Marcus runs talent acquisition for a 600-person regional hospital network, leading a team of four recruiters who collectively screen about 240 applicants a week across nursing, allied health, and administrative roles. Last quarter a vendor pitched him a "fully AI-powered hiring platform" with a price tag that would have consumed nearly a fifth of his tooling budget. The demo was slick. But when Marcus asked what the system was actually trained on, the sales engineer paused, said "it uses advanced AI," and changed the subject. Marcus did not buy. Not because he distrusts AI, but because he had learned to hear "AI" not as an answer but as a category, and to ask which technology inside that category the tool actually used. That single habit, knowing what the words mean, is what separates a recruiter who gets sold from one who buys well.

Why the Vocabulary Is Worth Your Time

When Marcus first heard "machine learning" and "large language model," they sounded like something from a research lab, not his inbox. But these technologies were already woven through his workflow. The system that surfaced "people you may want to source" inside LinkedIn Recruiter, the resume parser in his applicant tracking system, the tool that drafted a first pass at a rejection email: all of them were running on the technologies in this lesson. Treating them as interchangeable black boxes is what leads to a six-figure purchase nobody can explain six months later. The point of learning the distinctions is not to pass a quiz. It is so that when a vendor says "AI," Marcus knows exactly which follow-up question exposes whether the product fits his hospital network or just sounds impressive.

There is a useful parallel in the work itself. When Marcus screens a resume, he scans for credentials, looks for progression, and leans on years of pattern recognition about what a strong candidate for his market looks like. That is intelligent decision making applied to information. The technologies below do versions of that same work, at scale and at machine speed. Understanding the mechanics matters far more than memorizing the terminology, and the mechanics are not complicated once you stop letting the jargon intimidate you.

Artificial Intelligence: The Umbrella Term

Artificial intelligence is the broadest of the four words, and the least precise. It means, simply, computer systems designed to perform tasks that normally require human intelligence: recognizing patterns, making decisions, learning from experience, improving over time. It does not mean conscious machines or sentient algorithms. In a recruiting context, a system that ranks candidate profiles by predicted fit, or one that flags which offers are likely to be accepted, is an AI system. It is software built to do something a person would otherwise do with judgment, just faster and more consistently.

Here is the trap, and it is the exact trap Marcus avoided in that demo. Because AI is a category and not a specific technology, "we use AI" tells you almost nothing about what a product can reliably do. It is like a candidate writing "experienced professional" on a resume with no detail underneath. True, but empty. When a vendor leads with "AI" and stops there, that is your cue to ask which technology inside the umbrella is actually doing the work, because the answer changes everything about what you should expect.

What AI Is Not

Before going deeper it is worth clearing away the myths that vendors rarely volunteer to correct, because each one of them, left standing, leads to a specific kind of mistake in hiring.

• AI is not conscious. It does not think, decide, or intend. It processes information.
• AI is not all-knowing. It knows only what it was trained on. Ask a system whose training data ended in 2020 about events in 2025 and it will have nothing to draw on.
• AI is not objective. It is trained on human data, and human data carries human bias. The system inherits that bias.
• AI is not magical. It is mathematics and statistics. Powerful mathematics, but still math.
• AI is not a person. You cannot hold it accountable. You remain accountable for how you use it.

The distinction underneath all five is the one Marcus repeats to his team whenever someone starts talking about what the tool "thinks." When a person is intelligent, they understand context, meaning, intention, and consequence. When a system is called intelligent, it is recognizing and replicating patterns in data. Those are not the same thing, and a recruiter who forgets the difference will eventually defend a decision by saying the software recommended it, which is not a defense at all.

Machine Learning: Systems That Learn From Data

Machine learning is a specific way to build AI. Instead of a programmer writing fixed rules ("if the resume contains the word Python, add five points"), a machine learning system is shown many examples and discovers the patterns itself. You give it historical data, it analyzes the relationships in that data, and it develops its own internal decision rules. Nobody hand-writes those rules. The system infers them.

