When people talk about AI in schools, the conversation usually centers on students — are they cheating, are they learning, are they falling behind? But the quieter truth is that teachers adopted AI tools first, and they did it for a very specific reason: time.

Teaching is one of the few jobs where the visible part — standing in a classroom — is actually the smaller slice of the work. The bigger slice is everything that happens before and after: planning lessons, writing assessments, giving feedback, emailing parents, filling out paperwork, differentiating materials for students at different reading levels. A 2023 RAND Corporation survey of over 1,000 U.S. teachers found that the average teacher spends 10.5 hours per week on tasks that don't involve direct instruction. That's more than a quarter of a full-time job's worth of invisible work, every single week.

AI tools showed up and said, in effect: I can help with the invisible work. And teachers, who are chronically overworked, said yes — often quietly, often without training, and often without any school policy telling them whether they were allowed to.

The tools they reached for fell into a few clear categories, and understanding those categories is the first thing you need to do to make sense of this whole landscape.

By mid-2024, the EdTech company HolonIQ had tracked over 200 AI tools marketed specifically to K–12 teachers. Most of them fell into five buckets. Knowing these five categories means you can categorize almost any new tool the moment you hear about it — instead of treating every announcement like it came from nowhere.

1. Feedback and grading assistants — Tools like Writable, Turnitin's AI features, and EssayGrader analyze student writing and suggest feedback. They look at things like argument structure, sentence variety, and whether a claim is supported. They do not understand whether something is true or meaningful to a specific student's life.

2. Lesson planning generators — Tools like MagicSchool AI, Curipod, and Diffit let teachers type a topic and a grade level, and receive a full lesson plan — objectives, activities, discussion questions — in under a minute. MagicSchool AI reported 3 million teacher users by early 2024, roughly one in six U.S. teachers.

3. Differentiation tools — These take a piece of text and rewrite it at different reading levels, or translate it, or add vocabulary supports. Diffit is the most widely used. A teacher with a class of students reading anywhere from 4th-grade to 10th-grade level can get a version of the same article adapted to each level in seconds.

4. Chatbot tutors for students — Tools like Khanmigo (built by Khan Academy in 2023) and Synthesis Tutor put an AI directly in front of students. Instead of giving answers, these tools are designed to ask questions back — to guide thinking rather than replace it.

5. Detection and integrity tools — Turnitin added an AI detection score in April 2023. GPTZero launched in January 2023, built by a Princeton student named Edward Tian, and was downloaded 7 million times in its first week. These tools try to detect whether student work was written by AI. They are imperfect in ways that matter — we'll get to that.

Here is the part that almost no article about AI in schools mentions: most teachers who use these tools received no formal training on them. A 2023 survey by the nonprofit ISTE found that 69% of teachers who reported using AI tools said they were "self-taught." They learned from YouTube videos, from colleagues, from trial and error.

This is not unusual in the history of technology in schools. When the photocopier arrived in the 1960s, no one trained teachers on intellectual property. When the internet arrived in the 1990s, most schools let teachers figure out search engines on their own. When smartphones became universal around 2012, districts wrote policies years after students already had the devices in their pockets.

The pattern is always the same: the tool arrives, people start using it without a framework, and then the frameworks get built — sometimes well, sometimes badly — years later. Right now, we're in the "people using it without a framework" phase, and that has real consequences. A teacher who doesn't understand how an AI grading tool was trained might not notice when it consistently rates essays by English-language learners lower. A school that bans AI entirely might be cutting students off from skills they'll need in every job they ever hold.

The gap between "available" and "understood" is where most of the interesting problems in AI education live right now. And it's where your generation — the one that grew up with both the technology and the questions about it — has a genuinely unusual advantage.

Return to Amy Jeter for a moment. She used an AI tool to generate first-draft feedback on student essays. She edited that feedback before students saw it. She never mentioned that AI had been involved. Her students believed they were getting their teacher's assessment of their work.

Is that a problem?

On one hand: the feedback was accurate. Jeter reviewed it. Students got more detailed comments than they might have received if she'd been exhausted on Sunday night. On the other hand: students have a reasonable expectation of knowing who — or what — evaluated their work. If a doctor had an AI draft your diagnosis and never told you, you'd probably want to know that, even if the diagnosis was correct.

Different schools have landed in very different places. Some require teachers to disclose any AI-assisted feedback to students. Some require students to disclose AI use in their work, but have no rule about teacher use. Some have no policy at all. The inconsistency itself tells you something important: this is genuinely unsettled territory, and the people making the decisions are still figuring it out in real time.

