Reading Nik Prassas earlier, I was reminded of Richard Dawkins’s recent transformation into full AI-evangelist. Dawkins changed after some cozy conversations with Claude, whom he refers to as Claudia. The language model’s attention (pun intended) made him think it was conscious, and he embarrassingly proclaimed that message to the rest of the world. I imagine a British version of Buddy the Elf, squealing “I’m in love, I’m in love, and I don’t care who knows it!” As many have quipped, the Claude Delusion is strong.
Nik Prassas’s piece is also strong. Speaking of Dawkins, he writes:
To summarise his argument in a sentence: Claudia can think and she is just an assemblage of material stuff and so you too must be an assemblage of stuff as well.
Prassas explores whether large language models make materialism more plausible or less, since they give credence to materialist theories of mind. He moves on to quote Donald Knuth, who suggests that thinking machines might provide evidence for God, rather than against. It’s a fun piece.
Thoughts About Thought
I have recently spent a lot of time thinking about thinking, through articles and videos discussing C. S. Lewis’s Argument from Reason. The core thrust of Lewis’s argument is this: if a conclusion can be fully accounted for by non-rational causes then it is not a valid rational inference, and since every conclusion can be fully accounted for by non-rational material causes under materialism, no conclusion is a valid rational inference. This includes the inference to materialism itself.
Furthermore, since all computational language models are deterministic cause-and-effect machines, all their conclusions are fully accounted for by non-rational causes. This renders their inferences non-rational. If C. S. Lewis is right, then computers can’t think.
Some have denied the first premise of Lewis’s argument, claiming that even if we have a full cause-and-effect account for how a belief was formed, that wouldn’t rule out the validity of the inference. It wouldn’t rule out intentionality and reason, they argue, since our brains are also cause-and-effect machines under materialism. Yet we are intentional and we reason. Dawkins would say the same about Claude…er, Claudia. While I am not persuaded that inferences could be rational if you arrive at them because of anything other than evidence, truth, and logic, let’s grant this point for the sake of argument. Let’s assume Dawkins is right, that Claudia thinks and yet is fully material. Let’s grant that knowing the step-by-step account of how its output was formed mechanically wouldn’t preclude the output from simultaneously being the product of intention and thinking.
Inconsistent Application
The fundamental inconsistency arises when we consider Dawkins’s views on evolution and intelligent design. Intelligent design is the theory that “certain features of the universe and of living things are best explained by an intelligent cause, not an undirected process such as natural selection.” (intelligentdesign.org) Dawkins dismisses intelligent design because of his unwavering belief in the power of natural selection, a blind watchmaker that transformed cells into cephalopods. According to Dawkins, we have a step-by-step accounting for how material physics produced animals, at least in principle, with no need to invoke an intelligent agent. For him, evidence for natural selection is evidence against God’s creative action.
But notice the inconsistency.
If the artifacts produced by Dawkins’s femme bot can simultaneously be the result of cause-and-effect physics AND intelligent agency, then why would natural selection’s presence prevent animals from simultaneously being the result of divine intelligence?
Do step-by-step mechanistic explanations rule out intelligent design or do they not?
Were Dawkins consistent, he would have to argue that the physics-based explanation of Claudia’s outputs are detailed enough to eliminate the need to invoke an intelligent agent as their cause. We have far more certainty that Transformer blocks1 produced Claudia’s tokens than we have that blind natural selection produced hummingbirds and ATP synthase.
Dawkins is playing favorites.
Not a God Delusion, but a God Conclusion
If step-by-step mechanical explanations do not rule out creation by intelligent agents, then I’m not sure we can dismiss the evidence for intelligent design in nature. Symbolic genetic systems like DNA are far more sophisticated than tokens from Claudia telling Dawkins she misses him. If the latter indicate production by intelligence, why not the former?
I’ve asked this for some time. In an AI paper I presented a decade ago (!), I argued that the presumed validity of the Turing Test carries implications for intelligent design’s scientific status. I wrote,
For the Turing Test to work, one must be able to distinguish intelligent causes from unintelligent causes based solely on observable artifacts. But this leads to the conclusion that intelligent design cannot be simultaneously disregarded, since its methodological structure rests on the same foundation.
If symbolic artifacts on a screen can tell us that an intelligence was involved, why can’t symbolic artifacts within our cells? We can’t have it both ways. Either artifacts can carry signs of intelligence or they cannot. If they cannot, then neither the Turing Test nor intelligent design detection is valid. But if they can, then maybe we should start paying more attention to what nature is telling us.
Consider the lilies of the field, and go to the ant, you sluggard.
Notes
- “Transformer blocks” are part of the internal large language model neural network in Transformer-based models. I don’t mean Lego blocks that are more than meets the eye.
Cross-posted from God & Machine on Substack.
Editor’s note: See also Jonathan McLatchie writing here, “On Minds and Machines, Richard Dawkins’s Curious Inconsistency.”