Picture Marcus's nursing pipeline. Over several years his network has hired roughly 500 nurses, and the system has the outcomes attached: who passed probation, who was still on staff after two years, who left within six months. A machine learning model can study that history and surface correlations a human would never have time to tabulate. Perhaps candidates who came through a particular clinical rotation program stayed longer. Perhaps a certain sequence of prior roles predicted strong performance. The recruiter did not tell the system to look for those things. It found them by analyzing the examples. That is the defining trait of machine learning: the patterns come from the data, not from a person.

The strength here is scale and consistency. Marcus has good instincts about his market, but those instincts were built from the few hundred candidates he personally interviewed. A machine learning model can find patterns across thousands of decisions. The catch, and it is the one that matters most for evaluation, is that a machine learning system is only as good as the data it learned from. If the historical hiring data carries bias, the model can learn and repeat that bias at scale. That is precisely why "what was it trained on?" is not an idle technical question. It is a fairness question.

Rules You Write Versus Patterns the System Finds

A great many products sold as "AI" in recruiting are actually rule-based automation, and telling the two apart changes how you should evaluate them. Rule-based automation is what happens when you, or a developer working for you, write the logic explicitly. A rule might read: if a candidate has five or more years of Python experience, and currently works at a Fortune 500 company, and attended a top-20 university, move them to the fast track. You define the rule, and the system follows it exactly, every time, with no flexibility at all.

That rigidity is both the appeal and the problem. Rules are predictable, because you set them and you know the outcome. They are transparent, because you can see exactly why a candidate was included or excluded. They are controllable, because you can adjust or switch one off in a minute. What they cannot do is adapt to context or nuance, and they require you to already know which criteria matter. Worst of all, they miss signals that do not fit the rule, which is how a brilliant candidate who went to a state school disappears from a pipeline without anyone noticing.

Machine learning trades those properties away for different ones. Feed the system a thousand resumes of people you hired and a thousand of people you passed on, and it may surface something like "candidates who stayed at previous employers for three to five years tend to accept offers more often." It did not memorize a rule; it found a pattern. Its strengths follow from that: it can find patterns humans miss, it adapts as the data changes, it handles nuance rather than binary thresholds, and it is often more accurate than a hand-written rule set. Its limitations follow just as directly. It can be a black box, so you may not be able to say why it produced a given output. It is only as good as the data it learned from, so a history of biased hiring becomes a model that reproduces bias. It needs a large volume of historical data to work at all. And it can latch onto spurious patterns, correlations that hold in your data but are not causal and will not hold for the next candidate.

So the single most useful question Marcus asks a vendor is deceptively simple: is this automation, meaning rules I can read, or is this machine learning, meaning patterns learned from data? The answer tells you whether to ask for the rule list or the training set, and those are completely different diligence conversations.

The Three Phases of a Machine Learning System

Every machine learning tool you will ever be sold went through the same three phases, and knowing them lets you ask where in that sequence the vendor's claims come from. In training, the system is given a large dataset with known outcomes: five thousand resumes, say, each labeled hired or not hired, along with what happened afterward, whether the person accepted the offer and whether they succeeded in the role. The system analyzes which characteristics correlate with success and which with rejection, and builds a mathematical model, not a list of rules but a statistical representation of those patterns. In validation, the model is tested on data it has never seen. If it learned real patterns rather than memorizing the training examples, it should perform well on the new data. In deployment, the trained model is applied to live candidates, producing outputs like "this candidate has a 78 percent likelihood of accepting an offer, based on patterns from your past hires."

Probability Is Not Certainty

That 78 percent deserves a hard look, because it is where well-meaning organizations do the most damage. It is a probability drawn from patterns across historical data. For any individual candidate, the actual outcome is either zero or one hundred: they accept or they do not. The number describes a population, not a person.

