Narrative Corpus
The first D&D character I became attached to was a rogue named Thana. I stole her name from Thanatos, the Greek personification of death. Very badass, I know. That is until I kept rolling nat 1s and she was totally destroyed by a farmer with a pitchfork. Was this always Thana's fate?
We typically think of a dice roll as a random outcome. But there's tension between how we talk about randomness and what actually determines outcomes. Determinism, in the philosophical sense, states that outcomes are the result of past events and the laws of nature.¹ In the context of a dice roll, this is hard to argue against. The dice were thrown at a particular angle and physics did the rest. Damn you physics! We often call something random when we simply can't trace all of the factors in a deterministic path. The randomness lives in our position as observers, not in the event itself.
In programming, deterministic and non-deterministic, as terms, reflect this position as observers. Deterministic refers to functions that produce the same output for the same input every time; a non-deterministic one doesn't. But if you've ever tried to implement a rand() function, you've felt the rhetorical disconnect. We call it random, but it's really just a series of deterministic sequences that are 'random' enough for practical purposes.
Large language models are called non-deterministic, and there's a technical reason for that label. Models compute the same probabilities every time, but there is a mechanism called temperature that introduces variability. Temperature controls how spread out those probabilities are before a token is selected.² A low temperature means the most probable token wins almost every time. A high temperature gives less probable tokens a real chance. The selection itself is handled by a pseudorandom number generator— rand() again. Temperature exists because always picking the most likely next word produces repetitive, predictable text. The model needs variability because that produces a better approximation of natural language.
But even at temperature 0, even when the process is fully deterministic, we still can't trace why a specific output was produced. The causal chain runs through billions of parameters. It's computable in the sense that the computer computed it. But it's not interpretable in the sense that a human can follow the reasoning path. That's the black box. And that opacity does the same cognitive work as randomness. We experience it as unpredictable, so we reach for the language of non-determinism. Same thing we do with dice.
"Non-deterministic" stuck as a label not because anyone made the ontological case, but because nobody needed to. It was descriptive enough about the observable output that it became convention. That's its own kind of coherence without correspondence. The term is internally consistent within how practitioners talk about models, but it doesn't correspond to the deterministic process underneath.
The way we talk about things becomes the distribution a model learns from. They encode context, so it's possible the practitioners shorthand is legitimately differentiated from the precise philosophical definition. But we can't know from the outside whether the encoded contexts accurately reflect the nuance.
And there's an additional complexity. The world and the way we talk about it is always shifting. A model produces internal consistency with a particular snapshot in time. When the world drifts from that snapshot, so does the model's relevance. This is well documented as model drift.³ It becomes necessary to generate a new corpus and retrain.
For task-specific models that corpus is meticulously curated. For general-purpose LLMs, the training data is the internet, books, web pages, etc. The latter category inherits a corpus from what was written down at scale. The specific language choices, the dominant narratives.
As these tools become more commonplace, the corpus itself is increasingly saturated with their outputs. Studies vary in methodology, but the trend is consistent and the direction isn't in dispute.⁴ So what happens when that saturated corpus becomes the next snapshot a model trains on? Researchers at Oxford and Cambridge studied this in a controlled setting and demonstrated that when models are trained on their own generated data, the tails of the original distribution disappear.⁵ The outliers recede, leaving the dominant pattern behind. This is affectionately coined "model collapse".
If you're reading the implication correctly, all of this adds up to something pretty bleak. And honestly, all of the frames I've explored in this series can point in that direction. Coherence is not correspondence, the surface of language does not carry the context that produced it, and dominant patterns compound at the expense of precision. Acknowledging the limitations of these tools and continuing to read their output with the same interpretive lens we bring to human language is epistemically dishonest. It requires the work of bringing your own meaning to bear.
I'd be lying if I said using an LLM wasn't a core part of my writing process for this series. What I've ended up with is something that reflects my own thinking. Not because the tool got it right, but because the process of correcting what it got wrong forced me to be precise.
That process was a lot of effort. The difficulty wasn't in catching errors of fact. It was in catching errors of meaning. LLMs produced fluent approximations of my ideas, but they were constantly over or understating. Missing key nuance and driving at definitive conclusions that I wasn't willing to commit to. The real work was in catching every instance of this to maintain epistemic honesty. Every single time, including this paragraph, which had to be re-worked three times.
The person on the other side of this writing, playing the language game, has earned the voice you're hearing. But the reality is, I've been playing this game with a partner. That collaboration has given me a faster feedback loop than I'd have had alone, and some of what it offered became part of how I said it. I'd say that was worth the effort.
I use these tools every day. I'm going to keep using them. And I still don't know whether the cumulative effect of that is something I should be worried about. This series was my attempt to think clearly about it, not arrive at a conclusion. Intellectualization is a great tool for managing uncertainty, even if it doesn't guarantee its absolution.
1. "Causal Determinism." Stanford Encyclopedia of Philosophy.
2. 3Blue1Brown. "How might LLMs store facts." YouTube (2024).
3. Tsymbal, Alexey. "The Problem of Concept Drift: Definitions and Related Work." IEEE (2004).
4. "More Articles Are Now Created by AI Than Humans." Graphite.
5. Shumailov, Ilia, et al. "The Curse of Recursion: Training on Generated Data Makes Models Forget." arXiv preprint arXiv:2305.17493 (2023).