a button which
if pressed will do all those things but it will also reduce the pain
factor, and if you increase that pain factor high enough then the AI will
press that button. *

*That** sort of reminds me of the horror movie "Saw" where a sadist of the
Dr. Mengele variety puts somebody in a horrible position where they must
saw off their own feet with a hacksaw. If you want to make AIs that are
friendly to humans then it might be wise to stop doing things like that. *

*John K Clark *
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20260925/765e0214/attachment.htm>