The misuse is predictable. A team treats a 78 percent prediction as a high-confidence decision, and starts auto-rejecting everyone the model scores low. But the model is working with radically incomplete information about individuals. A candidate rated 45 percent likely to accept might have circumstances that make them nearly certain: a recent relocation to the area, a long and engaged recruiting process, a role that fits something specific they have been looking for. None of that is in the training data. Machine learning models are genuinely useful for thinking about populations and trends. They are dangerous when used to make binary accept or reject decisions about individuals without human judgment in the loop, and that danger is a fairness problem as much as an accuracy one, because the candidates most likely to be misread by a population model are the ones whose paths look least like the historical majority.

Natural Language Processing: Machines That Read Language

Natural language processing, or NLP, is machine learning aimed specifically at human language. It is how software makes sense of the words in a resume, an email, a job description, or an interview transcript. Recruiting runs almost entirely on unstructured text, so without NLP, recruiting AI would have nothing to work with.

When the parser in Marcus's applicant tracking system ingests a resume, NLP is doing several jobs at once. It extracts structured facts: job titles, employers, dates, certifications. It interprets context, working out whether "led a team" meant directly managing people or coordinating across departments. It reconciles vocabulary, recognizing that "RN," "registered nurse," and "R.N." all point to the same credential, and that "BLS" and "Basic Life Support" are the same certification. It picks up on nuance, distinguishing "basic proficiency in medication administration systems" from "expert." None of that is generation; it is comprehension. NLP is the reading layer of recruiting AI, the step that turns a messy document into something the rest of the system can act on.

The Five Jobs NLP Does in Recruiting

If you want a checklist for spotting NLP at work in your own stack, look for these five tasks. It extracts information from text, reading a resume and identifying name, contact details, skills, experience, and education. It classifies text, reading a job description and categorizing it as technical, sales, operations, or clinical. It matches text, comparing the skills stated on a resume against the skills required in a posting. It generates text, producing job descriptions, emails, or interview questions, which is the job the large language models in the next section specialize in. And it summarizes, reading a long resume or transcript and producing something short enough to act on. Almost every "AI feature" in a recruiting platform is one of those five wearing a product name.

How NLP Actually Works, Simplified

Traditional NLP relies on rules and dictionaries: if the word "Python" appears on a resume, tag it as a programming language. Modern NLP uses machine learning instead. The system learns from analyzing millions of texts that words appearing near "Python," such as "code," "library," and "project," indicate a technical context, and it uses that surrounding evidence to decide what a word means. That is far more flexible than a dictionary, but it is still pattern matching rather than understanding. Give it a resume line reading "I am fluent in Python" and the system may get confused about whether the candidate means the programming language or something else entirely, because fluency is a word that usually travels with spoken languages. Errors of that shape, plausible and quiet, are the ones that survive a quick skim.

Large Language Models: Generating and Summarizing Text

A large language model, or LLM, is a particular kind of machine learning system trained on an enormous volume of text: web pages, books, articles, and other written material. The name describes it plainly. "Large" because it has billions of internal parameters. "Language" because it works in text. "Model" because, at bottom, it is a very sophisticated pattern matcher predicting what words should come next.

What makes LLMs different from the nursing-outcomes model above is where their knowledge comes from. That earlier model knew only what Marcus's hospital network taught it. An LLM arrives already trained on a vast body of general text, which is why it can draft a job description, summarize a 40-minute interview transcript into a tight set of notes, generate behavioral questions for a charge nurse role, or produce a first draft of a rejection email in a respectful tone, all without ever having seen Marcus's specific hiring data. The general tools your team likely already touch, including assistants built on models such as Anthropic's Claude or comparable systems, are LLMs in this sense. The strength is fluency and versatility across language tasks. The limit, and it is the mirror image of machine learning's strength, is that a general LLM knows nothing about your particular roles, your market, or what success looks like in your organization unless you feed it that context.

What the Model Is Actually Doing

The mechanism is worth holding onto because it explains every strength and every failure that follows. Imagine showing a system billions of sentences drawn from the internet, books, job postings, resumes, and emails, and telling it only this: learn which words and phrases commonly follow other words. It builds a statistical model of language patterns. When you then ask it a question, it produces an answer by predicting, one word at a time, which word should come next, making a probabilistic decision at every single step and doing so thousands of times across a single response. There is no lookup, no retrieval of a stored fact, no checking. There is prediction.

