Table of Contents
An AI Learning Tools Review Found None Ready to Teach Alone
A review of 20 AI learning tools across 16 school systems found purpose-built tools beat chatbots. What the AI learning tools review means for your child.
The most useful AI learning tools review published this autumn did not crown a winner. On September 14, 2026, the nonprofit Instruction Partners reported mixed results from an analysis of 20 AI learning tools across 16 school systems, with three findings a parent can act on: general-purpose chatbots posed the biggest risks, purpose-built tools for writing feedback and math showed more promise, and none of them was ready to teach independently. That last clause is the one to tape to the fridge. Every tool in the set needed an adult.
Key Takeaways
- No tool was ready to teach on its own. Not the chatbots, not the specialists. The finding is about supervision, not about software quality.
- Purpose-built beat general-purpose. Tools designed for one job, writing feedback or math practice, outperformed open chatbots on both promise and risk.
- The review covered 20 tools across 16 school systems, which makes it a landscape map rather than a controlled trial. Treat it accordingly.
- Instruction Partners’ own published work frames these as “learning conversations,” explicitly not endorsements, and the project is ongoing with new entries added monthly.
- This matches the experimental literature almost exactly: a purpose-built tutor beat classroom instruction in one trial, while an unguarded chatbot reduced retained skill in another.
What the review was, and what it was not
A landscape review is a structured observation of tools in real settings, designed to describe variation rather than to measure an average effect.
The AI-in-education policy tracker entry dated September 14, 2026 summarises the result: an analysis across 16 school systems found general-purpose chatbots posed the biggest risks, purpose-built tools for writing feedback and math showed more promise, and none was ready for independent teaching.
Instruction Partners describes itself as a nonprofit working to change “the way that schools, school systems, and states develop and support instructional leaders,” with a focus on English language arts, early literacy and math, and particular attention to underserved students. Its public AI in Action Learning Tour page carries video conversations with ed-tech developers, with 13 products profiled at the time of writing, including Brisk Teaching, Coursemojo, Enlighten AI, Goblins and Google Gemini. The page states plainly that the videos “are not endorsements; they are learning conversations, intended to help us all understand the range of instructional jobs developers are building products to do,” and describes the project as ongoing with new conversations added monthly.
So be precise about what you are reading. This is not a randomised controlled trial, and the organisation that produced it says so. It is practitioners watching tools work in classrooms and describing patterns. For a parent trying to decide which questions to ask at a conference, that is arguably more useful than an effect size. For a district deciding whether to spend $400,000, it is not sufficient on its own.
Why purpose-built tools keep winning
The mechanism behind this finding is not mysterious, and it is worth understanding because it predicts the next five years of school software.
A general-purpose chatbot is optimised to produce a satisfying response to any prompt. That objective is in direct tension with teaching, because the most satisfying response to “solve this for me” is to solve it. The model has no stake in whether the child can do it next week. It has no curriculum, no sequence, no record of what the child got wrong on Tuesday, and no instruction to withhold an answer.
A purpose-built tool constrains the model. It defines a narrow job, writing feedback on a rubric, or practice problems within a math progression, and it wraps the model in rules: do not give the final answer, ask a diagnostic question first, flag this response pattern to the teacher. Those wrappers are where the pedagogy lives. The model is the engine; the wrapper is the car.
This is exactly the distinction the experimental literature found. Bastani and colleagues, publishing in PNAS in 2025, ran a large field experiment with high school math students: GPT-4 tutoring improved performance during practice but produced reduced skill development once access was removed, and that harm was mitigated when the tutor was configured to give teacher-designed hints instead of answers. Kestin and colleagues, publishing in Scientific Reports in 2025, built an AI physics tutor on the same pedagogical principles used in their active-learning classroom and found students “learn significantly more in less time” than peers in the in-class condition.
Same underlying technology. Opposite results. The variable was the wrapper.
| Tool type | What it is optimised for | Review finding | The question to ask school |
|---|---|---|---|
| General-purpose chatbot | A satisfying answer to any prompt | Biggest risks of the categories reviewed | ”Is this the consumer version or a configured school version?” |
| Purpose-built writing feedback | Comments against a rubric | More promise | ”Does the teacher see the feedback before my child does?” |
| Purpose-built math practice | Problems inside a skill progression | More promise | ”Does it withhold the answer, and does it report what my child missed?” |
| Any of the above, unsupervised | Nothing in particular | None ready for independent teaching | ”Who is in the room while this runs?" |
"Not ready for independent teaching” is the operative phrase
This is a finding about staffing, and it collides with the main reason districts buy AI.
The pitch for AI in schools is usually capacity: too many students, too few adults, too little time. A tool that requires an adult present does not solve the capacity problem. It changes what the adult does, which may be worth paying for, but it is a different product than the one in the sales deck.
For a parent, the practical translation is a scheduling question. If your child’s class uses an AI writing tool while the teacher circulates and reviews flagged responses, the setup matches what the review found promising. If your child uses the same tool in a computer lab supervised by a rotating aide, or at home as assigned practice, the adult is not in the loop and the review’s warning applies directly.
