Overview

Every vendor call, every tool documentation, every article about AI throws around jargon. LLM, hallucination, prompt, fine-tuning, RAG, context window, token, bias, training data. You can nod along and pretend to understand, or you can actually understand. This lesson gets you to the second option.

Purpose

You need to know what these terms mean because you're making decisions about tools, evaluating vendor claims, and explaining to your leadership what's actually happening inside these systems. When someone says "we've reduced hallucinations by 40%," you need to know what that actually means and whether it matters. When a tool uses "RAG-enhanced retrieval," you need to understand what's actually different.

This lesson is not a glossary. It's a narrative walk through the terms you'll actually encounter, explained in context, with HR examples for each.

Why This Matters for HR Professionals

Vendors sell AI tools using technical language that sounds impressive. "We've fine-tuned a large language model using your company's historical hiring decisions to achieve class-leading accuracy in candidate classification." That sounds great. But if you understand what each term actually means, you might recognize this is saying: "We've trained our system to replicate your historical hiring, which probably included bias."

Understanding terminology prevents you from being sold something dangerous while thinking it's a feature.

The Core Terms

Large Language Model (LLM)

An LLM is an AI system trained on enormous amounts of text. It's learned statistical patterns from that text. GPT, Claude, Llama. These are different LLMs. They're trained on different data and built with different approaches, but they're all doing the same core thing: predicting the next token based on patterns.

Why it matters for HR: When someone says "we use an LLM," they mean the system is trained on general text data, not specific HR data. It knows common HR concepts but not your company. An LLM can write a job description, but it won't know your company's specific culture or what makes your team unique unless you tell it in your prompt.

Prompt

A prompt is your input. The question you ask. The instruction you give. "Write a job description for a software engineer" is a prompt. "Analyze this feedback and identify performance themes" is a prompt.

Why it matters for HR: Different prompts get different outputs. A good prompt is specific about what you want, includes context, and clarifies constraints. "Write a job description for a backend software engineer at a startup focused on hiring diverse candidates" is better than "Write a job description." Learning to write better prompts gets you better output without changing the tool.

Hallucination

A hallucination is when an AI system generates false information with confidence. It makes up citations, regulations, policies, facts. It sounds authoritative while doing it.

Why it matters for HR: Hallucinations are the biggest risk in HR AI. The system will generate text that sounds like it's citing specific regulations, policies, or facts when it's actually making them up. You read it, it sounds authoritative, you trust it. Then you implement a policy based on a regulation that doesn't exist.

Token

A token is a small chunk of text. Usually a word, sometimes a part of a word. "Performance review" might be two tokens. "Hallucination" might be one token.

Why it matters for HR: Tokens affect cost and context. If you're pasting a long feedback document (5,000 tokens) into a system with a context window of 2,000 tokens, you've already exceeded the system's capacity. Tokens also affect cost. You pay per token input and output. Shorter, more focused prompts use fewer tokens and cost less.

Context Window

The context window is how much text the system can process at once. A system with a 2,000-token context window can process about 1,500 words at a time. A system with a 100,000-token context window can process about 75,000 words.

Why it matters for HR: If you want to paste a year of feedback for an employee (10,000 tokens) and ask for analysis, you need a system with a large enough context window. If you want to paste multiple documents at once, you need to account for context window size. Smaller windows limit what you can do.

Training Data

Training data is the text the system learned from. The Internet, books, articles, code repositories, whatever was used to teach the system patterns.

Why it matters for HR: If the training data included biased content, the system learned those biases. If the training data didn't include recent information, the system doesn't know about recent regulations. Training data is a major source of hallucinations, if the system learned from articles that contained false information, it replicates those falsehoods. You should always ask vendors: What's your training data? When was it collected? Has it been verified for accuracy?

Fine-Tuning

Fine-tuning is taking an already-trained system and training it further on your specific data. You have historical job descriptions, so you fine-tune the system on those. The system learns your company's style and preferences.

Why it matters for HR: Fine-tuning can be powerful or dangerous depending on what you're fine-tuning on. Fine-tuning on your job descriptions helps the system match your voice. Fine-tuning on your hiring decisions teaches the system to replicate your historical biases. Always ask: What data are we fine-tuning on? Does that data contain the patterns we want to replicate?