A Worked Example: The Outreach Email

Suppose you prompt an LLM with this: "Draft a compelling outreach email to a senior backend engineer with five years of experience. We are hiring for a Principal Engineer role at a Series B startup." The model does not understand that backend engineering is a distinct discipline, or why a Series B company might be appealing to someone at that stage, or what a Principal Engineer actually does day to day. What it has instead are learned patterns: which words tend to appear near "senior engineer," including experience, technical, leadership, and impact; the customary structure of professional outreach, moving from greeting to value proposition to specific opportunity to call to action; and the shape of the thousands of recruiting emails it saw during training. So it predicts an opening like "Hi [Name], I noticed your impressive work at [Company]," because that sequence commonly follows similar openings.

The output might read beautifully and be genuinely useful. But it arrived there through pattern matching, not through any grasp of your candidate or your opportunity, and that is exactly why the last mile of judgment stays with you.

Hallucination: Confident and Wrong

Because an LLM predicts from patterns rather than retrieving real information, it can generate false statements with total confidence. The industry calls this hallucination, and it is the failure mode most likely to reach a hiring manager unnoticed.

Picture using an AI tool to summarize a candidate's professional networking profile. The summary comes back reading: "Candidate has 12 years of marketing experience and currently leads the marketing team at a large technology company." You trust it and move the candidate forward. During the interview the candidate mentions they have been at that employer for two years and have about seven years of experience in total. The tool hallucinated. It did not lie in any intentional sense; it predicted plausible text and the prediction happened to be wrong. Everything downstream of that summary, including the comparison you drew against other candidates, was built on a fabricated foundation.

How the Four Terms Nest Together

The cleanest way to hold these four words in your head is as a set of nesting dolls. Artificial intelligence is the largest doll, the broad idea of systems doing intelligent things. Machine learning sits inside it, one specific approach to building AI by learning from data instead of following hand-written rules. Natural language processing sits inside machine learning, that same learning approach pointed at human language. And large language models sit inside NLP, a powerful and recent kind of language system trained on billions of words.

You will hear a fifth term thrown around in the same breath, and it fits in the same way. Deep learning is a subset of machine learning that uses neural networks, structures loosely inspired by how brains are wired, to learn patterns. LLMs are deep learning systems. For recruiting purposes you rarely need that level of detail; the distinction that actually changes your decisions is still machine learning versus rule-based automation.

They are not four competing products you choose between. They are layers of the same idea. So when Marcus uses almost any modern recruiting tool, he is using machine learning, frequently using NLP, and increasingly using an LLM somewhere in the stack. Walk through one of his real workflows and you can see each layer doing its job. He is filling a senior charge nurse role. His sourcing tool uses machine learning, tuned on the network's past hires, to rank incoming applicants by predicted fit. It uses NLP to actually read what each candidate wrote and to normalize the credentials and skills. And if an LLM is bolted on, it summarizes each shortlisted candidate's background into a consistent one-paragraph brief and drafts tailored phone-screen talking points. Different layer, different job, one continuous pipeline.

Run the numbers on that single pipeline and the value is concrete. Across a week, Marcus's team handles roughly 240 applicants. Manually reading and writing a structured summary for each one runs about eight minutes; the AI-assisted version, which the recruiter reviews and corrects rather than writes from scratch, takes closer to two. That is about 24 hours of recruiter time saved across the team in a week, time that moves from data entry to actual conversations with candidates. The figures are illustrative of his scenario, not a benchmark, but the shape is the real lesson: the savings show up only when each layer is doing the job it is actually suited for.

Tracing the Layers Through One Real Screen

Abstractions get slippery, so trace a single requisition end to end. Marcus's finance office opens a senior accountant role and 500 applications arrive. The platform he uses advertises "AI-powered resume screening." Here is what is actually happening underneath that phrase, in order.

