What Is a Large Language Model, and What Does It Get Wrong?

AI Recruiting Academy › Module 1: AI basics for recruiters

How does a large language model work?

Think of the autocomplete on your phone, then scale it up enormously. During training, the model reads a vast collection of text and learns, word by word, which words tend to follow which. It does this billions of times until it has a detailed statistical sense of how language fits together.

When you type a request, the model writes its answer one small chunk of text at a time, each time picking a likely next piece based on everything before it. Those chunks are called tokens. A token is roughly three quarters of an English word.

The important point for recruiters is what the model is not doing. It is not searching a database of facts about your candidate. It is producing text that sounds right given the words in front of it. When the words in front of it are good, such as a real resume you pasted in, the output is usually good. When it has to fill a gap from memory, it may invent something plausible.

Which language models are behind recruiting tools?

A handful of companies build the big general models. Most recruiting products do not build their own. They rent one of these models through an API, then add recruiting data, screens and workflows on top.

Two tools can feel very different even when they use the same model underneath, because the instructions, data and guardrails the vendor adds matter as much as the model. When you evaluate a product, ask which model it uses and what it adds on top.

What is a context window, and why does it matter?

A model can only pay attention to a limited amount of text at once. That limit is called the context window. Everything in the conversation, your instructions, the documents you pasted and the model’s own replies, has to fit inside it.

Modern windows are large, but recruiters still hit the edge. Paste 200 resumes into one chat and ask for the top ten, and the model may skim, skip candidates in the middle or quietly work from only part of the list. Long, messy chats also drift, with early instructions carrying less weight later on.

Two habits help. Give the model one job at a time, such as one resume against one scorecard, and start a fresh chat when you change tasks.

What is a hallucination?

A hallucination is when a model states something false as if it were true. It is not lying in the human sense. It is predicting text that fits the pattern, and a confident wrong answer can fit the pattern perfectly.

OpenAI’s research on why models hallucinate argues that standard training and testing reward guessing over admitting uncertainty, much like a multiple choice exam rewards a guess over a blank answer.

In recruiting, hallucinations tend to look like this:

The fix is not to avoid AI. It is to give the model the source material and ask it to use only that, then check any fact that will reach a candidate, client or hiring manager.

Why do language models carry bias?

Models learn from text written by people, and that text reflects the biases of the world it came from. A model can pick up patterns such as which names tend to appear next to which jobs.

University of Washington researchers put this to the test. They swapped 120 first names across more than 550 real resumes and had three language models rank them against over 500 real job listings, more than three million comparisons in all. The models favored white associated names 85% of the time and female associated names only 11% of the time, and never favored Black male associated names over white male associated names.

That is why “the AI decided” is never a defense. If a tool ranks or rejects candidates, someone needs to know how it was tested for bias. Module 4 covers bias audits in detail.

What happens to the data you paste in?

It depends on the product and the plan. Free consumer chat apps may use conversations to improve their models unless you change a setting. Business plans usually promise the opposite. OpenAI, for example, states on its enterprise privacy page that it does not train on business customers’ data by default.

Resumes and interview notes are personal data. Before pasting them into any AI tool, check which plan you are on and what your employer’s policy allows. When in doubt, remove names and contact details first.

How do you check what an LLM gives you?

A simple rule works: the closer the output gets to a real person, the harder you check it.

Quick answers

Does AI know what happened this week?

Not on its own. A model only knows what was in its training data up to a cutoff date. Some assistants can search the web, and then they are only as current as what they find.

Is one model better for recruiting than the others?

The leading models trade places often. For most recruiting tasks, how you set up the task matters more than which model you pick. Lesson 4 covers that.

Can AI tell if a resume was written by AI?

Not reliably. OpenAI withdrew its own AI text classifier in 2023 because of its low accuracy. Judge the candidate on evidence, not on a detector score.

Is an LLM the same as an AI agent?

No. The LLM is the engine. An agent is a system built around it that can take actions. Lesson 3 explains the difference.