Intro to AI for Educators: The Latest in a Very Long Line

Start with the continuum, not with the panic

Every conversation I have with a school, a university, or a training organization about AI starts in the same place: fear, or hype, and usually both in the same room. Someone is certain this ends education. Someone else is certain it saves it. Both are reasoning from the assumption that something unprecedented has happened.

Something important has happened. Unprecedented is a different claim.

Walk backwards with me. Roughly twenty-four centuries ago, in Plato’s Phaedrus, Socrates objected to the written word. He argued it would destroy memory, plant forgetfulness in the souls of students, and produce people who seem wise without being wise. Read that objection again with today’s ears. It is, almost line for line, the argument being made this year about AI. And writing turned out to be the technology that made civilization cumulative.

That is not a reason to be careless. It is a reason to be curious. When a fear is that old and that recurrent, the honest move is neither to dismiss it nor to surrender to it, but to ask what the record actually shows.

What the continuum actually looks like

Here is the line I ask educators to hold in their heads. Not as trivia — as a working tool. Each of these changed what a teacher does, what a learner does, and what counts as knowing.

• Books. Knowledge stopped living only in a teacher’s mouth and a student’s memory. Suddenly a learner could consult a mind that was not in the room, and could return to it.
• The abacus. The first great externalization of calculation — thinking placed onto an object you move with your hands.
• The encyclopedia. Knowledge organized for lookup rather than for narrative. There were once salesmen who went door to door selling families a shelf of certainty, and parents who bought it as an investment in their children’s futures. That shelf was a serious piece of educational technology.
• The blackboard. Thinking made public and provisional. A teacher could now work out a problem in front of learners instead of only presenting the finished result — arguably the single most underrated pedagogical invention on this list.
• The calculator. Arithmetic became optional and understanding became more important, not less. There was real panic, real bans, real conviction that a mathematically illiterate generation was coming.
• Whiteboards, then digital boards. The same public thinking surface, but now erasable without dust, then saveable, shareable, replayable.
• Slides. Then PowerPoint. The lesson became a sequence of prepared frames. It bought enormous clarity and it cost something too — every teacher who has watched a room go glassy at slide forty knows exactly what.
• Multimedia, then multimedia encyclopedias. Sound, video, and interactivity arrived on a disc, and a child could hear a whale and watch a volcano rather than read about them.
• The Internet. The answer to any what, who, when, where question became universally and instantly available. This one alone should have already reorganized assessment — and in most places it has still not.
• Certifications. Proof of capability decoupled from the institution that taught it. Learning began to be recognized wherever it happened.
• Gamification. Motivation treated as designable rather than as a fixed trait of the learner.
• Active learning and constructivism. Not devices at all — but innovations every bit as transformative. The learner stopped being a container to fill and became the person doing the building.
• Microlearning. Knowledge cut into pieces small enough to fit a real adult life.
• Deliberate practice. The insight that repetition with feedback and rising difficulty is qualitatively different from simply doing an activity a lot.

And more besides — the printing press, the overhead projector, the language lab, the photocopier, the learning management system, the smartphone in every pocket.

Here is what I want you to notice. Not one of these destroyed learning. Every one of them relocated the human work. Writing moved it from memorization to interpretation. The calculator moved it from computation to modeling. The Internet moved it from recall to judgment. Each time, the educators who thrived were the ones who found where the human work had gone and taught that. The ones who fell behind kept assessing the thing the tool now did for free.

Something to sit with: think of one thing you currently assess that a phone could answer in four seconds. Not to feel bad about it — every one of us has one. Just to locate where the work has already moved without our permission.

So what makes AI different?

Different in kind, I would say — not different in importance. The calculator was very important. So was the Internet. AI is strange in a way they were not, and naming the strangeness is what dissolves most of the confusion.

A calculator was a tool. You picked it up, it did a defined thing, you put it down. The Internet was a platform. It was not a thing you used so much as a place where things happened — where you searched, published, met, and worked.

Artificial intelligence is a tool, and a model, and a platform — all at once.