1. NLP extracts the information. The system reads each resume and identifies skills, years of experience, job titles, and education. This is pattern matching against known categories, and it is the step that turns 500 documents into 500 comparable records.
2. Rules do the hard filtering. The system applies criteria Marcus set: must hold a CPA, minimum five years of accounting experience. These are hard filters, which is to say automation rather than machine learning, and their great virtue is that he can read them, explain them, and change them.
3. A machine learning model ranks what survives. For the candidates who pass the filters, a model trained on roughly a thousand of the network's past hires and non-hires scores them for fit. It learned something like "candidates with specific accounting software skills, plus large-firm experience, plus a CPA, tend to succeed here," and it applies that pattern to each new applicant.
4. Marcus reviews the top 50. The system surfaces the fifty highest-scoring candidates. He does not treat the ranking as a verdict. He reads them, weighs what the role actually needs this year rather than what it needed historically, and forms his own judgment.

Score that workflow honestly and the AI contributions look like this: NLP for information extraction, which is genuinely useful; rules for filtering, which are helpful and transparent; and machine learning for ranking, which is fine as a guide and dangerous as gospel. The human contributions are the rest: deciding which ranked candidates to interview, recognizing context the model could not see, and carrying accountability for the outcome. Notice that no layer of this stack is capable of holding that accountability, which is why it sits with Marcus by default whether or not anyone writes it down.

Drafting a Job Description: Where the LLM Helps and Where It Stops

Now take the opposite kind of task. Marcus needs a job description for that same senior accountant role, and he asks an LLM: "Write a job description for a senior accountant. We need someone with a CPA, five or more years of experience, strong spreadsheet skills, and the ability to mentor junior accountants." Back comes a clean draft built from the patterns of thousands of job descriptions the model saw in training, with the standard sections in the standard order: about the role, key responsibilities, required qualifications, preferred qualifications, why join us.

What is good about that is real. The blank page is gone, the structure is sound, and Marcus has something to react to rather than something to invent. What is missing is equally real. The model does not know his network's culture, the actual compensation band, the growth trajectory of the finance team, or the honest reason a strong accountant should want this job over the one down the road. It generated plausible text, not truthful specifics. So his role is the one that never transfers: take the draft, replace the generic with the actual, set the tone, and add the differentiation only someone inside the organization can supply. The AI generated options; the human applied judgment.

That division generalizes. The most effective use of AI in recruiting is not replacing your judgment. It is absorbing volume and routine work so that your attention lands on judgment, relationship building, and the context only a person can hold.

Why the Distinction Changes How You Buy

Knowing which layer a tool relies on changes the questions you ask and the claims you believe. Three habits follow directly from the nesting model.

First, never accept "AI" as a description of capability. It is a category, not a feature. When the vendor in Marcus's demo said "advanced AI" and moved on, the right response was the one he gave: which technology, trained on what data, built for what problem? A tool that screens with machine learning trained on 50,000 industry hires and a tool that "uses an LLM for screening" are doing genuinely different things, with different accuracy profiles and different failure modes. The marketing word is identical; the products are not.

Second, do not assume newer means better for your task. LLMs are the most sophisticated layer, so it is tempting to treat an LLM-based tool as automatically superior to a plain machine learning one. For many recruiting jobs the opposite holds. A focused machine learning model trained on your own hiring outcomes will often rank candidates for your specific roles more accurately than a general-purpose LLM that has never seen your data. The LLM is more versatile; that is not the same as more accurate for the narrow thing you need. Match the layer to the task: machine learning on your data for ranking and prediction, NLP for reading and extraction, an LLM for drafting and summarizing.

Third, remember that a definition is necessary but not sufficient. Knowing what machine learning is does not tell you whether a particular resume screener was trained on your network's hires or on generic public job postings, and those are wildly different products wearing the same label. The definitions give you the foundation. The diligence is in pressing each vendor on the specifics: the training data, the problem it was built to solve, and the constraints it carries. That is what turns vocabulary into buying power, and it is exactly the muscle Marcus used to keep a slick six-figure pitch from becoming a six-figure mistake.

