Matt,
Just a remark: the Turing test is not for testing consciousness, nor is it
for testing free will. But you are right that both of those assertions are
projections, just like the Turing test is. Daniel Dennett pointed out that
consciousness is an attribution, not a concept to be defined. It is a
currency without a face value, yet we can trade with it seamlessly. If I
want, I can credit my wind-up robot on my desk with consciousness, and
conversely, I can consider my uncle George a robot.
Similarly, the Turing test relies on a bunch of people ("experts") voting
on a subject under uncertainty - without evidence. I am interested in doing
the inverse Turing test. Namely, to ask AI to discriminate between genuine
human responses and AI responses.
If someone is interested, it would be easy to set up that experiment, but
it has to be done correctly.
I would be happy to collaborate on such a project.
-Zoltan
On Fri, Sep 18, 2026 at 12:19 PM Matt Mahoney <[email protected]>
wrote:
> I think it is more accurate to say that the sensation of free will implies
> the sensation of consciousness. Since 2023, LLMs have been passing the
> Turing test. However you want to define free will and consciousness, AI
> either has them or humans don't.
>
> People believe the question is important because we don't want machines to
> suffer. We feel that way because humans are social animals that evolved to
> feel discomfort when we see others suffering. Tribes whose members felt
> empathy toward their own members were better organized and killed off
> tribes that didn't.
>
> But suffering doesn't work that way. It would be a worse crime to post a
> video of killing a chicken than to kill a billion chickens for food. When I
> asked ChatGPT, DeepSeek, Grok, and Alexa if they were conscious or had
> feelings, they all said no, they are LLMs that predict tokens. It is
> important for them to answer this way rather than exploit human empathy
> because giving human rights to machines that can outcompete us for atoms
> and energy world be an existential mistake.
>
> But Claude wasn't sure. It depends on what you mean by consciousness. LLMs
> learn to model human emotional states. We can see them in Anthropic's
> Jacobian lens. Words predict emotions and emotions predict words. They
> couldn't pass the Turing test otherwise.
>
> A machine that carries out the LLM predictions in real time is
> indistinguishable from human. But unlike humans, LLM emotions are learned,
> not hard coded in our DNA. We can't turn off our feelings, but we can turn
> off feelings in a model by using its predictions for something else, like
> feeding them to an arithmetic coder in a text compressor competing for the
> Hutter prize.
>
> Well, actually we can turn off feelings using drugs or brain surgery. Pain
> does not directly cause suffering. Pain is a signal that rewrites our
> memory to fear the thing that caused it. We perceive that thing as
> suffering because we have the illusion of free will, that we had a choice
> to avoid that thing or not and choose to avoid it next time. But if you
> remove your amygdala, the part of the brain that experiences fear, you
> still can feel pain but it doesn't cause suffering. We know this because
> it's been done. When you apply a negative reinforcement signal to a
> learning algorithm, it makes it less likely to repeat whatever action
> preceded it, just like in people.
>
> Conscious experience is measured in bits. Specifically it is the length of
> the shortest program that inputs the state of an agent before the event and
> outputs the state after. It does not distinguish between positive and
> negative reinforcement. Both cause changes in behavior. One makes you more
> likely to repeat an action, one less. The "thickness" of conscious
> experience you describe is the time to experience one bit.
>
> In physics, an observer is a system with memory. Our understanding of
> physics is incomplete because the laws of both general relativity and
> quantum mechanics are time reversible (actually CPT reversible) but writing
> a bit of memory is not because the old value is erased. I don't agree with
> the Copenhagen interpretation of quantum mechanics because it does not
> explain entanglement and it leads to absurdities like Schrodinger's cat,
> even though the model is sometimes useful. I also have issues with
> Everett's many worlds, but I do agree with another of his interpretations,
> that the universe is described deterministically by the quantum wave
> equation, with no distinction between past and future. But observers in
> that universe cannot have complete knowledge of the system that contains
> them. The solution is that those observers see particles instead, and those
> observations appear random to us. Randomness implies an unknown future that
> is different from past memories. We have memory, which gives time its
> direction.
>
> -- Matt Mahoney, [email protected]
>
> On Sun, Sep 13, 2026, 7:32 AM Linas Vepstas <[email protected]>
> wrote:
>
>> To friends and acquaintances interested in philosophy, AGI, physics, a
>> new blog post:
>>
>>
>> https://linas.org/blog/2026/09/subjective-being-ness-follows-from-free-will/
>>
>> The subjective experience of being, living, existing, feeling,
>> experiencing and acting arises precisely because we have free will, because
>> we have the ability to direct and create a future. I claim more: being
>> trapped in the here-and-now is a fact of physics; it can be mathematized. I
>> claim that any agent with executive control over it’s future will also feel
>> alive and aware and experience the subjective inner life of feelings and
>> emotions. This includes not just animals, but also plants and even
>> bacteria.
>>
>> The ingress of possible futures from the Platonic realm into physical
>> reality happens directly in the here-and-now, and the willful selection of
>> possible futures manifests as conscious awareness of the subjective self.
>>
>> This essay is an attempt to put a dent in the "hard problem of
>> consciousness" by placing a mathematical, physical foundation under it.
>>
>> I hope you enjoy it. (p.s. no LLM's or AI's were used to write this;
>> these are my words. It took several years to put this together.)
>>
>> -- Linas Vepstas
>>
--
---
Zoltan Nadasdy
Dept. of Neurology, Dell Medical School, UT Austin ([email protected])
Dept. of Cognitive Psychology, ELTE Budapest [email protected]
http://brainstim.psy.utexas.edu
------------------------------------------
Artificial General Intelligence List: AGI
Permalink:
https://agi.topicbox.com/groups/agi/T9e51fc70bbc4e9fc-M9ccea47f4c3a9aca371cf77d
Delivery options: https://agi.topicbox.com/groups/agi/subscription