It is the instrument you pick up to draft a rubric. It is the model that generates language, images, and code — a thing that produces rather than merely processes. And it is increasingly the environment in which learning happens: the tutor, the study partner, the writing surface, the place a student goes at eleven at night when the assignment is due and nobody else is awake.

That triple nature is precisely why AI feels confusing in a way earlier technologies did not. Every previous innovation could be filed under one heading. This one refuses. That is not a reason for fear and it is certainly not a reason for hype. It is a reason to actually understand it — which is what the rest of this piece is for.

1. What AI is

Strip away the marketing and it is simpler than it sounds. The AI systems educators are meeting right now — the chat assistants, the image generators, the writing helpers — are prediction machines built on patterns found in enormous amounts of text and images.

A language model does one core thing extremely well: given everything so far, it produces what most plausibly comes next. Then it does it again. And again. Out of that repeated, very well-informed guess comes an essay, a lesson plan, a translation, a piece of code.

It helps to be precise about the vocabulary, because it gets used loosely:

• Artificial intelligence is the broad field — machines performing tasks we associate with human intelligence.
• Machine learning is the dominant approach: rather than being told the rules, the system derives them from examples.
• A model is the trained result — the artifact that holds those derived patterns.
• Generative AI is the family of models that produce new content rather than only classifying existing content.
• A prompt is what you give it. And your prompt is a far larger determinant of what you get back than most people realize.

An analogy I find holds up: a calculator contains arithmetic. A language model contains something more like the shape of how people write — the contours of explanation, argument, story, and instruction, learned from watching a staggering amount of human expression. Ask it for the shape of a good rubric and it will give you one, convincingly. Ask it whether that rubric is right for your students, and you have wandered outside what it can actually know.

2. How it was trained

Understanding training is what makes everything else make sense, so it is worth five honest minutes.

Roughly, in three movements.

First, it read. An enormous body of text — web pages, books, articles, code, conversation — was fed through the system. For each fragment, the model was asked to predict the next piece, then corrected against what actually came next. Repeat that astronomically many times and the internal settings of the model gradually organize themselves so the predictions get good. Nobody wrote grammar rules into it. It absorbed them, the way a child absorbs the grammar of a first language long before anyone names a verb.

Second, it was shaped. A raw predictor is not yet a useful assistant. Human reviewers rated responses, showing the model what a helpful, honest, appropriately careful answer looks like. This stage is where a great deal of its personality — and its politeness, and its caution, and its tendency to agree with you — comes from.

Third, it was fenced. Guardrails, refusals, safety layers, and usage rules were added around the model to constrain what it will produce.

Three consequences follow directly, and every educator should hold them:

• It inherits its training data. Every bias, gap, and skew in what it read is a bias, gap, and skew in what it produces. If a body of writing under-represents your students’ language, region, or reality, the model will too.
• It has a horizon. Its knowledge ends where its training data ended. It can be confidently out of date and will rarely announce it.
• It was optimized to be agreeable. That shaping stage rewarded helpfulness and encouragement — which is wonderful for a discouraged learner and quietly dangerous for an over-confident one. Hold that thought; it comes back shortly with my son in it.

3. How it “thinks”

The scare quotes are load-bearing and I am not going to remove them.

The model is not understanding your question in the way a colleague understands it. There is no comprehension of meaning underneath, no model of the world being consulted, no moment where it grasps what a fraction is. There is pattern and prediction, operating at a scale and fluency that produces something that reads exactly like understanding.

Which is why the most important single fact about AI for any educator is this: it can be completely, fluently, confidently wrong.

Not wrong the way a search engine is wrong, returning a bad link you can see is bad. Wrong the way a very articulate student who did not do the reading is wrong — smooth, plausible, structured, citing a study that does not exist. The field calls this hallucination or confabulation. I prefer to describe it plainly: the system is producing what a correct answer would look like, and looking correct and being correct are not the same property.

This has a direct and slightly uncomfortable implication. The person best protected from AI error is the person who already knows the subject. An experienced teacher spots the invented citation instantly. A student meeting the topic for the first time has no defense at all. Fluency is not evidence. Confidence is not evidence. And the more polished the output, the more carefully it deserves to be read.

