AI in the classroom is worth a teacher's attention, but not every tool labeled "AI" deserves it. The useful question is not whether a product uses artificial intelligence — nearly everything now claims to — but whether it does something specific and measurable for your students without adding hours to your week. Requirements and results vary by district, subject, and grade level, so no single tool fits every room.
This guide explains what AI tools actually do, which classroom uses have held up, and which claims to treat with caution. It ends with a short list of questions to ask before any adoption, whether you are choosing a tool for your own use or sitting on a district committee.
What does "AI" actually mean in an education product?
Artificial intelligence is the capability of computational systems to perform tasks associated with human intelligence, such as learning, reasoning, and problem-solving, according to Wikipedia's overview of artificial intelligence. The field dates to 1956, and it has cycled through waves of excitement and disappointment — the periods of lost funding are known as "AI winters." Knowing that history is useful. It explains why the current boom, driven since the 2020s by generative AI that can create and modify text and media, feels both genuine and familiar.
Most education products today are built on large language models, or LLMs. These systems learn from enormous amounts of text rather than from explicit rules. That design has a consequence every teacher should understand: LLMs can produce incorrect outputs, often called hallucinations, even when they sound confident. A chatbot that explains the causes of World War I fluently can also invent a quotation or misstate a formula in the same paragraph. Fluency is not accuracy.
So when a vendor says a tool is "AI-powered," translate it into a concrete question: what task does the AI perform, and how often does it get that task wrong? A tool that drafts reading questions at five reading levels is doing something narrow and checkable. A tool that promises to "personalize learning" for every student is making a much larger claim that deserves much more evidence.
Which classroom uses of AI have held up?
The durable uses tend to share two traits: the AI handles a narrow task, and a teacher can check the output quickly.
- Drafting and differentiation. Tools that rewrite a passage at different reading levels, generate practice items, or produce first-draft lesson materials save real planning time. The teacher reviews everything before students see it, so a hallucination gets caught at your desk instead of in front of the class.
- Feedback support. Some tools give students first-pass comments on writing, which the teacher then confirms or overrides. This works when the tool flags issues rather than assigns grades. It fails when a district treats machine feedback as a substitute for teacher judgment.
- Practice and tutoring. AI tutors are the fastest-growing and least-proven category. Some show promising results in math practice; many show nothing you could not get from a worksheet. Our analysis of the vendor landscape is simple: treat every tutoring claim as unproven until the company shows its sample and its controls. We cover the evaluation questions in detail in How Districts Should Evaluate AI Tutors Before Buying.
- Accessibility. Text-to-speech, speech-to-text, translation, and captioning are older AI applications that quietly help many students. These are among the least hyped and most reliably useful tools in the category.
Adaptive learning software — programs that adjust difficulty based on student responses — predates the current boom, and the research on it is mixed. We review what the evidence shows in What Is Adaptive Learning Software and What Does Research Say. The pattern there is a useful template for AI claims generally: a few studies with strong designs, many with weak ones, and vendors quoting the flattering ones.
What are the warning signs of passing hype?
Passing hype has a recognizable shape. Watch for these signals:
- Broad promises, narrow evidence. "Transforms learning outcomes" with no named study, no sample size, and no comparison group. A vendor's claim about its own product is the vendor's claim, not an independent finding.
- Pilot pricing that hides renewal costs. Free first years are a common entry strategy. Budget for the year the invoice arrives, and read the contract for auto-renewal terms before signing. We catalog the traps in EdTech Contract Red Flags Districts Should Catch Before Signing.
- Data practices you cannot explain. If you cannot state in one sentence what student data the tool collects, where it goes, and who can see it, that is a problem. Federal law already sets the floor: FERPA and COPPA govern how schools and vendors handle student records and the data of younger children, as we explain in How FERPA and COPPA Protect Student Data in the Digital Classroom.
- No exit plan. Any tool that stores student work needs a way to export it. A product that locks in your data locks in your district.
- Teacher workload left out of the pitch. If the demo shows what students see and never what the teacher has to set up, monitor, and grade, the vendor is hiding the real cost.
One more sign sits outside the product entirely: procurement pressure. When a tool arrives because a salesperson worked the district rather than because teachers asked for it, adoption usually stalls. How school districts actually buy edtech platforms is its own process, and understanding it helps teachers push evidence into the decision. We walk through it in How School Districts Actually Buy EdTech Platforms.
What questions should you ask before adopting any AI tool?
Use this list whether you are evaluating a tool alone or on a committee. It takes about twenty minutes per product.
- What specific task does the AI perform, and what is the error rate? Ask the vendor to describe a failure case. A company that cannot name one has not tested honestly.
- What evidence supports the outcome claim? You want a named study with a sample and a control group, not a testimonial reel. If the only evidence is the vendor's own pilots, label it accordingly.
- What student data does it collect, and where does it go? Get the answer in writing, and check it against your district's data privacy agreements.
- What does the teacher have to do to make it work? Setup time, monitoring, and grading load all count. A tool that saves students ten minutes but costs the teacher an hour a week is a net loss for the classroom.
- What does it cost per seat after the pilot, and how do we exit? Ask about renewal terms and data export before the first invoice, not after.
- Does it work on the devices and bandwidth we actually have? A tool that assumes home internet excludes students without it.
The last question matters more than it sounds. Access gaps did not end with school laptop programs, and districts still manage uneven connectivity in ways that shape which tools can work at all. We cover that context in The Digital Divide After E-Rate: How Districts Keep Students Connected.
What this means for teachers
The honest picture is this: AI tools that draft, differentiate, and support accessibility are already useful and likely to stay. AI tutors and "personalization" platforms are the category where marketing runs furthest ahead of evidence, and they deserve the most skepticism, not the least. Hallucination is a structural property of the models underneath these products, not a bug that a version update removes, so human review stays non-negotiable for anything students see.
None of this guidance promises results. Your students, your state standards, and your workload differ from every district described in any study. The durable habit is the one good evaluators already use: ask what the tool changes in a real classroom, ask who measured it, and ask what it costs — in dollars and in teacher time — to keep. For more reporting and guides in this area, the publication's edtech section collects ongoing coverage, and new teachers weighing where the field is heading can also find career-side context in careers.
AI will keep changing. The evaluation discipline should not. Grade every tool the way you would grade a student's claim: show your work, show your sample, and no credit for confidence alone.