Five Questions That Separate Capability From Marketing

Marcus now runs every demo through the same short list, and he asks the questions out loud in front of whoever else is in the room.

• Is this rule-based automation, where we set the rules, or machine learning, where the system learns from data?
• What data was it trained on?
• Can you explain how it works?
• Can we audit or override its decisions?
• What happens when it is wrong?

A vendor with a real product can answer all five in plain language. A vendor who cannot, and who retreats into "advanced algorithms" or "proprietary AI," is usually telling you either that they will not explain it or that they cannot. Either way you have learned what you needed to know. And the bias question rides along with the second one: if a model learned from your historical hiring, and that hiring favored one group over another, the model will faithfully reproduce the pattern. It is not being malicious; it is doing exactly what it was built to do. That is why knowing which data trained a model is a fairness question first and a technical question second.

Is Any of This Real, or Is It Hype?

It is honestly both, and holding both at once is the mature position. Pattern matching at scale is genuinely powerful: a system can process thousands of resumes in seconds, surface correlations in hiring data no person would have time to find, and produce drafts that save real hours. Those capabilities are not imaginary, and the speed and scale are the whole point. What is overblown is the suggestion that AI will replace recruiters, make flawless hiring decisions, or solve problems it structurally cannot touch. Know what it is good at, which is volume, variation, and summarization, and what it is not good at, which is judgment, context, and accountability. The realistic future of recruiting is not AI or humans. It is humans and AI together with clear boundaries: the system handles volume and structure, the person provides judgment, context, and answerability.

Related Lessons

This lesson is the vocabulary floor. Several others build directly on top of it.

• AI in Recruiting: Realistic Capabilities and Common Use Cases takes the layers you just learned and maps them onto the specific tasks in a hiring workflow, so you can see where each one earns its place.
• The Hype Cycle and How to Think Critically About AI Claims extends the five vendor questions into a fuller method for testing marketing claims before they become budget commitments.
• How AI Can Perpetuate or Amplify Bias picks up the point that a model trained on biased history will reproduce that history, and works through the mechanisms in detail. If only one follow-up lesson fits your schedule, make it this one.
• What AI Cannot Do: Hallucinations, Limitations, and Failure Modes goes deeper on the hallucination example above and on the other ways confident output turns out to be wrong.
• Third-Party Tools and Vendors: Due Diligence and Contracts turns the diligence habit into a repeatable process for evaluating and contracting with the vendors whose demos you are about to sit through.

Key Takeaways

• AI is a category, not a capability. "We use AI" tells you nothing actionable on its own. It is the umbrella term for any system doing work that normally requires human intelligence, so treat it as the prompt for a follow-up question, never as the answer.
• Machine learning learns its rules from data rather than from a programmer. Show it thousands of past hiring outcomes and it infers the patterns itself, at a scale no recruiter could match. Its accuracy and its fairness both depend entirely on what data it was trained on.
• NLP is the reading layer of recruiting AI. It is machine learning applied to language, the part that extracts facts from a resume, reconciles "RN" with "registered nurse," and interprets context. Without it, none of the text-based tooling could function.
• LLMs are the writing and summarizing layer. Trained on huge volumes of general text, they draft emails, summarize transcripts, and generate questions without ever seeing your specific hiring data. That breadth is their strength and the reason they know nothing about your roles unless you tell them.
• The four terms nest, they do not compete. An LLM is a kind of NLP, which is a kind of machine learning, which is a kind of AI. Most modern recruiting tools use several layers at once, each handling the job it is suited for.
• Newer is not automatically better for your task. A focused machine learning model trained on your own outcomes often outperforms a general LLM for ranking candidates. Match the layer to the job rather than chasing the most advanced-sounding option.
• Definitions are the floor, not the ceiling, of good evaluation. Knowing the vocabulary lets you ask the questions that matter: trained on what data, built for what problem, constrained how. That diligence is what separates buying well from getting sold.