What Are LLMs? How They Work and Their Limitations
Every article in this Artificial Intelligence section assumes you understand, at least roughly, what's actually happening when you type a question into ChatGPT or Claude and get an answer back. This article is that foundation. It's written for authors, not engineers — no computer science background required — and once you understand the handful of core ideas here, the strengths, weaknesses, and quirks of every AI tool covered elsewhere in this section will make a lot more sense.
Before going further, it's worth being direct about where ScribeCount stands on AI, since that position shapes how every article in this section is written. We believe AI is a genuinely useful production and organization tool for the business of being an author — research, editing feedback, brainstorming, scheduling, formatting assistance, marketing copy drafts you'll revise yourself. We do not believe AI should write your story for you. The creative work — the decisions, the voice, the thing readers are actually paying for — should remain yours. Every article in this section is written from that position, and understanding how these tools actually work, covered below, makes clear exactly why that distinction matters.
What "LLM" Actually Means
LLM stands for large language model. Breaking down each word helps: "language" means it works with text; "model" is simply a computer science term for a system built to represent or simulate something, in this case how language works; "large" refers to the sheer scale of the system — modern LLMs are built from hundreds of billions of internal numerical settings, called parameters, that shape how the model processes and generates text. Put simply, an LLM is a very large, very complex mathematical system that has studied an enormous amount of text and learned statistical patterns in how language works — and uses those learned patterns to predict and generate new text in response to whatever you type in.
That last phrase — predicts and generates new text based on learned patterns — is the single most important idea in this entire article. An LLM is not thinking, reasoning, or understanding your book the way a human editor would. At its core, it's an extremely sophisticated pattern-matching and prediction system. Everything else in this article follows from that one fact.
How an LLM Actually Learns
Before an LLM can generate anything, it goes through a process called training. The model is shown a staggering volume of text — a meaningful fraction of all the publicly available writing on the internet, plus licensed books, articles, and other sources — and is repeatedly given a simple task: predict the next word, based on everything that came before it. When it predicts wrong, its internal settings (the parameters mentioned above) get nudged slightly toward a better prediction. This happens billions of times, across billions of examples, until the model has internalized deep statistical patterns about grammar, facts, reasoning structures, and writing style — not because anyone explicitly taught it grammar rules, but because it absorbed the patterns from sheer repetition and scale.
After this initial training, most modern AI tools go through an additional refinement stage, where human reviewers evaluate the model's responses and guide it toward being more helpful, accurate, and appropriately cautious. This is part of why a raw, untrained prediction engine and a polished product like ChatGPT or Claude feel so different to use, even though the underlying technology is fundamentally the same prediction process.
Tokens: How AI Actually "Reads" Your Words
Computers don't understand words the way you do — they work with numbers. So before an LLM can process anything you type, your text gets broken into small chunks called tokens, which are roughly (but not exactly) equivalent to word pieces — sometimes a whole word, sometimes part of one. Each token gets converted into a string of numbers that captures something about its meaning and how it tends to be used, based on patterns learned during training. Words used in similar contexts end up with similar numerical patterns — which is part of how the model is able to recognize, for instance, that "furious" and "livid" are related concepts, without ever being given a dictionary.
This tokenization process matters practically because it's also how AI tools measure and charge for usage, and how they define their context window — the topic covered next.
The Context Window: An AI's Working Memory
The context window is the total amount of text — measured in tokens — that an AI model can actively consider at one time. Think of it as the model's working memory. Everything within that window (your prompt, the conversation so far, any document you've shared) is available to the model when it generates a response. Anything beyond that window is effectively invisible to it — not stored somewhere else and consulted later, just genuinely not part of what the model can currently "see."
Why AI Tools Sometimes Make Things Up: Hallucination
"Hallucination" is the term for when an AI tool confidently states something false — a fake citation, an incorrect fact, a book that doesn't exist — presented with exactly the same fluent, confident tone as something true. This isn't a bug that occasional updates will fully fix; it's a direct consequence of how these models fundamentally work. Remember, an LLM is a prediction system, not a fact-lookup system. When it generates text, it's predicting what words plausibly come next based on learned patterns — and a plausible-sounding wrong answer and a plausible-sounding right answer can look, structurally, very similar to the model. The system has no built-in mechanism to know with certainty that it's wrong, because it isn't retrieving stored facts the way a search engine or database does — it's generating text that fits the pattern of what an answer should sound like.
⚠ This is the single most important practical limitation to understand before using any AI tool for research, fact-checking, or nonfiction content. Never treat an AI tool's factual claims as verified simply because they're stated confidently and fluently. Always verify specific facts, dates, statistics, and citations against a real source before they go in your book. Tools like Perplexity, covered later in this section, are specifically built to reduce this risk by grounding answers in cited, real-time sources — but even those tools aren't infallible, and verification habits matter regardless of which tool you use.
Why Different AI Tools Give Different Answers
If you've ever asked ChatGPT and Claude the same question and gotten meaningfully different answers, that's not a malfunction — it's expected, and understanding why helps explain the differences covered throughout this section. Different tools are built on different underlying models, trained on different data, refined through different processes, and configured with different default instructions (often called a system prompt) that shape their tone and behavior. There's also a genuine, deliberate element of randomness in how a model selects the next token at each step, which is part of why even the same tool can give you a slightly different answer twice to the identical question.
What This Means for How You Should Use These Tools
Treat AI output as a draft or a starting point from a knowledgeable but sometimes-wrong collaborator — not as a finished, verified answer
Use AI tools for what they're genuinely good at: brainstorming, organizing, summarizing, restructuring, explaining concepts, and generating first-pass drafts of things you'll personally revise — not as an unsupervised source of final facts or final prose
Understand that a tool's context window limitation is a real constraint on long projects, and look for tools (covered throughout this section) that are specifically built to manage that limitation for book-length work, rather than assuming any general-purpose tool can simply "hold" your whole manuscript
Remember that the creative judgment — what your story actually says, who your characters are, what your book's voice sounds like — is not something an LLM has any genuine grasp of beyond pattern-matching against similar text; that judgment remains yours, which is the core philosophy behind every article that follows in this section
Conclusion
Understanding what an LLM actually is — a large-scale prediction system trained on patterns in text, not a thinking or fact-retrieving entity — demystifies a lot of what otherwise feels like magic or mystery in today's AI tools. It also explains exactly why these tools are powerful for some tasks (organizing, drafting, summarizing, brainstorming) and genuinely risky for others (unverified facts, long-range consistency, anything requiring real judgment). Every article that follows in this section builds on these fundamentals, walking through specific platforms — general-purpose assistants, purpose-built fiction tools, editing software, and audio production tools — with this same grounded, practical understanding of what AI can and can't actually do for your author business.
- Randall