The “AI tutor for kids” category went from speculative to crowded in the past three years. Khanmigo from Khan Academy launched as the first major one. Schmoozy, Question.AI, MagicSchool’s student-facing tools, the Duolingo Max conversational practice, and a dozen smaller startups have followed. Some of these are genuinely useful for specific purposes. Some are not, and not because they don’t work technically, but because the use case they’re solving doesn’t actually need an AI to solve it.

Worth being specific about what makes an AI tutor useful for a kid, what makes one a waste of time, and where the responsibility sits between the tool and the adult overseeing the kid’s learning.

The use cases where AI tutors earn their place

A few patterns where these tools have produced clear value.

Math practice with Socratic-style hints rather than answers. The AI tutor that asks “what would you try first?” instead of giving the solution is meaningfully different from a static answer key. Khan Academy’s Khanmigo has moved aggressively in this direction; the Khanmigo math sessions function as a patient one-on-one helper that doesn’t get tired of the same kid asking the same question seven times. For motivated students working through hard problems, this is real.

Reading comprehension with adjustable depth. An AI that can read a passage with a kid and ask progressively harder questions about it, while staying in the kid’s zone of proximal development, is closer to what a one-on-one reading tutor does than what an app can usually provide. The Diffit tools and a handful of others have built around this.

Foreign language conversation practice. Duolingo Max and Speak both let kids practice conversational language with an AI partner who never gets impatient and won’t judge mistakes. For language learning specifically, the willingness to make mistakes in front of the tutor is half the work, and the AI version of that is genuinely freeing for many kids.

Writing feedback at the diagnostic level. An AI that reads a kid’s draft and points out structural problems (your second paragraph contradicts your first, your conclusion isn’t supported by the body, your evidence is thin in section three) without rewriting the work is a useful complement to a teacher’s or parent’s feedback. The AI doesn’t always get it right, but it gets it right often enough to be a reasonable first reader for revision purposes.

The use cases where AI tutors don’t help

Several patterns where the marketing exceeds the value.

“Personalized learning” as a substitute for instruction. The AI tutor as the entire teaching mechanism, where the kid sits at a screen and the tool guides them through a curriculum, is an old idea (computer-assisted instruction since the 1970s) wearing new clothes. The format has consistently underperformed expectations across decades, and AI hasn’t fundamentally changed what was hard about it.

Early elementary anything. A 6-year-old learning to read needs an adult reading aloud, asking questions, modeling reading. An AI tutor for a 6-year-old is a screen, and screens at that age are mostly displacement. The case against more elementary screens applies particularly hard to AI tools marketed for younger kids.

Conceptual subjects where the kid hasn’t yet seen the concept. AI tutors are good at responding to a kid’s questions about a topic the kid has some grasp on. They are weaker at introducing a concept from scratch, where the explanation needs to land in a particular order, with particular examples, calibrated to a particular kid’s prior knowledge. A book or a video or a teacher introduces a concept; the AI tutor helps practice it.

Anything where the kid uses the AI to evade the cognitive work. The classic version: kid asks the tutor for help, gets a complete answer, copies it, moves on. The AI didn’t tutor; it produced. This is a usage problem more than a tool problem, but the tools that don’t actively guard against it tend to enable it.

The supervision question

The single most important variable in whether an AI tutor helps a kid is the adult oversight wrapping around it. The pattern works like this: the kid sits down with the tool, has a learning session, and then talks with a parent or teacher about what they did. The conversation is where the learning consolidates; the tool is the input.

Without that conversation, the tool’s value drops. The kid stays inside the AI’s framing, doesn’t notice when the AI made an error, doesn’t cross-reference the explanation against other sources, and accumulates a fragile understanding that hasn’t been tested.

This is true of any one-on-one tutor, AI or human. The difference is that a human tutor naturally builds a small social relationship that creates the post-session conversation; an AI doesn’t, so the parent or teacher has to do that work explicitly.

Specific tools, briefly

A short, opinionated take on the major options.

Khanmigo (Khan Academy). The most ambitious and the most carefully designed. The Socratic-style guardrails are real, the integration with Khan’s existing curriculum is the cleanest in the category, and the privacy posture (school-mediated, with parental controls) is reasonable. Best fit: kids in grades 4 and up using Khan Academy as part of their math or science practice.

Magic School AI student-facing tools. Built into the same ecosystem as the teacher-facing lesson-planning tools; the student tools are guided by the teacher’s curriculum. Best fit: classroom contexts where a teacher is wrapping the AI use into a defined lesson rather than letting the kid roam.

Duolingo Max conversational practice. Specific use case (foreign language practice) and meaningfully better than the no-AI alternatives. Best fit: language learners at any age who need conversation practice and don’t have a human partner available.

Brisk Teaching’s student-facing features. Tighter integration with Google Classroom; useful for the specific contexts where teachers are already using the Brisk teacher tools. Best fit: classrooms running Google Workspace where the teacher is in the loop on what the AI is doing.

The pure-chat AI tutors (ChatGPT, Claude, the consumer-grade tools). These work for self-directed older kids who can guide their own learning and have the metacognition to evaluate the AI’s responses. They are the wrong tool for younger kids who don’t have those skills yet, even though they are the most accessible.

What we’d tell a parent considering an AI tutor

Three rules of thumb.

Use it as supplement, not substitute. The AI tutor is one of several inputs into the kid’s learning, alongside a teacher, a parent, books, peers, and unstructured practice. Setting up the AI tutor as the primary teacher of any subject is the wrong setup, regardless of how engaging the tool is.

Sit with the kid for the first several sessions. Find out what the kid asks the AI, what the AI says back, where the AI gets it wrong, and what the kid does with the answers. The tool’s value depends on how the kid uses it, and you can only see that by watching.

Build the post-session conversation. Ask the kid what they learned, what was confusing, what the AI said that they’re not sure about. The conversation is where the actual learning consolidates. Without it, the tool is doing less work than it appears to be doing.

The category will keep evolving. Some current tools will improve and others will fade. The framework for evaluating any of them is the same as the framework for evaluating older edtech: does it help a kid develop real skill, with appropriate adult support, and in proportion to the time it consumes? Most don’t. The few that do are worth using carefully.