A year ago, asking a computer to draw a labeled diagram of the water cycle, in the style of a 1950s textbook, with one deliberate error for students to catch, would have sounded like a party trick. Now it is a Tuesday. The newest wave of AI image tools, from Google's Nano Banana and Nano Banana Pro to OpenAI's image models and Midjourney, can produce usable pictures in seconds from a sentence of plain English. The internet is full of prompt lists promising you can generate anything you can imagine.
That is true, and it is also not very helpful. "You can make anything" is the kind of sentence that sounds exciting and tells a teacher nothing about Monday morning. So let us be specific. What are these tools actually good for in a school, where do they quietly fall apart, and what should a teacher or parent keep in the back of their mind while the whole category keeps moving?
The boring uses are the good ones
The most valuable classroom applications of AI image generation are unglamorous, and that is exactly why they work.
Take worksheets and slide decks. Every teacher has spent an evening hunting for a clip-art image of a specific thing, a cross-section of a leaf, a Roman aqueduct, a friendly cartoon thermometer, and either come up empty or landed on something with a watermark and uncertain licensing. Image generators collapse that search into a request. You describe the thing, you get the thing, and you get it sized for the slide you are building. The picture does not need to be a masterpiece. It needs to be clear, on-topic, and ready in under a minute.
Differentiation is another quiet win. The same vocabulary word can be illustrated three ways: a simple icon for an emerging reader, a detailed scene for a stronger one, a labeled version for a student who needs the scaffolding. Producing three variants used to mean three separate searches. Now it means three lines in the same prompt.
Then there is the editing trick that the newer models, Nano Banana Pro among them, have gotten genuinely good at: you hand the model an existing image and tell it what to change. Make the background a desert. Add a snowstorm. Show this same room a hundred years later. For a history or science class, the ability to take one base picture and run it through a sequence of changes is a small superpower. It turns a static image into a before-and-after, a then-and-now, a what-if.
Where it gets interesting for student work
The classroom uses above are teacher-facing. The student-facing ones are where the conversation gets more careful, and more interesting.
Used well, image generation is a thinking tool, not a shortcut. A literature class can ask students to generate the setting of a novel exactly as the text describes it, then defend every choice against the words on the page. Why is the door green? Where does it say that? The image becomes evidence of close reading, or evidence of skipping it. A design or engineering class can iterate through twenty product concepts in the time it once took to sketch two, then explain why one survived. The value is not in the picture. It is in the reasoning the student has to do around it.
The trap is obvious and worth naming. If the assignment was the picture, the AI did the assignment. If the assignment was the argument about the picture, the AI did the easy part and left the hard part to the student. Good prompts for student use are the ones that make the AI output the raw material and leave the judgment to the human. That distinction is the whole game.
What these tools are still bad at
It is easy to watch a clean demo and forget the failure modes. They are real, and in a school they matter more than usual.
Text inside images is the classic weak spot. It has improved a lot, and Nano Banana Pro markets legible text as a feature, but generators still mangle words, especially anything long, technical, or in a non-Latin script. Never trust an AI-generated image to spell a key term correctly on a handout without checking it yourself.
Accuracy is the bigger issue. These models produce something that looks like a labeled diagram, not necessarily a correct one. The bones of a generated skeleton may be in the wrong places. The countries on a generated map may be the wrong shape. For decoration, fine. For instruction, every factual image needs a human who actually knows the subject to sign off before it reaches a student, because a confident, polished, wrong diagram is worse than no diagram at all.
And the models carry the biases of what they were trained on. Ask for "a scientist" or "a CEO" or "a nurse" and watch who shows up by default. In a classroom that is not a neutral quirk. It is a teachable moment if you catch it, and a quiet lesson in stereotypes if you do not.
The rights question nobody wants to read
Two things are worth knowing, even briefly, because they will come up.
First, the training-data fight is unresolved. These models learned from enormous piles of images, much of it scraped from the open web, and artists and stock companies have lawsuits in progress arguing that was not okay. None of that is settled. A school does not need to take a side, but it should not pretend the tools arrived clean.
Second, generating in a living artist's style is easy and ethically loaded. "In the style of" a famous illustrator is a single phrase away. For a lesson on artistic technique, that might be defensible. For a poster you publish under the school's name, it is a conversation worth having first. Teaching students to ask "should I" and not just "can I" is arguably the most durable thing in this entire topic, because the tools will change and that question will not.
How to actually start
If you want one practical on-ramp, it is this: pick a single recurring task you already do and hand it to the tool. Not your whole workflow. One thing. The illustration for next week's vocabulary list. The header image for the class newsletter. A visual for the one slide that always feels empty.
Write the prompt like you are briefing a fast, literal, slightly careless intern. Say what it is, what style you want, what to include, what to leave out. Look hard at what comes back, fix it or rewrite the prompt, and check anything factual before it goes near a student. Do that a few times and you will develop an instinct for what these tools do well and where they need a human standing over them.
That instinct, not any specific prompt list and not any specific model, is the skill worth building. Nano Banana will be replaced by Nano Banana Pro will be replaced by whatever ships next quarter. The teacher who knows what to ask for, and what to double-check, keeps every bit of that knowledge no matter which logo is on the button.
The picture takes a second. Knowing whether it belongs in front of a child is still the job.