AI study tools students actually use, and what teachers can see

Ask a class what AI they use and you get a shrug. Ask them what they used last night to get a specific assignment finished and you get a list. The gap between those two answers is where most school AI policy fails: it is written about a category, and students use products.

This is an inventory of the AI study tool types that actually appear in a middle or high school, what each is genuinely good at, what it hides from you, and what a teacher can realistically see. No brand rankings, because the specific products churn every few months and the categories do not.

1. The solver

Point a camera at a question, receive the answer with working. Descended from the photo-math apps that predate the current wave, and by far the most widely used category in math and science.

Genuinely good at: unblocking a student at 11pm who has attempted six questions and is stuck on the seventh. Showing a worked method in a format a textbook does not.

What it hides: everything. The output is indistinguishable from a student who understood. There is no signal about where the student stalled, and after a few weeks of use there is no signal about whether they can do any of it.

What a teacher can see: almost nothing directly. Indirectly: a widening gap between homework performance and in-class performance. That gap is the most reliable indicator you have, which is the entire argument for keeping one short device-free checkpoint each week. We go through the mechanics in AI homework help that does not do the homework.

2. The summarizer

Paste a chapter, a PDF or a lecture recording; receive bullet points. Heavily used for studying, and by students who have not read the text.

Genuinely good at: orientation before reading something difficult, and recall after reading it. Turning a forty-minute recording into something a student can scan on a bus.

What it hides: whether the student read anything. Also, quietly, the argument structure of the original. Summaries flatten hedged claims into confident ones and drop the sentence where the author explains why the obvious answer is wrong, which is usually the sentence the assessment is about.

What a teacher can see: a distinctive flatness in written responses. Students who worked from a summary produce accurate, generic answers with no textual detail and no quotation. Assessment design fixes this faster than policy does: ask for a line reference and the summary stops being sufficient.

3. The flashcard and quiz generator

Turn notes into practice questions or a spaced-repetition deck. Broadly the least controversial category in the building.

Genuinely good at: retrieval practice, which is one of the small number of study techniques with strong evidence behind it. Getting a student to do any active recall at all is a win over rereading highlighted notes.

What it hides: the quality of the source. A deck generated from a student's own incomplete notes rehearses the gaps. Generated distractors are also frequently too easy, so the student's confidence rises faster than their competence.

What a teacher can see: usually nothing, and usually that is fine. The useful intervention is to supply the source material rather than let each student generate from their own notes.

4. The writing assistant

Drafting, restructuring, tone changes, grammar. The category that dominates the academic integrity conversation in English, history and any subject with extended writing.

Genuinely good at: getting a student who freezes at a blank page to produce a first sentence. Second-language support. Feedback on structure at a level of detail no teacher can provide for thirty essays.

What it hides: authorship, obviously, but more importantly the drafting process, which is where most of the learning in writing happens. A polished submission tells you nothing about whether the student can plan, revise or notice their own weak paragraph.

What a teacher can see: more than in any other category, if the assignment collects the process rather than the artifact. Version history, an in-class plan, a paragraph written under supervision, or a short oral defense. Detection software is not the answer here; it produces false positives and disproportionately flags students writing in a second language.

5. The conversational tutor

A general chat product used as a tutor: ask questions, get explanations, go back and forth. Increasingly the default for capable students, and increasingly voice-based.

Genuinely good at: patience. A student can ask the same thing five times in five different ways without embarrassment, which is a real advantage over asking a teacher in front of a class.

What it hides: the whole session. This is the category that produces the most learning and the least evidence. Fifty minutes of genuine cognitive work leaves a transcript that nobody reads, and the teacher's picture of that student is unchanged the next morning.

What a teacher can see: nothing, unless the tool is designed to hand something back. This is the gap Coeducate was built for, and it is the difference between a transcript and a per-skill mastery record. The demo on the homepage produces the second from a piece of student work so you can see what the teacher-facing artifact is.

6. The voice explainer

Increasingly common, especially with younger students and with students who find reading effortful. A spoken conversation about a topic, often while walking home.

Genuinely good at: accessibility. For a student with a reading difficulty, a spoken explanation removes a barrier that has nothing to do with the subject being studied.

What it hides: the same as the conversational tutor, plus the fact that it happened at all. There is no artifact, no browser history entry a parent would notice, and no written trace.

What a teacher can see: nothing. Worth knowing about mainly so that "I did not study" is not taken at face value.