There is also a quieter implication for feedback. A tool that generates comments on an essay is producing something a child reads as authoritative. If no teacher reviews those comments first, the child is being taught by the tool, which the review says none of them is ready to do. Ask about the review step specifically. It is usually a setting, and it is usually off by default because it costs teacher time.
What to do at home
Get the tool’s actual name and tier
“We use AI for writing” tells you nothing. “We use Brisk Teaching on the district licence, teacher-reviewed” tells you everything. Ask for the product name and whether it is a school-configured version or a consumer account. The difference between those two is the difference between the two arms of the PNAS experiment.
Ask whether the answer is withheld
For anything math-adjacent, this is the single highest-signal question. A tool that gives the final answer when asked is operating in the condition that harmed retention. A tool configured to give hints designed by a teacher is operating in the condition that did not. Teachers generally know which setting is on.
Watch the transfer, not the output
The PNAS finding was that performance looked fine with the tool and worse without it. You can run that test at home in ten minutes. Have your child do three practice problems with the tool, then three similar problems with the device closed. If the second set collapses, you have found the illusion the researchers found, and you have found it in your own kitchen.
Treat writing feedback as a draft, not a verdict
AI comments on an essay are a hypothesis about the essay. Read them with your child and ask which ones they agree with. Disagreeing with a machine, out loud, with a reason, is a skill. Our guide to AI homework tools and when they help or hurt has more on separating help from substitution.
What not to do
Do not ask the school to ban the category. The review’s finding was not “AI tools are bad,” it was “these tools are not ready to run alone.” A ban removes the configured, supervised version and leaves the unconfigured consumer version your child can reach on a phone, which is the version that performed worst. Ask for supervision and configuration instead. Our comparison of SchoolAI, Khanmigo and Class Companion shows how differently purpose-built tools can be set up.
What to Watch For Over the Next 3 Months
- Week 4: Look for new entries on the learning tour. The project adds conversations monthly, and a tool your district uses may appear with a developer explaining what job it was designed for.
- Month 2 red flags: A school describing a consumer chatbot as an “AI tutor.” Writing feedback that reaches students without teacher review. Any AI practice assigned as homework with no adult in the loop, which is the exact configuration the review warned about.
- Month 3 self-check: Do a closed-device check on one skill your child has been practising with AI help. If performance drops sharply without the tool, raise it with the teacher using that specific observation rather than a general worry.
Frequently Asked Questions
Which AI learning tool came out on top?
None did, and that was the point. The review reported mixed results across 20 tools in 16 school systems, found purpose-built writing-feedback and math tools more promising than general-purpose chatbots, and concluded that none was ready to teach independently.
Is this a scientific study?
It is a practitioner landscape review, not a controlled trial. Instruction Partners frames its published tool conversations as learning conversations rather than endorsements, and the project is ongoing. Use it to generate questions, and use experimental studies to settle effect claims.
Why are general-purpose chatbots riskier?
Because they are optimised to give a satisfying answer to any prompt, which conflicts with teaching. They hold no curriculum, no record of what a child missed, and no instruction to withhold an answer unless someone configures one.
What does “not ready for independent teaching” mean for homework?
It means AI practice assigned as unsupervised homework is the configuration the review cautions against. If the tool is used, the useful version has a teacher reviewing its output before or alongside the child.
How do I know if my child’s tool is configured properly?
Ask two questions: does it withhold final answers, and does a teacher see its feedback first? Both are settings. A school that cannot answer has probably not changed the defaults, and defaults favour answering.
Does any of this apply to the chatbot my child uses at home?
Directly. A consumer chatbot at home is the least configured version of the least promising category in the review. That is not a reason to forbid it, but it is a reason to sit beside your child the first few times and watch what it does when asked to just give the answer.
About the author
Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.
Sources
- Pursuit. (2026). “AI in Education: News, Policies, Innovations” (entry dated September 14, 2026). https://www.pursuit.us/news/ai-in-education-news-policies-innovations
- Instruction Partners. “AI in Action Learning Tour.” https://instructionpartners.org/ai-in-action/
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). “Generative AI without guardrails can harm learning: Evidence from high school mathematics.” Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.2422633122
- Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). “AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting.” Scientific Reports. https://www.nature.com/articles/s41598-025-97652-6
- Nickow, A., Oreopoulos, P., & Quan, V. (2020, July). “The Impressive Effects of Tutoring on PreK-12 Learning: A Systematic Review and Meta-Analysis of the Experimental Evidence.” NBER Working Paper 27476. https://www.nber.org/papers/w27476
- Sidoti, O., Park, E., & Gottfried, J. (2025, January 15). “About a quarter of U.S. teens have used ChatGPT for schoolwork — double the share in 2023.” Pew Research Center. https://www.pewresearch.org/short-reads/2025/01/15/about-a-quarter-of-us-teens-have-used-chatgpt-for-schoolwork-double-the-share-in-2023/
- UNESCO. (2024). “AI Competency Framework for Students.” https://www.unesco.org/en/articles/ai-competency-framework-students
