Free Resources

A large language model (LLM) is a statistical system trained to predict what text comes next given what came before. It has processed enormous quantities of human-written text — books, articles, code, websites — and learned patterns about how words, sentences, and ideas relate to each other.

It is not a database. It is not a search engine. It does not look things up when you ask a question. It generates a response based on learned patterns about what a plausible answer to your question looks like. This distinction matters enormously in practice.

The "large" in LLM refers to the number of parameters — internal numerical weights — that the model contains. Modern frontier models have hundreds of billions of parameters. More parameters generally (though not always) means more capability and more nuanced understanding of context.

LLMs generate text one token at a time. A token is roughly a word or word-fragment. At each step, the model calculates a probability distribution over its entire vocabulary — every possible next token — and selects one, then repeats the process. This is called autoregressive generation.

The selection isn't always the single highest-probability token. A parameter called temperature controls randomness: low temperature makes the model more deterministic and predictable; high temperature introduces more variety and creativity (and sometimes more errors).

The model processes your entire conversation — every message — as one combined input called the context window. This is why LLMs can reference earlier parts of a conversation. Context windows have a maximum length. Once exceeded, earlier content is no longer "visible" to the model.

Hallucination is when an LLM states something false with apparent confidence. It might invent a citation, misquote a statistic, or describe a law that does not exist. This is not a bug that will be fully eliminated — it is a structural consequence of how these systems work.

Because an LLM generates text based on learned patterns rather than verified facts, it has no internal mechanism for distinguishing "I know this is true" from "this sounds like something that would be true here." The model produces fluent, plausible text — fluency and accuracy are not the same thing.

Hallucination risk increases when you ask about: very specific facts (exact dates, names, numbers), recent events past the model's training cutoff, niche or specialized topics, and citations or sources. It decreases when you provide source material and ask the model to reason from it rather than from memory.

How you phrase your request significantly affects what you get. This is called prompt engineering — a somewhat inflated term for what is largely careful writing. Key principles:

LLMs have a training cutoff — a date past which their knowledge was not updated. Events, laws, and data after that date are unknown to the model unless you tell it. Ask a model what its knowledge cutoff is before relying on it for recent information.

LLMs cannot take actions in the real world by themselves (unless connected to external tools, which you will typically know about). They cannot check your email, access your files, browse the web (unless explicitly enabled), or make purchases. They respond to text and produce text.

LLMs reflect patterns in their training data, which means they can reproduce biases, outdated perspectives, and culturally specific assumptions present in that data. This is worth being aware of, particularly when using them for decisions affecting people.

OCF's research focuses specifically on how AI displacement is affecting workers — particularly in rural and economically vulnerable communities. This is not abstract. LLMs are already being deployed in roles that involve writing, customer service, data entry, paralegal work, and administrative tasks.

Understanding how these tools work is, increasingly, a form of workforce readiness. Not because everyone needs to become a developer, but because workers and institutions who understand what these systems can and cannot do are better positioned to advocate for themselves, adapt, and contribute to decisions about how AI is used.

OCF's Workforce Readiness research series is available free on our Research page. The papers examine these dynamics through an access-centered, evidence-based lens.