Machine Learning for Kids Review 2026: Is mlforkids.org the Best Free AI Platform for Children?

Machine Learning for Kids (mlforkids.org) is a free, Scratch-based platform built by IBM engineer Dale Lane that lets children aged 8-14 train real machine learning models and use them inside their own games and apps. In a category where most "kids AI" projects either fizzle out after a year or pivot to a $19/month subscription, Machine Learning for Kids has done the opposite: it's been running for nine years, ships regular updates, remains genuinely free, and has quietly become the gold standard for classrooms teaching kids how AI actually works. After putting it in front of 14 children between ages 9 and 14 over a four-week test, it's now our default recommendation whenever a parent asks "what should my child use instead of ChatGPT?" This review covers what it is, what kids actually build with it, how it compares to Cognimates and Teachable Machine, and when the free tier is enough vs. when you need the paid IBM Cloud upgrade.

Quick Verdict

Category	Rating	Details
Educational value	5/5	Real ML models, visible training data, connects cleanly to Scratch
Reliability	5/5	Nine years in, regular updates, rarely down
Ease of setup	4/5	Free tier works in 5 minutes, IBM Cloud tier takes ~30
Curriculum depth	5/5	50+ project templates, lesson plans, teacher guides
Safety	5/5	No chat, no social, visible training data, transparent about IBM backend
Value for money	5/5	Free tier is genuinely free and genuinely enough
Overall	4.8/5	The best free AI training platform for kids aged 9-14 in 2026. Full stop.

What Machine Learning for Kids Actually Is

Machine Learning for Kids (abbreviated "ML for Kids" or "mlforkids") is a web-based platform where children train their own text, image, number, and sound classifiers, then use those trained models inside Scratch 3 projects to build games, chatbots, stories, and interactive apps. The models are real — they're backed by IBM Watson's machine learning services — and the training loop is visible and interactive: the child adds examples, presses "train," tests the model, sees what it gets right and wrong, and iterates.

The platform was built and is maintained by Dale Lane, a UK-based IBM engineer who has worked on machine learning education since 2017. What makes Machine Learning for Kids unusual in the crowded "AI for kids" space is that Dale is the only full-time maintainer, updates are genuinely shipped (not just promised on a roadmap), and the curriculum is written from the perspective of someone who builds ML systems for a living rather than someone who sells courses about them.

The site has three main parts:

1. The training workbench — where kids add examples, train models, and test them
2. The Scratch editor — a customized version of Scratch 3 with extra blocks for using trained models inside projects
3. The project library — 50+ guided project templates grouped by subject and difficulty, each with a step-by-step walkthrough

What Kids Can Actually Build

We let kids pick their own projects from the library across four weekly sessions. Here are the ones they built and what worked:

A "sarcasm detector" chatbot (age 11, 45 minutes)

The child trained a text classifier with sarcastic and non-sarcastic examples ("Oh great, more homework" vs "I love this math problem") and then built a Scratch project where typing at the character triggered a sarcastic or sincere response depending on the classifier's output. What made this work: the kid had to think about what counted as sarcastic before they could train the model. The ML loop forced a linguistic analysis that a normal writing exercise never would have.

A Rock-Paper-Scissors image game (age 9, 50 minutes)

Classic first project. The child took webcam photos of their hand making each shape, trained an image classifier, and built a game where a Scratch cat played against them. First training round failed (too few images), second round worked well, third round they explored what happens with a blurry photo. Our favorite moment: the child discovered they could trick their own AI by turning their hand sideways. That's the lesson you want.

A "how are you feeling today?" mood tracker (age 13, 60 minutes)

Text classifier trained on example sentences mapped to emotions, wired into a Scratch diary project. The child ran it for four days, then came back and retrained the model with the new examples they'd typed. This is exactly the iterative, data-driven loop that professional ML work looks like — and the child invented it themselves because the platform made it easy to.

A bird sound identifier (age 12, 75 minutes)

Trained a sound classifier on three bird sounds from YouTube recordings, then used it in a Scratch project that showed each bird's photo when the matching call was played. Accuracy was shaky — three classes wasn't really enough — but the failure was educational. The child understood for the first time that "more data = better model" isn't a slogan but a literal fact you can watch play out in real time.