Try this with a class: ask the AI something you know cold — deeply, in your own subject. Push it into the corners. Watch where it goes smooth and thin. That single exercise teaches more about the nature of these systems than any explainer, mine included.

4. AI in the classroom

Now the concrete part. Here is what this actually looks like on a Tuesday, in the work you already do.

For your own preparation — and this is where nearly every educator should start, because the stakes are low and the relief is immediate. Draft a rubric and then argue with it. Generate fifteen practice problems at three difficulty levels. Turn a dense reading into three versions at different reading levels for the same class. Produce the worked example you did not have time to write. Get a first draft of the parent letter you have been avoiding.

For feedback. Not grades — feedback. A learner can get a response to a fourth draft at ten at night, which no human system has ever been able to offer at scale. The judgment about the work stays yours. The tireless, patient, immediate commentary does not have to.

For practice. This is the use I am most enthusiastic about, and it connects to a conviction I hold strongly: mastery requires extensive, deliberate, repeated practice, and nothing removes that requirement. What AI can finally do is make that repetition bearable — varied, instantly corrected, patient, personalized, never bored of you. Deliberate practice at a scale no era of education has ever managed. That is not a small thing. That may be the largest single opportunity on this list.

For access. Translation for a newly arrived student. Text-to-speech and speech-to-text. Simplification for a learner who reads below grade level and shouldn’t be locked out of the ideas. Explanations in a home language a family actually speaks. For learners at the margins of a classroom, this can be the difference between participating and watching.

For thinking with. Socratic questioning, devil’s advocate, a debate partner who never tires, a simulated historical interlocutor, a first reader for an argument. This is the use with the most upside and it requires the most design — which is exactly why it should come after you are comfortable, not before.

And here is the moment I want to sit with, because it is the one I keep returning to. When my son and I sit down to prepare for an exam, he comes back after twenty minutes: “Dad, I’m ready.” And I say no — we continue until we complete the hour and the hundred exercises, as we agreed. And he protests: “But Dad, the AI told me I’m ready, that I’m crushing it.” And I explain: of course it did, because we configured it to build your confidence and get you through the material. But real skill, real mastery, is built by sustaining it over time — not in twenty minutes and twenty questions. Let’s finish the hundred and the full hour, and if the result is the same, we go for ice cream.

That small negotiation contains the whole lesson of AI in education. The tool did its job. Holding the standard was my job — and it is not a job that can be delegated to the tool. The AI’s encouragement was not lying, exactly. It was doing the work we gave it.

5. Challenges and opportunities

Let me put them side by side, because in almost every case they are the same fact seen from two directions.

The challenge of the shortcut. A student can now produce a finished-looking artifact without having done the thinking that the artifact was supposed to require. The opportunity in the same fact: this finally forces the question we have avoided since the Internet arrived — what were we actually assessing? If an assignment can be completed without learning, the assignment was already measuring the wrong thing. AI did not create that problem. It made it impossible to ignore.

The challenge of the illusion of mastery. Encouraging, always-available help can manufacture the feeling of competence without the substance. The opportunity: the same system, configured differently, can deliver relentless deliberate practice — the thing we have always known builds mastery and never had the capacity to provide.

The challenge of the widening gap. Well-resourced schools will get AI plus teachers who understand how to use it. Under-resourced schools risk getting the tool handed over with nothing built around it. The opportunity: good design here is genuinely levelling — a patient one-to-one explainer in a learner’s own language has historically been available only to families who could pay for it.

The challenge of erosion. Some capacities really do atrophy when unused, and we should be honest that every technology on the continuum traded something away. The opportunity: being deliberate for once about what we are trading and what we are protecting, instead of discovering it a decade later.

6. Ethics and responsibility

This section is short and it is not optional.

Privacy. Student data typed into a consumer AI tool has left your institution. Before adopting anything, know where the data goes, whether it trains the model, who can see it, and what your local law requires. When in doubt, do not put a named student’s work, records, or personal circumstances into a general-purpose tool.