Knowing this landscape — the five tool categories, the policy lag, the disclosure question — means you're no longer just a person things are being done to in a classroom. You understand the system you're sitting inside.

When a company calls their product an "AI tool for teachers," the word AI is doing a lot of work. Most of the tools teachers use in schools are built on a specific kind of AI called a large language model, or LLM — the same underlying technology as ChatGPT. Understanding what an LLM is, at a basic level, makes every AI claim in education immediately easier to evaluate.

Here is the key thing to hold onto: an LLM has read more text than any human ever could, but it has never had an experience. It has never been confused by a teacher's explanation, never felt the relief of finally understanding something, never written an essay under pressure the night before it was due. This means it is very good at producing text that resembles expert feedback — and genuinely bad at understanding why a specific student wrote what they wrote.

When Writable generates feedback on a student essay saying "your argument in paragraph three lacks supporting evidence," it's pattern-matching: it has seen enough essay feedback to know that weak paragraphs often get this note. It is not reading the essay the way a teacher who knows the student reads it. This is not a flaw that will be fixed by a better version. It's a fundamental feature of what the tool is.

Every AI tool used in schools was trained on a dataset — a collection of examples the model learned from. For feedback tools, that's usually a large collection of essays and the feedback teachers wrote on them. For lesson plan generators, it's lesson plans, curriculum guides, and educational standards documents. For AI tutors, it's conversations, textbooks, and subject-matter content.

The training data problem is this: whatever was unequal in the real world shows up in the training data, and therefore shows up in the tool. The College Board story illustrates this exactly. Students from well-resourced schools had better AP preparation, produced higher-scoring essays, and those essays became the model's definition of "good." A student whose essay reflects a different but equally valid way of constructing an argument — common in many non-Western rhetorical traditions, for example — may be rated lower not because their argument is weaker, but because it doesn't match the pattern the model learned.

This problem appeared in a different context in 2018 when Amazon scrapped an AI hiring tool that its engineers had built. The tool had been trained on a decade of Amazon's hiring decisions. Amazon had historically hired more men than women in technical roles. The AI learned that and began down-ranking resumes from women. Amazon shut it down. The same logic applies to educational AI: garbage in, inequity out — even when nobody decided to put the inequity there.

Fair is fair: these tools do several things very well. Lesson plan generators produce structurally sound, curriculum-aligned plans in seconds. Differentiation tools produce readable, grade-appropriate text quickly. Writing feedback tools catch common surface-level issues — unclear topic sentences, missing citations, passive voice overuse — accurately and consistently. These are real time savings for real work.

But there are specific places where they break, and knowing those breakpoints is what separates a teacher who uses these tools well from one who uses them blindly:

They break on context. An AI feedback tool doesn't know that a student just moved schools, lost a family member, or is writing in their third language. The feedback it generates is context-blind. A teacher reading the same essay with that knowledge might respond very differently.

They break on novelty. LLMs are good at patterns they've seen before. An unusual essay structure, an unconventional argument, a creative risk — these often get penalized or misread because they don't match the training distribution. Students who take interesting intellectual risks can be scored lower than students who write competently within familiar patterns.

They break on factual currency. Most of these tools have a training cutoff — a date after which they have no knowledge of new events. A lesson plan about climate change might omit the most recent major IPCC report. A history lesson plan might not include recent scholarship that revised older interpretations.

Here is the ethical question from this lesson: When a student receives feedback, do they have a right to know whether a human being actually thought about their specific work — or whether a pattern-matching system processed it?

This question gets more complicated when you add the possibility of error. AI feedback tools produce incorrect or irrelevant feedback a meaningful percentage of the time. In a 2023 study by researchers at Stanford, AI feedback on student essays matched expert human feedback closely in roughly 65% of cases — which sounds good until you realize it was significantly off in 35% of cases. A student who trusts the AI feedback and rewrites their essay accordingly might make it worse.

The deeper question is about what feedback is for. If feedback is just a correction mechanism — find the errors, fix them — then a tool that's right 65% of the time is useful. If feedback is also about a student feeling seen, feeling like an adult engaged seriously with their thinking, then no AI tool yet built can do that. And that distinction affects how much we should rely on these tools, how transparent teachers should be about using them, and what we're actually losing when the human step gets compressed.

