Nonprofit Tested 20 AI Learning Tools. None Was Ready to Teach.

What the review found

The project is called the AI Learning Tour, run by Instruction Partners, a US nonprofit that coaches school districts and states on curriculum and instruction. Led by co-founder and CEO Emily Freitag, the team evaluated 20 AI learning tools and produced 16 deep-dive product profiles, built from classroom observation and interviews across 16 school systems, talking to teachers, students, district and building leaders, and the tool developers themselves. It plans to review at least 20 more in the coming school year.

The team sorted the tools into three kinds: general-purpose chatbots like ChatGPT, Claude, and Gemini; education-focused tools with many applications, like MagicSchool or Google Classroom; and instructional tools with a discrete purpose, like Amira Learning, which only offers reading practice. The general-purpose chatbots came out worst. Teachers reported students using them to “get out of work that requires effortful thinking,” and Freitag described watching kids game the tools outright:

“We definitely saw the biggest risks from the general-purpose chatbots. It’s also the way AI is showing up in search.”

The purpose-built tools showed more promise, with real limits. The effective ones did something specific: they could “diagnose which phonics skills students had mastered and which they still struggled with,” or in math “identify students’ misconceptions, or collect and analyze many students’ explanations.” A weakness Freitag flagged in the multi-purpose category was tools “that rewrote text at a lower reading level, which could potentially decrease rigor.” But across all three kinds, the verdict was the same:

“We did not see anything that is ready to do the pedagogical job independently. We also heard from students, especially, that they value the relationships with their teachers, and they want to learn from their teachers.”

One boundary matters for how you read the whole thing. The review was, in its own words, “focusing solely on tools’ instructional capabilities, not other concerns such as data privacy.” Freitag said that exclusion was deliberate, and that groups like Common Sense Media, Digital Promise, and ISTE should do the privacy work. Lewis Ferebee, CEO of the curriculum-evaluation nonprofit EdReports, offered a second voice on how young this whole field is: “We don’t know yet what it would mean to consistently judge quality.”

The frame the science supports

Here is the trap, and it is an easy one to fall into. A finding like this gets compressed, on the way to a parent, into a verdict on a category: AI is not ready, or AI is risky, or the opposite, that the vetted tools are the good ones. But the review does not divide the world into AI and not-AI. It divides tools by what they do with a child. The chatbots let a child offload the thinking. The promising tools made a child do a specific piece of it, and responded to the actual work. That distinction, not the technology label, is the whole story.

And it lines up with what the learning research has said for years. What separates tutoring that teaches from software that merely quizzes is granularity, whether the tool responds to each step a child takes rather than only marking the final answer right or wrong (VanLehn; Kulik and Fletcher). That is exactly the line Instruction Partners drew, between a tool that diagnoses which phonics skill is missing and a chatbot that hands over a finished answer. The same body of work draws a second hard line, between engagement and learning: time-in-app, streaks, and lessons completed measure whether a child is busy, not whether a child learned, and treating the first as proof of the second is the oldest move in ed-tech marketing. A tool that keeps a child busy is not the same as a tool that builds a skill. The OECD’s 2026 education outlook has drawn the same line independently, that finishing a task with an AI tool is not the same as learning from it.

One more word deserves a hard look, because it is on nearly every AI product box: ‘personalized.’ The popular version of personalization means matching teaching to how a child supposedly learns best, a visual learner, an auditory learner. That idea was tested and it failed, first in 2008 (Pashler and colleagues) and again in a 2024 meta-analysis, and yet a 2020 review of educators across 18 countries found nearly 9 in 10 still believe it (Newton and Salvi). Real personalization is not matching a style. It is sequencing, working out which skill a child has not built yet and which one to teach first, which is the question a label never answers. That is the logic behind a roadmap that names which skill to build first rather than sorting a child into a type.

What it means for your child

Start with the reassuring part, because the children in this review said it themselves: they value learning from their teachers, and no tool was ready to take that job. The honest read is that these tools are a supplement, not a substitute. The strongest of them free a teaching adult to do the relational work; the weakest are sold as a stand-in for one. The relationship is not the soft, replaceable part of learning. It is the load-bearing part, and it is the part you have the most leverage over at home.

So carry a sharper question into your child’s classroom, and into any app you are handed, than whether it uses AI. Ask what it asks of your child. Does it make them recall, build, and explain, or does it hand over an answer and log the minutes? Does it respond to the step they are stuck on, or reward a streak? This is the same shape as the debate over mandatory dyslexia screening: a policy or a product that sounds unambiguously good still turns entirely on how it is used, and a parent deserves to understand that decision rather than take a label for the answer. As Freitag put it, “Eventually, the field needs raters and regulators. Right now, the field needs sense-makers.” Until the raters arrive, the sense-maker in your child’s life is you.