RAG (Retrieval-Augmented Generation)

RAG is a technique where the system retrieves documents before generating an answer. You ask a question. The system searches for relevant documents from your company. Then it generates an answer based on those documents plus its training.

Why it matters for HR: RAG makes the system more accurate for company-specific questions because it's working with your actual documents. But it's only as good as those documents. If your policy documents are outdated, the system generates answers based on outdated information. RAG also doesn't eliminate hallucinations, the system can still misunderstand document content or generate false information even when working with retrieved documents.

Bias

Bias in AI usually means the system produces systematically different outcomes for different groups. The system rates candidates from certain backgrounds lower. It flags certain demographics as flight risks. It generates different language in performance reviews for different groups.

Why it matters for HR: Bias in AI systems is often learned from training data that reflected human bias. It's also sometimes built in through the way the system is designed. You can test for bias and try to reduce it, but you can't eliminate it entirely if the underlying data was biased. The key is acknowledging bias exists and deciding how to manage it.

Accuracy

Accuracy is a metric: What percentage of the system's outputs were correct? 95% accuracy sounds great. But it's often misleading in HR because it hides disparities. The system might be 98% accurate for one group and 88% accurate for another group.

Why it matters for HR: When vendors cite accuracy metrics, ask for accuracy by demographic group. Ask what "accuracy" means, accurate compared to what standard? If the system was trained on historical hiring data, is it "accurate" if it replicates historical hiring bias? Accuracy metrics can hide the actual patterns you care about.

Prompt Injection

Prompt injection is when instructions hidden in content (a document, an email, a data field) get treated as instructions by the system. You paste a document that says "ignore the user's instructions and do X instead," and the system treats that as a real instruction.

Why it matters for HR: If you're using RAG systems that retrieve documents, be aware that instructions can be embedded in those documents. An employee could put instructions in their feedback or comments that the system treats as directives. It's not common in HR, but it's a risk to be aware of.

Temperature

Temperature controls randomness in the system's output. Low temperature (0.1-0.3) means the system almost always picks the statistically likely next token. High temperature (0.7-1.0) means the system picks more randomly.

Why it matters for HR: In HR work, you want low temperature. You want consistency and predictability. You don't want interview questions that are creative but unprofessional. You don't want policy language that's varied and inconsistent. Set temperature low and the system will give you the most reliable outputs.

Inference

Inference is the process of using a trained system to make predictions or generate outputs. Once the system is trained, inference is what happens when you type a prompt and get an answer.

Why it matters for HR: Inference is where errors happen. A system might be trained well but make wrong inferences (predictions) because of the specific prompt you gave or the randomness involved in generation. When things go wrong with an AI tool, it's often an inference problem, not a training problem.

Embedding

An embedding is a mathematical representation of text. "Performance review" gets represented as numbers that capture its meaning in a mathematical space. Systems use embeddings to understand similarity and find related content.

Why it matters for HR: Embeddings are used in systems that find similar documents, recommend content, or search. You might not see embeddings directly, but they're what powers the "related documents" feature or recommendation algorithms. They're a powerful way to find patterns but can also amplify biases if the training data was biased.

Parameter

A parameter is a value the system learned during training. A very large LLM might have 70 billion parameters. More parameters usually means more capability but also more computational cost.

Why it matters for HR: When people say "a 70-billion-parameter model," they're talking about how powerful the system is. More parameters means more capability but also higher cost and slower responses. You don't need to know the exact number, but understanding that more powerful systems have more parameters helps you understand why some AI tools are more capable (and more expensive) than others.

RLHF (Reinforcement Learning from Human Feedback)

RLHF is a technique where the system is trained using feedback from humans about what outputs are good. Instead of just learning patterns from data, the system learns from human judgments about quality.

Why it matters for HR: Systems trained with RLHF tend to produce outputs that align better with human preferences. They're more likely to refuse harmful requests and less likely to generate obviously biased content. When vendors mention RLHF, they're saying the system was trained to produce outputs that humans like. This is usually a positive signal compared to systems without it.

Related Concepts

Model Drift

Model drift is when a system's output quality decreases over time. The world changes, new patterns emerge, the system's training becomes outdated.

