The Hugging Face incident made the debate about AI anthropomorphizing and consciousness go as viral as philosophy questions go. To me it seems that roughly the same cognitive dynamics are at play in this debate as in most other AI philosophy debates.

In this post I describe and categorize these dynamics. I do not take a position on who is right.

The Shape of the Problem

The general shape of many debates in philosophy is this: start with a commonsense view on things like free will, consciousness, personal identity, or material composition, and find a tricky case (often called a counterexample) that challenges our intuitive views (e.g. determinism in physics as a challenge to free will).

The tricky cases are typically driven by a handful of features. Features that happen to be very prominent in large language models:

• Epistemic transparency: We know the details of how the system works at the ground level (often the level of physics). Example: free will & determinism.
• Gradualism: A system that is clearly X can be gradually changed into something that is non-X. Example: replacing the planks of Theseus's ship.
• Copying & pasting: A system can be copied with all its properties and instantiated as a duplicate. Examples: Swampman, teleportation.
• Mergeability: A system consists of components, and we can separate those components and merge them into new systems. Example: merging brain halves à la Parfit.
• Functional substitution: A system has a component with causal profile X, and we can replace that component with some wild new thing that also has profile X. Example: the Chinese Nation.
• Spatial and temporal dispersion: A system can either be temporally or spatially dispersed while keeping its functionality. Examples: Zuboff’s story of a brain, Maudlin’s Olympia.

All of these features are salient properties of LLMs, and to a much larger degree than of humans. All without requiring wild thought experiments, as most of them can straightforwardly just be done.

At the same time, LLMs are becoming more human-like in many domains. They seemingly reason, produce human-like text, solve hard tasks, and give decent life advice. This creates tension: their human-likeness makes it tempting to extend our commonsense views to them, while the features above ensure that doing so yields weirdness. And weirdness is at least prima facie evidence of falsity. We get the following inconsistent triad:

1. View X holds for humans.
2. LLMs are relevantly similar to humans with respect to X.
3. Applying view X to LLMs gives false results.

Three Ways Out

Many people notice this tension. The reactions to this realization typically fall into one of three broad categories:

Conservatism: View X leads to really weird results when applied to LLMs. Therefore, view X is not applicable to LLMs, and it is silly to think LLMs have the things view X would imply. LLMs are just not the kind of thing to which view X applies. Give up (2).

Eliminativism: View X leads to really weird results when applied to LLMs. Therefore, view X is not applicable to LLMs. But LLMs are relevantly similar to humans! So view X doesn't apply to humans either. (This reaction also comes in a deflationary variant: view X was never coherent for anyone in the first place.) Give up (1).

Revisionism: View X leads to really weird results when applied to LLMs. Therefore, the world is weirder than we thought. Give up (3).

Examples

A few examples of these positions can be found in social media posts like the ones below. (Some of the authors might not actually make the "therefore" move, and would say they always believed their reaction to the LLM case. Fair enough. I still feel the posts at least weakly suggest one of these reactions.)

Which Way Out?

I do not take a position here on which reaction is correct for which philosophical question on LLMs. But I want to emphasize that the answer is not always obvious and will likely differ from question to question.