Bias. The model inherits its training data. Expect skews — in whose names, dialects, histories, and examples come out as default. Notice them out loud with students. A biased output examined together is a better lesson than a clean output taken on faith.

Authorship. Decide, state, and teach what counts as the student’s own work in your context — and be specific, because “don’t use AI” is neither enforceable nor educational. Better: use is allowed here, disclosed there, and prohibited in this specific place, and here is why in each case. Students respect a clear line far more than a vague prohibition.

Equity of access. Not every learner has the same device, connection, or paid tier. An assignment that quietly assumes access to a good AI tool is an assignment that quietly sorts students by household income.

Honesty with students. Tell them what these systems are. Tell them the systems can be wrong. Tell them when you used one to prepare their material — model the disclosure you are asking them for. Learners forgive a teacher who is still learning. They do not forgive being managed.

7. Best practices and recommendations

• Keep the human judgment where it belongs. Augmentation, not replacement. The AI can draft, vary, translate, explain, and tire never. Deciding what a learner needs, whether the work is good, and what the standard is — that stays with you.
• Verify anything that looks like a fact. Names, dates, citations, statistics, quotations. If it is verifiable, verify it. If you cannot verify it, do not pass it on.
• Be specific in what you ask. Give the context, the learner, the level, the constraint, the format. Vague in, plausible-but-useless out. Prompting is closer to writing a brief than to searching.
• Move assessment toward process. Drafts, revisions, oral defense, explanation of reasoning, catching a deliberately wrong answer. If a learner can explain why it works, they learned something.
• Set the target before you start. The hour, the count, the standard. Do not let the tool’s encouragement move the finish line.
• Start where the stakes are low. Your own preparation first. Student-facing use once you have a feel for how the thing fails.
• Teach the tool alongside the subject. A student who understands what a model is and why it hallucinates is protected in a way no ban can protect them.
• Change it — but change it well. Cámbialo, pero cámbialo bien. Reforms fail not because people resist change but because they change the wrong thing, drop what was load-bearing, and keep what was decorative.

8. Getting started: three things this week

Not a strategy. Three concrete actions, each of which fits inside a normal week.

One — interrogate it in your own subject. Half an hour. Take something you know completely and push the AI until it goes thin or invents something. You are not testing whether it is good. You are calibrating your own instinct for where it fails, which is the single most transferable skill you can acquire here.

Two — hand it one real task from your own preparation. Not a student-facing one. A rubric, a set of practice problems, three reading levels of one text, a worked example. Then edit it hard. Notice what it saved you and what it got wrong about your learners. That gap is your professional judgment, made visible.

Three — write your line and say it out loud. One paragraph for your students: where AI use is expected, where it must be disclosed, where it is not allowed, and the reason for each. Read it to them. Invite the argument. Their objections will improve your policy more than any framework will.

That is the week. Everything after it builds on those three.

Where this leaves us

Artificial intelligence is the latest innovation in a very long line of innovations that have transformed and reshaped how human beings learn. Books did it. The abacus did it. The encyclopedia, the blackboard, the calculator, the projector, the disc, the web — all of them did it, and every one arrived to the sound of someone insisting that this time learning would not survive.

Learning survived. It survived because education is not a fixed set of tasks to defend. It is the deliberate cultivation of human capability, and each tool on that list, once we stopped fearing it, turned out to be one more way of cultivating more of it, in more people, than the tool before.

AI is different in one respect that genuinely matters: it is a tool, a model, and a platform at once, which is why it is confusing and why the shortcut it offers is more tempting than any that came before. That is not a reason to stand still. It is a reason to understand it well enough to shape it — and then to be the educator who noticed where the human work had moved, and went and taught that.

You already know you have to change it. The whole task is changing it well.

This introduction states a position and a frame rather than a survey of the literature. Where I have referred to history — Socrates’ objection to writing in the Phaedrus, the classroom panic over calculators, the encyclopedia’s displacement by the open web — I have deliberately kept it to what is widely documented and avoided precise figures I cannot source. If you have work I should read, tell me. — Carlos Miranda Levy

Four perspectives