You now understand something that most parents and many teachers don't have a precise language for: the difference between a tool that produces output that looks like expertise and a tool that actually applies expertise. Those are not the same thing. The gap between them is where most of the real risks in educational AI currently live.

The first question most schools, parents, and news articles reached for was: Is this cheating? It's an understandable question, and it's not irrelevant. But starting there meant most institutions got stuck there, spending energy on detection and enforcement rather than the more difficult underlying question.

The more difficult question is: What were we actually trying to accomplish with the assignment that the AI can now do?

Consider a standard five-paragraph essay assigned in English class. The stated purpose is usually "to develop the student's ability to construct a written argument." But AI can construct that argument in thirty seconds. If the tool can do the thing the assignment was measuring, then either the thing being measured wasn't valuable to begin with — or we need to get much more specific about what part of the process we actually care about.

Ethan Mollick, a professor at the Wharton School of the University of Pennsylvania, published a paper in 2023 arguing that the arrival of capable AI writing tools should force educators to do something they'd been avoiding: be precise about which cognitive skills assignments are actually developing, and redesign assignments around the parts AI can't do. He called this "assignment renegotiation" — and it's the frame that the most thoughtful educators started using in 2023.

The media narrative about students and AI in 2023 was almost entirely about cheating. The actual picture was more complicated. A large-scale survey by the Walton Family Foundation in early 2023 — covering over 1,000 students ages 12–18 — found that student AI use fell roughly into four patterns:

Cheating-adjacent: Using AI to write full assignments and submitting them without disclosure. This was the pattern everyone was worried about, and it did exist. Surveys suggested 10–20% of students had done this at least once.

Research scaffolding: Using AI as a starting point — asking it to explain a topic, then doing further reading. Students described this as "like asking a friend who knows a lot" before going to the actual sources. This was the most common pattern.

Feedback seeking: Submitting their own draft to AI and asking for suggestions before turning it in. The students doing this often described it as filling a gap — they didn't have adults available to read their drafts and give feedback before submission.

Exploration: Using AI for topics unrelated to school — creative writing, game design, personal questions, creative projects. This was especially common among students ages 12–14.

These four patterns have very different implications. The "cheating-adjacent" pattern is the one that erodes learning. The others are, in many ways, exactly what we'd want students doing with a powerful information tool — if there were adults helping them use it critically. The problem is that most students in 2023 were using these tools entirely without adult guidance, because schools were busy writing bans.

Here's where the conversation gets genuinely difficult, because not everyone agrees on what learning is.

One view: learning is fundamentally about building cognitive patterns through practice. When you struggle to write an essay — when you can't find the right word, when you realize your argument doesn't hold together, when you rewrite a paragraph four times — you are building something in your brain. That struggle is the mechanism by which skill gets formed. If AI removes the struggle, it may remove the learning, regardless of whether the output looks good.

Researchers studying this effect point to a parallel with GPS navigation. Studies in the early 2010s showed that people who relied heavily on GPS navigation performed significantly worse on spatial memory tasks and had lower activation in the hippocampus — the brain region involved in navigation — than people who used paper maps or memorized routes. The GPS produced the desired output (you got to your destination) but removed the cognitive work that builds the underlying skill.

The counter-argument: schools have historically taught skills that turned out to be less critical than believed, once technology changed. Handwriting instruction dominated elementary school for generations after typewriters existed. Memorizing state capitals is still a classroom staple even though everyone carries a device that knows them instantly. Maybe "constructing a written argument from scratch" is like knowing the capital of Nebraska — a skill with diminishing returns in a world where the tool does it better than most people ever will.

The honest answer is that we don't have enough long-term data yet to know which view is right. What we do know is that this question is being answered — by default, by inaction, by individual teacher choices — in schools right now, without the research to back up any confident conclusion.

Here is the ethical tension from this lesson, and it's one that isn't talked about nearly as much as the cheating question:

AI tools are not equally available to all students. A student with a reliable internet connection, a laptop, and parents who can help them use these tools thoughtfully has a significant advantage over a student using a shared family phone on a slow connection with no adult guidance. This isn't new — the same gap exists with tutors, with test prep, with home libraries. But AI accelerates it, because the quality difference between "using AI well" and "not using AI at all" is larger than the quality difference between "having a tutor" and "not having a tutor."

Schools that allowed AI use freely in 2023 may have inadvertently widened the advantage gap between students who had support and students who didn't. Schools that banned AI may have penalized students who were using it responsibly while doing nothing to stop students with the knowledge and tools to hide their use.