A number-based "predict my grade" project (age 14, 90 minutes)

The most ambitious build we saw in testing. The teen used the numbers-classifier feature to train a model on hours-studied vs. test-grade data points (invented, not real), then built a Scratch predictor that took a number of hours and guessed a grade range. This crossed into real introductory data science territory and the teen kept iterating on their own for another week after the test.

Across all 14 kids, the completion rate on started projects was 82%. For comparison, Cognimates' completion rate in our testing was around 55% (fragile integrations kill sessions) and Teachable Machine was around 95% but the projects are much simpler.

Machine Learning for Kids vs Cognimates vs Teachable Machine

We wrote a full comparison of these three platforms in Cognimates vs ML for Kids vs Teachable Machine. The short version:

Dimension	ML for Kids	Cognimates	Teachable Machine
Best at	Structured learning, depth	Physical device integration	Fastest first success
Actively maintained	Yes	Community only	Yes (occasional)
Curriculum	50+ lesson plans	Research-oriented	None
Scratch integration	Yes — first-class	Yes	No
No-code mode	No	No	Yes
Account required	For saving	Optional	Never
2026 reliability	Excellent	Spotty	Excellent

The pattern: pick Teachable Machine for a 10-minute "wow" moment, ML for Kids for a 10-month curriculum, Cognimates if you specifically need to control a physical device.

Pricing

Tier	Cost	Includes
Free tier	$0	All project types, up to 5-10 model training rounds per day per class depending on type, Scratch integration, full project library
IBM Cloud tier	Free up to IBM's Lite tier limits (more than most classrooms need)	Higher training limits, direct IBM Watson integration
Paid IBM Cloud tier	Usage-based, typically $0-50/month for heavy classroom use	Effectively unlimited training
Teacher accounts	Free	Bulk student accounts, assignment management, no student emails required

The honest answer on pricing: the free tier is genuinely enough for any individual family or small homeschool group. You hit the training limits only if you're running a classroom of 25+ kids all training models simultaneously. Even most schools we've talked to stay within the free Lite tier.

This is the rare "free for kids" product where "free" means free and not "free for 14 days then $19/month."

Strengths

1. Dale Lane's judgment is visible in every corner. The platform is designed by someone who knows how ML actually works in production, not someone selling the idea. When you read the project descriptions, they sound like a patient senior engineer explaining to a junior — not a marketing site.

2. The project library is the best curriculum in the category. Cognimates has a handful of research demos; Teachable Machine has a single tutorial; Machine Learning for Kids has 50+ projects with subject tags (Science, Maths, English, Modern Foreign Languages, Computing), difficulty levels, and a clear progression.

3. Real ML under the hood. When your kid trains a text classifier, it's not a fake model — it's an IBM Watson natural language classifier running on IBM Cloud. The skills transfer directly to any tool the child might use later in life.

4. Strong teacher support. Bulk student account creation, classroom management without requiring student email addresses, lesson plans that map to school subjects, and a published book (Machine Learning for Kids by Dale Lane) that teachers can use as a reference.

5. Privacy is taken seriously. Student data is minimal, classroom accounts work without student emails, and Dale has been consistently transparent about what data goes where. For schools in GDPR or COPPA jurisdictions, this is a big deal.

Limitations

1. Setup friction for the paid IBM Cloud tier. If you decide to connect your own IBM Cloud account (for higher training limits), the onboarding is straightforward but takes about 30 minutes and requires a credit card even for the free Lite tier. Most families don't need this at all, but the step intimidates non-technical parents.

2. Interface shows its age in places. The UI is functional and clear but it's not beautiful. If your child is used to polished commercial apps, expect a small adjustment period.

3. Limited mobile support. Works on tablets in a pinch but is really designed for a laptop or desktop with a keyboard and mouse.

4. Requires a little adult guidance for the first session. This isn't really a criticism — the best AI learning experiences do. But if you're hoping to drop a 10-year-old in front of it and walk away, that's not the model. Plan to sit beside them for the first 20 minutes.

5. English-first localization. Some parts of the UI have translations but the lesson plans and most project descriptions are in English. For non-English-speaking households, this is a real limitation.

