An AI doesn't know what it's going to say next. It predicts it. Understanding that one fact changes everything about how you work with these tools.
In the last issue, we established that an AI doesn't read your words — it eats them, chopping your carefully crafted sentences into tokens before processing a single syllable of meaning. Today, we go one level deeper. Because once the machine has its tokens, it has to do something with them. And what it does is so simple, so almost embarrassingly straightforward, that it will either make you laugh or make you stare into the middle distance for a few minutes.
It guesses what comes next.
That's it. That's the whole trick. Every word an AI produces — every sentence, every story opening, every piece of feedback on your manuscript — is the result of the model asking itself, over and over again: given everything I've seen so far, what token is most likely to come next?
One token. Then another. Then another. Until it's done.
You've seen this logic before. It's the autocomplete on your phone — the little grey suggestions that appear above your keyboard as you type. Type "Happy birth—" and it offers day, to, day! Your phone has learned, from your own messages and millions of others, which words statistically follow which.
A large language model is this idea taken to a scale that is almost impossible to comprehend. Instead of learning from your texts, it learned from an enormous swath of the written internet, digitized books, academic papers, forums, news articles, code repositories — hundreds of billions of words. Its predictions aren't just about what follows "Happy birth—." They encompass the deep statistical patterns of how human language works, across subjects, styles, tones, and centuries of writing.
The result is autocomplete that can write a sonnet, summarize a legal brief, and argue both sides of a philosophical debate. But at its core, the mechanism is identical to your phone keyboard. Predict. Output. Repeat.
Here is the thing that baffles most people when they first use these tools: the output is so smooth. It doesn't sound like a machine guessing. It sounds like a confident, capable writer. How?
The answer lies in what the model learned. Language, it turns out, is deeply, almost tyrannically predictable at the local level. After the word dark, certain words are far more likely than others. After dark and stormy, the field narrows further. The model has internalized these patterns at a depth and precision no human could consciously achieve — having processed more text than any person could read in a thousand lifetimes.
So when it predicts the next token, it is drawing on an extraordinarily rich map of how words relate to each other. The fluency isn't artificial smoothing. It's the emergent result of pattern mastery at enormous scale.
When a sentence flows, something is working — the syntax is right, the rhythm holds, the word choices belong in the same register. Fluency is evidence of craft and intention.
When an AI sentence flows, it means those tokens statistically follow each other with high confidence. Fluency is evidence of pattern frequency — not understanding, not intention, not truth.
That distinction matters. A lot. We'll come back to it.
If the model always picked the single most probable next token, every output would be deterministic — the same input would produce the same output, every time. Perfectly consistent, perfectly boring, and dangerously prone to getting stuck in repetitive loops.
So there's a setting — called temperature — that controls how adventurous the model's predictions are. Think of it as a dial between cautious and wild.
At low temperature, the model almost always picks the highest-probability token. Output is predictable, safe, and coherent — but can feel flat and repetitive. At high temperature, the model is more likely to reach further down its list of candidates and pick something surprising, unexpected, even a little strange.
Most consumer AI tools like ChatGPT or Claude sit somewhere in the middle — balanced enough to be coherent, loose enough to feel generative. When you're asking for creative writing assistance, that balance is doing a lot of work. And when the output surprises you in a good way, you can thank a touch of temperature for that happy accident.
Now for the confession buried inside all that fluency.
Because prediction is local — one token at a time, each one conditioned on what came before — the model has no grand plan. It is not holding a mental outline of your story. It is not checking the last paragraph against the first. It is not asking whether the claim it's about to make is actually true.
It is asking: what token is statistically likely here?
This is why AI "hallucinates" — a term the tech world uses when a model confidently states something false. It isn't lying. It isn't confused. It is doing exactly what it was designed to do: produce the most plausible-sounding continuation. And sometimes, the most plausible-sounding continuation happens to be wrong.
Notice the pattern: the model is most reliable when the task is generative — creating, shaping, imagining. It is least reliable when the task is factual retrieval of specific, obscure information. That's because generating plausible text and accurately recalling specific facts are very different challenges. The model was built for the former, not the latter.
Understanding prediction changes how you prompt. When you ask an AI to continue your story, it is not reading your intent — it is reading your words and predicting what comes statistically next in a passage that looks like yours. This means the specificity of your prompt matters enormously.
A vague prompt gives the model enormous latitude to follow high-probability paths — which tend to be the most generic ones. A precise prompt narrows the probability landscape and guides the model toward outputs that feel tailored rather than templated.
Think of it like casting. The more specifically you describe the role — the character's wound, their speech patterns, what they want and can't say — the better the performance you'll get. Leave it vague and the model will cast whoever it cast last time in something similar.
Three things worth taking into your practice:
1. Specificity is your steering wheel. The more context and detail you give, the more the model's probability landscape shifts toward what you actually want. Vague in, generic out.
2. The first few words carry enormous weight. Because each token conditions the next, your opening sets the trajectory. If you want a melancholy tone, start melancholy. Don't hope the model will find it three sentences in.
3. Treat factual outputs as drafts, not sources. The confident tone is a statistical artifact, not a warranty. Any specific fact an AI gives you — a date, a quote, a name — deserves a quick check before it ends up in your manuscript.
END OF ISSUE NO. 2 — AI & THE CRAFT