There is no clean answer here. But recognizing that the equity question and the cheating question are different questions — and that most schools conflated them — is the kind of precision that makes a difference when these policies get made. That precision is now yours.

When you sit in a classroom that uses an AI grading tool, or get feedback that was AI-assisted, or use a lesson that a teacher built with MagicSchool AI — a chain of decisions was made before that experience reached you. Understanding that chain is part of understanding the landscape.

The decisions happen at multiple levels, and they don't always align:

The technology companies building these tools make the first and most consequential set of decisions: what data to train on, what features to build, what to optimize for, what to measure, what to disclose. These decisions happen inside private companies and are largely not visible to schools or students. The companies have a financial incentive to sell the product and a reputational incentive to make it look good. These are not the same as an educational incentive.

School districts decide which tools to adopt, usually through a procurement process that involves a committee, a vendor demonstration, and a contract. Most district committees are not equipped to evaluate AI systems technically. They're making decisions based on marketing materials, peer recommendations, and price.

Individual teachers decide how to use the tools once they exist in a school — and, as Lesson 1 established, often adopt their own tools independently without district approval. This creates a situation where the actual AI use in a school can be significantly different from what the official policy says.

Students are almost never formally included in these decisions. They are the people whose learning, privacy, and futures are most directly affected, and they are the least represented voice in the governance structures.

When a student uses a school-provided AI tool — a writing assistant, a tutoring chatbot, a feedback platform — that tool collects data. It collects what the student typed, how long they spent, what they revised, what questions they asked, how their writing changed over time. This is, in many ways, more intimate data about a student's thinking than anything stored in a grade book.

In the United States, student data is nominally protected by a law called FERPA (Family Educational Rights and Privacy Act), passed in 1974 — almost fifty years before LLMs existed. FERPA was designed for paper records. It does not cleanly cover the kinds of behavioral and interaction data that AI tools collect. A 2023 report by the Electronic Frontier Foundation found that dozens of AI tools widely used in schools had vague or inadequate privacy policies, and that many retained student interaction data in ways that could be used for purposes unrelated to education.

This matters because the data collected by an AI tutoring tool doesn't disappear when a student graduates. It exists somewhere — on the vendor's servers, possibly shared with partners, possibly used to train future versions of the model. A student's pattern of questions, their areas of confusion, their writing over time — this is a rich portrait of how that person thinks. Who owns it? Who can access it? Who can sell it? These questions are mostly unanswered right now.

It's worth being concrete about what a well-designed, well-governed AI integration in schools would actually look like — not as an ideal nobody reaches, but as a standard against which current practice can be measured.

In 2024, a coalition of educators, researchers, and AI ethics specialists released a set of principles they called the Student-Centered AI Framework. Their core criteria for a trustworthy educational AI tool were four:

Transparent: Teachers and students know when AI is involved in producing content they receive — feedback, explanations, lesson materials. No silent AI.

Audited: The tool has been independently tested for demographic bias before it's used in schools, and those results are publicly available. Not just the vendor's internal testing.

Controlled: Teachers can override, adjust, or disable AI outputs at any point. The AI is a tool under human control, not a system making unilateral decisions about students.

Privacy-preserving: Student interaction data is used only to improve that student's experience, is not retained beyond the school year, and cannot be sold or shared with third parties.

Looking at the landscape described across all four lessons, most currently-deployed educational AI tools meet some of these criteria — and fail others. Lesson planning generators are generally transparent (teachers know they used them) but rarely audited for bias. AI grading tools often aren't transparent to students and rarely have published independent bias audits. AI tutors often have unclear data retention policies.

The framework exists. The implementation is uneven. The gap is where the work is.

Here is the final ethical tension of this module: Should students have formal input into the AI tools that are used to evaluate them?

At most companies, products are tested with representative user groups before launch. At most schools, AI tools are adopted without anyone asking the students who will use them — or be evaluated by them — what they think. The Department of Education's 2023 report said educators should be primary designers of educational AI. It did not say students should be included in that process.

There are reasonable arguments on both sides. Students might lack the technical knowledge to evaluate AI systems. They might have short-term preferences that don't align with long-term learning goals. On the other hand: students are the only people in the system who simultaneously experience being a learner, a person being evaluated, and a person whose data is being collected. They have a perspective that no educator, administrator, or AI ethics researcher can fully substitute for.

The question of who gets to shape the tools that shape education is not a technical question. It's a political one. It gets answered through policy, through advocacy, through the voices that get included in the room when decisions are made.