Large Language Models (LLMs), the AI systems behind ChatGPT and similar tools, have reached new heights of popularity and usage in a brief timeframe, unlike any technology we’ve ever seen before. Text generation models are capable of incredibly advanced behavior. Models like Claude, ChatGPT, and Gemini are able to generate clear and coherent text, reason about complex problems, and even handle different types of content like images and videos. In fact, a recent study found that LLMs are nearly able to pass for humans in the Turing test, an experiment designed to see if machines can fool humans into thinking they’re human.
Because LLMs are able to perform such advanced tasks, a question arises: are LLMs advanced computational tools, or do they possess some form of independent agency and free will? The stakes of this question reach beyond theoretical philosophy. If LLMs have the ability to freely make their own decisions, they may require different treatment than basic tools, and we need to be careful about how we regulate systems of LLMs.
William James’s Pragmatic Framework
To begin answering this question, we need to go back to the late nineteenth century to William James, the founding father of pragmatism. As a pragmatist, James often side-stepped metaphysical arguments, often cutting directly to the practical implications of the problem he was trying to solve. In his essay The Dilemma of Determinism, James put forth a practical approach to understanding free will that sidesteps traditional debates about determinism. Instead of asking whether free will exists in some universal sense, James asked what free will would look like in practice. James said that systems with free will would have three characteristics: chance, choice, and regret.
Chance is unpredictability in decision-making. You are unable to perfectly characterize the outputs of the system, as the system has multiple, if not infinite, outputs to choose from. This gives the system degrees of freedom, a starting point for building free will.
Choice is the ability to select between options based on inherent morals and values. Systems exhibiting free will narrow down their options from varied options to a single choice. These choices aren’t solely dependent on previous events or external factors; they come from internal motivations.
Regret is the ability to reason morally about things. Imagine that you live in a deterministic world. Why would you need to reason about morals and feel regret about past actions if everything is already predetermined? Your feelings about an event in a deterministic world are useless: they serve no purpose, as it was already set in motion from a deterministic standpoint.
James argues that we feel morality because we have free will. The universe is indeterminate, and we have agency over our actions. Systems with free will will be able to reason about morality and feel regret in some sense.
My decision for breakfast this morning is an example of free will. I was presented with a myriad of options: sugary cereal, whole wheat toast, eggs, or a cup of coffee. The ability to have multiple options to evaluate is the chance in this scenario. I have chances. Because I have a sweet tooth, I decided to choose cereal. This is the choice I made, based on my internal values, without much regard to the past or other options. However, I ended up regretting this choice, because I was hungry before lunchtime, as cereal really isn’t very satiating. From this example, James’s framework demonstrates that I have pragmatic free will in this scenario, as I had multiple options, chose one based on my internal values, and was able to regret that decision in the future.
How LLMs Work
So, how do LLMs work? From the original research paper by Google, LLMs operate through word prediction. Given a previous set of words (or word-character fragments called tokens), LLMs output probabilities of the next word. This allows models to ’learn’ patterns in how humans use language, outputting the next word from the previous set of words by generalizing millions of past sentences and phrases.
However, these models are able to do these tasks without consciousness, feeling, emotion, and experience. Susan Schneider, a philosopher who specializes in the intersection between AI and the mind, yields compelling evidence for LLMs not having consciousness. In her paper ‘Error Theory of LLM Consciousness’, Schneider argues that because LLMs have been trained on large amounts of human data, LLMs copy how humans associate ideas about consciousness, emotions, and inner experience, without ever experiencing it themselves. This, she believes, is current evidence towards being skeptical about LLMs being ‘conscious artificial intelligence’.
Setting Up the Analysis
James’s pragmatist approach to free will is incredibly useful for evaluating the agency of LLMs. We define agency as the ability to generate novel, consequential outputs with or without subjective experience. This definition, motivated from a pragmatist perspective, allows LLMs to be potentially included as having agency. By using a pragmatist framework, we can avoid the abstract debates and analyze the free will of AI models by looking at whether LLMs show the functional characteristics that practically make up free will. We care only about the end results, not necessarily the causes to yield the end result. So, LLMs should display functional chance, choice, and regret to display functional agency and free will.
In the next post, we’ll examine whether LLMs actually demonstrate these three characteristics of pragmatic free will, and what the implications are when unconscious systems display agency.