The pattern across all six

Read the "what it hides" lines together and the shape is clear. Every category on this list either removes information from the teacher or leaves it uncollected. None of them add any. That is a design choice by the products, not a law of nature: they are built for the student as the customer, and the student is not asking for a report to be sent to school.

The problem is not that students use AI. It is that every tool they use makes them less legible to the person responsible for teaching them.

Which suggests the actual requirement for anything a school recommends: it should be the only tool in the student's stack that produces something for the teacher. Not surveillance of what they typed. A per-skill record of what the session showed, with the student's own reasoning attached as evidence, which is a categorically different artifact.

Four moves that work better than a ban

1. Change what you collect. Collect the plan, the reasoning, the annotated draft or the line reference. Almost every category above becomes harmless when the artifact is the process.
2. Keep a short device-free checkpoint. Five minutes, one problem, once a week. Not surveillance, calibration. It tells you whether the homework signal is real.
3. Teach the categories explicitly. Students distinguish these six far better than adults expect when someone names them. "The solver is not studying, the flashcard generator is" is a sentence a fourteen-year-old can act on. A blanket ban is not.
4. Recommend one tool that reports back. If the school-provided option is the only one that produces evidence, students use it for the work that gets graded, and the teacher's picture of the class becomes accurate again.

What to say to parents

Parents ask a version of the same question: is my child cheating with this, and should I take the phone away? The useful answer distinguishes the solver from the tutor. Taking the phone away removes both, and the second one is the one families have historically paid for.

A more productive line: "We would rather they use something that makes them do the thinking and tells us what they found hard, than something that finishes the work. Here is the one we recommend, here is what it does when they get stuck, and here is what happens to their work." That last clause matters more than most schools expect. See where general chat tools break student privacy and our own student privacy page for the version we hold ourselves to.

Summary

• Six categories, not one: solver, summarizer, flashcard generator, writing assistant, conversational tutor, voice explainer.
• Only the solver and the writing assistant are usually treated as integrity problems, but every category removes information from the teacher.
• Assessment design closes more gaps than policy does, and far more than detection software.
• The requirement for a school-recommended tool is that it reports back with evidence, not that it blocks anything.
• Name the categories for students. They are better at this distinction than the policy usually assumes.

A worked example: one homework, four tools, four different blind spots

The categories above are easier to believe when you watch them applied to the same task. Four grade 8 students are set the same five percent-change problems. Each reaches for a different tool. All four hand in correct work, and what the teacher learns from each is different, and in every case inadequate.

The solver

Student one photographs question two and gets a complete worked solution back. She copies it out. What the teacher receives is a page of correct method with no information in it: it is indistinguishable from the page a student who understood it would have produced. The blind spot is total, and it is the one every integrity conversation is about.

The summarizer

Student two does not use a solver at all. He pastes the textbook section on percent change and reads the summary instead of the chapter. His homework is wrong in a specific way, treating the two 20s as cancelling, because the summary compressed away the one sentence that distinguishes the two bases. Nobody thinks of this as an integrity issue, and it is the case where the tool actually caused the misconception rather than hiding it.

The flashcard generator

Student three generates twenty cards from the chapter and drills them. She can now state the rule for percent change verbatim and still cannot apply it, because the cards tested recall and the homework tests choosing a base. Her homework is wrong and her confidence is high, which is the hardest combination to spot from a grade book and the one most likely to survive to the exam.

The conversational tutor

Student four asks a chat product to explain rather than solve. It does, patiently, and he gets the homework right. This is the best outcome of the four and it still leaves the teacher with nothing: no record that he arrived not knowing which base a percentage is taken of, and no reason to check next week whether it stuck.

What the four have in common

Every one of them removed information from the teacher, and only the first is usually treated as a problem. That is the argument of this whole piece in one homework: the integrity framing catches one of four cases, while the diagnostic loss runs through all of them. A recommended tool is not one that blocks anything. It is one that hands back what the session showed: the base confusion, the additive thinking, the notation slip. Then the four students stop looking identical on paper.

Where to read the two parts that decide this

In math the whole thing turns on how specific the skill names are. The four students above have four different faults that produce one cross each. Which fault lines a per-skill record is built to separate, across fractions, decimals, ratio, percent, algebra and measure, is set out on the AI math tutor page.

The second part is what happens to the work once a school-recommended tool reads it. Before you put a name on a letter home, get the answer in writing on training, control of the record, deletion and end-of-year retention. Ours is on the student data privacy and COPPA compliance page, including the attestations we do not hold.