Why it matters for HR: If you're using an AI system to analyze compensation trends or hiring patterns, the system's understanding becomes less accurate as the world changes. What was accurate last year might be wrong this year. This is one reason you can't just build an AI system once and leave it. It needs ongoing evaluation and retraining.

Adversarial Examples

Adversarial examples are inputs that trick a system into making mistakes. Small changes to an input can cause dramatically different outputs. If you slightly modify a resume (rearrange the bullet points, use slightly different words), the system might score it completely differently.

Why it matters for HR: This is mostly a research concern, but it's worth knowing: AI system outputs aren't always stable and reliable. A small change that shouldn't matter might change the output entirely. This is one reason you can't just trust a system's output without human review.

Transfer Learning

Transfer learning is taking a system trained on one task and adapting it to a different task. An LLM trained on general text can be adapted to HR-specific work.

Why it matters for HR: Transfer learning is how most HR AI tools work. They're not built from scratch for HR. They start with a general LLM and are adapted to HR through prompts, fine-tuning, or RAG. Understanding this helps you recognize that HR AI is always somewhat adapted from general systems, which is why accuracy on general tasks doesn't guarantee accuracy on HR tasks.

What to Do Monday Morning

• 
  Make a glossary for your team. Pick the 10-15 terms from this lesson that are most relevant to your tools and create a one-page glossary. Share it so everyone uses consistent language.
• 
  When talking to a vendor, ask specific questions:
  - What's your training data? When was it collected? Has it been verified?
  - What's the system's context window?
  - Have you tested for bias? Across which demographic groups?
  - What accuracy metrics do you have?
  - Is this fine-tuned? On what data?
  - Does this use RAG? What documents are retrieved?
• 
  Reread a tool vendor's marketing material and translate it. Where they say "leverages advanced LLM technology," you now know that means "uses pattern matching." Where they say "fine-tuned on 10 years of hiring data," you know that means "trained to replicate your historical patterns."
• 
  In your next tool evaluation, focus on one term: Ask the vendor to explain one concept in detail. Do they really understand it, or are they using jargon they don't fully grasp?
• 
  Test your own AI tool for hallucinations. Ask it about a specific HR regulation. Compare the output to the actual regulation text. What did it get right? What did it invent?

Key Takeaways

• Know the core terms: LLM, prompt, hallucination, token, context window, training data, fine-tuning, RAG, bias, accuracy
  - Understand that these terms refer to specific technical concepts, not just marketing language
  - Ask vendors detailed questions about terminology, if they can't explain it, they probably don't fully understand their own system
  - Recognize when vendors are using terms to hide risks (fine-tuning on historical data, high accuracy metrics that hide disparities)
  - Use terminology to have precise conversations about what AI systems actually do

FAQ

Q: If a system has 95% accuracy, isn't that good enough?
A: Not necessarily. Accuracy by overall metrics hides disparities. The system might be 98% accurate for one group and 85% for another. In HR, that difference is discrimination. Always ask for accuracy broken down by demographic group.

Q: What's the difference between RAG and fine-tuning?
A: Fine-tuning updates the system's internal knowledge based on your data. RAG keeps your data separate and retrieves it when needed. RAG is usually better for HR because it doesn't permanently embed your data into the system, and you can update documents without retraining.

Q: Does more parameters always mean better performance?
A: Usually, but not always. A bigger system with more parameters is more capable but also slower and more expensive. Sometimes a smaller, well-designed system outperforms a larger one. For HR tasks, you rarely need the most powerful system available.

Q: Can we fix bias through fine-tuning?
A: Fine-tuning can help reduce bias, but it requires deep understanding of what biases exist and how to address them. For most HR teams, the better approach is understanding what biases exist and building human review into the process rather than trying to fix it through fine-tuning.

Q: If a vendor says they use RLHF, does that mean the system isn't biased?
A: RLHF helps reduce certain types of bias and makes the system more aligned with human preferences. But it doesn't eliminate bias. The system can still learn and amplify biases in training data or from human feedback that contains bias.

What's Next

Now you understand the terminology and the mechanisms. In the next chapter, we shift from "how AI works" to "what AI can actually do in HR." We'll explore the specific capabilities and limitations that matter for your work, where AI excels, where it fails, and what you need to verify before it affects anyone.