AI Terms HR Needs to Know: A Plain-English Glossary for Small Employers

You do not need to memorize hundreds of AI words. Learn twenty-five, sorted by the decision each one affects: what the tool is, how you instruct it, where its answers come from, what can go wrong and what rules apply to people decisions. For each term, know the one question it should make you ask.

I am Mac Malbrough. I hold a J.D. and an MS-HRM, I am not an attorney, and this is general information, not legal advice. I wrote this glossary for HR leaders and owners at small employers who keep hearing vocabulary in vendor demos and team meetings and nod along. A term you cannot explain is a risk you cannot manage, so each entry below ends with the question that turns the word into a decision.

The problem this solves: Vendors and coworkers keep using AI words, your staff are already pasting work into free tools, and nothing in your handbook says what is allowed.

How UHR solves it: The California employee handbook template ($149) gives you the editable policy home for an AI rule, and the AI Use and Disclosure Policy Kit ($29) gives you a short use and disclosure standard to adapt.

Which terms describe what the tool is?

These words describe the kind of system in front of you. They matter because the legal and practical risks change with the type of tool.

• Artificial intelligence (AI). Software that performs tasks that normally need human judgment, such as sorting, predicting or writing. Ask: what exactly is this tool doing on our behalf?
• Machine learning. A type of AI that learns patterns from examples instead of following rules a person wrote. Ask: what examples did it learn from?
• Generative AI. AI that produces new text, images or audio from a prompt. Ask: who reads the output before it leaves HR?
• Large language model (LLM). The kind of generative AI behind most writing assistants. It predicts likely words, which is why it can sound confident and be wrong. Ask: how do we check a confident answer?
• AI agent. A tool that pursues a goal by taking several steps, such as reading an inbox, drafting replies and scheduling meetings. Ask: what is it allowed to do without asking us?

Which terms describe how you instruct it?

These words describe your side of the conversation. Good instructions are the cheapest safety control you have.

• Prompt. The instruction and context you give the tool. Ask: does our prompt say what the tool must not do?
• Context window. The amount of text the tool can consider at once. Long conversations can push earlier instructions out. Ask: did the tool forget a rule we gave it earlier?
• Prompt template. A reusable prompt with blanks to fill in. Ask: who owns this template and when was it last reviewed?
• Iterative prompting. Improving a result by revising your request after seeing the output. Ask: are we keeping the final version or only the first draft?
• Skill or custom assistant. A saved set of instructions for one task, so each person gets consistent results. Ask: what does it do when it does not know?

Which terms describe where answers come from?

This group decides whether an HR answer is trustworthy. A policy answer that is not drawn from your policy is a guess.

• Training data. The material a model learned from. Ask: does the vendor tell us what it was trained on, and was any of it our employees’ data?
• Grounding. Tying a tool’s answer to a source you supplied, such as your handbook. Ask: can the tool show which paragraph produced the answer?
• Retrieval-augmented generation (RAG). A setup where the tool looks up your documents first and then writes the answer from them. Ask: how current is the document library?
• Hallucination. A confident answer that is made up. In HR, a hallucinated leave rule or a made-up statute is the common example. Ask: who verifies legal and policy statements?
• Fine-tuning. Further training of a model on selected examples. Ask: are employee records being used to train anything outside our control?

Which terms describe what can go wrong?

These are the words behind most HR risk. They are worth learning first.

• Bias. A tool that produces systematically different results for different groups. Ask: has anyone tested the results by group?
• Algorithmic bias. Unfair results that come from the way a system is built or the data it uses. Ask: what proxies for protected traits could be hiding in the inputs?
• Automation bias. The habit of trusting a computer’s recommendation too quickly. Ask: does the reviewer have the time and authority to disagree?
• Black-box tool. A system that gives results without a clear reason. Ask: can we explain the result to an affected employee?
• Shadow AI. Employees using unapproved AI tools for work. Ask: what are people already pasting into free tools?
• Personally identifiable information (PII). Details that identify a person, such as a name with a Social Security number or an address. Ask: is any of it going into a tool we have not approved?

Which terms describe the rules and controls?

The last group connects the technology to your responsibilities as an employer in California.

• Human in the loop. A person reviews and decides before the output is used. Ask: does the person have the authority to change the result, or only to approve it?
• Guardrails. Limits on what a tool can access or produce. Ask: where are the limits written down?
• Automated-decision system. California’s civil rights regulations use this term for a computational process that makes a decision or helps a person make a decision about an employment benefit. A resume screener or a video interview scorer can qualify. Ask: are we using anything that sorts, ranks or scores people?
• AI governance. The policies, owners and review habits that keep AI use legal and fair. Ask: who is accountable when the tool is wrong?

For how those rules work in practice, read the California civil rights rules for AI hiring tools and the piece on SB 947 and automated decisions in discipline and termination.

Which three terms cause the most trouble for small employers?

If you only have time for three, start with hallucination, shadow AI and human in the loop. Hallucination explains why a confident answer about leave, pay or breaks cannot be trusted without a check against the law or your policy. Shadow AI explains where your real exposure sits today, because employees are often using free tools before anyone approves them, and what they paste in may include names, medical details or complaint facts. Human in the loop explains what makes a tool defensible, but only when the person has time, training and authority to change the result. A reviewer who clicks approve on forty outputs in ten minutes is not a control. Write those three into your policy first, and the rest of the vocabulary becomes much easier to place.

This list covers the 25 terms I would learn first. The Urban Glossary section on AI for HR teams defines 47 more, including refusal rule, data minimization, prompt injection and model drift.

How do you use this glossary with your team?

A glossary is only useful if people share it. Try three simple moves.

1. Pick the five terms your team uses most and write one-sentence definitions in your own words.
2. Add the matching question to your vendor checklist, so every demo ends with the same five questions.
3. Choose one term per month for a ten-minute team discussion, starting with shadow AI.

Notice that none of these steps needs a technical background. You need a shared vocabulary and the habit of asking where an answer came from and who checked it.

How does the UHR handbook template turn these terms into policy?

The problem for most small employers is that the vocabulary arrives before the rules do. Staff are already using tools, nobody wrote down what may be pasted into them, and a vendor demo ends with a signature and no policy. The California employee handbook template ($149) gives you the annotated, editable policy structure where your AI rules belong, alongside the confidentiality and employee relations clauses those rules must fit. Write your AI use rules into that structure, using the terms above, so the words your team uses and the words your policy uses are the same.

For the AI rules themselves, the AI Use and Disclosure Policy Kit for Solo Consultants is $29 and shows how a short use and disclosure standard is built, and the free AI Prompt Pack Sampler: 12 Free Claude Prompts lets you practice the prompt vocabulary above on safe tasks before you write your policy.

What should you do this week?

1. Ask each manager which AI tools they already use for work and what they paste into them.
2. Pick the five terms from this glossary that matter most to your team and write your own definitions.
3. List every tool that sorts, scores or ranks applicants or employees, and note who reviews its output.
4. Draft a one-paragraph AI use rule and put it in your handbook notes.

If you would like an outside review of where AI already touches your people decisions, the HR compliance check-up is the place to start, and fractional HR support can help you write and maintain the policy.

Related reading: The HR AI Field Guide, my book of plain-English terms, prompts and safe rules for AI in HR.

For a related angle, see people analytics terms in plain English.

Sources

• California Civil Rights Department, automated decision system regulations
• California SB 947 (2026), read on the California Legislative Information site

McKinley holds a J.D. but is not a licensed attorney. Articles here are general information, not legal advice. For your specific situation, talk with an employment attorney.

Data current as of November 2026. Sources are linked where each figure appears.