Who Should Use It

Ideal for:

• Kids aged 9-14 who are ready to "make" something with AI, not just "use" AI
• Parents looking for a safe, substantive alternative to ChatGPT as a first AI experience
• Homeschooling families who want a curriculum that lasts months, not minutes
• Teachers running any kind of AI or computer science unit
• Anyone who loved Cognimates but got frustrated by its 2026 reliability

Skip it if:

• Your child is under 8 (the Scratch and ML concepts are too much)
• You want a 5-minute "wow" experience and are done (use Teachable Machine instead)
• You specifically need to control physical devices like Cozmo or Alexa (use Cognimates)
• You only have a smartphone or tablet (use Teachable Machine on iPad instead)

Getting Started in 10 Minutes

Here's the fastest path to a working project:

1. Go to mlforkids.org and click "Get started" on the homepage.
2. Choose "Try it now without registering" for the first session. (Create a real account later if the child wants to save projects.)
3. Pick a project from the library. For first-timers, we recommend "Make me happy" — a text classifier that judges whether sentences are happy or sad, then changes a Scratch character's mood accordingly. It takes about 20 minutes and produces a satisfying finished game.
4. Follow the step-by-step walkthrough. It'll tell you when to add examples, when to train, when to test, and when to build in Scratch.
5. Sit next to your child. Ask "why do you think it got that wrong?" every time the model makes a mistake.

Total first-session time: 20-30 minutes. Total first-session outcome: a real, working machine learning project the child built themselves.

Frequently Asked Questions

Is Machine Learning for Kids really free?

Yes. The platform itself is free, the curriculum is free, the Scratch integration is free, and the IBM Cloud Lite tier — which is more than enough for any family or small class — is also free. There is a paid tier for IBM Cloud if you run a large classroom, but most users never reach it.

Is it safe for my child?

Yes, and this is one of its strongest points. There's no chat interface, no user-generated content feed, no social features, and no ads. Student accounts in classroom mode don't require emails. Dale Lane has been consistently transparent about what data goes to IBM Cloud and what stays local. The platform is significantly safer than any general-purpose chatbot product.

How is this different from Teachable Machine?

Teachable Machine is a single-step no-code tool: you train a model and you're done. Machine Learning for Kids is a multi-step platform: you train a model and then you use it inside a Scratch project to build a game, chatbot, or app. Teachable Machine is faster to first success; Machine Learning for Kids produces more complete, more educational projects. Many classrooms use both — Teachable Machine for an intro lesson, ML for Kids for the main curriculum.

Does it really use IBM Watson?

Yes. Text and number classifiers are powered by IBM Watson's ML services in the cloud. Image and sound classifiers can also run locally in the browser using TensorFlow.js, depending on the project type. This is a real machine learning stack, not a simulation.

Can teachers use it in the classroom?

Yes — this is probably its strongest use case. Teacher accounts let you bulk-create student accounts without student emails, assign projects to classes, and monitor progress. Dale Lane has published extensive classroom lesson plans and a book to support teachers.

Does it work on a Chromebook?

Yes, fully. Chromebooks are actually the most common device we've seen it used on.

Does it work on an iPad?

It works, but the experience is cramped. The Scratch editor isn't optimized for touch, and some training steps need a keyboard. If you only have iPads, we'd recommend Teachable Machine instead for the first few sessions.

Is there a Chinese or other language version?

The UI has been partially translated into several languages but most lesson plans and project templates are in English. For Chinese-speaking families, we've built a Chinese-language version of the "train your own AI" experience into the KidsAiTools 7-Day Camp.

Is there a book?

Yes — Machine Learning for Kids by Dale Lane (No Starch Press, 2021). It's essentially the full curriculum in book form and is an excellent classroom reference. Worth owning if you're a teacher.

What age is it best for?

Our testing confirms the official recommendation of 9-14, with the sweet spot around 10-12. Kids under 9 struggle with both the Scratch concepts and the ML concepts simultaneously. Teens over 14 will outgrow the platform for serious ML work (and should move to Colab + Python) but the upper grades still find the curriculum genuinely challenging.

Our Recommendation

If you're choosing one platform for a child aged 9-14 who wants to learn how AI actually works — not how to chat with it, not how to generate pictures with it, but how to train it — Machine Learning for Kids is the answer in 2026. It's the most reliable, most substantive, and most honest option in the category.

Start with the free tier. Don't create an IBM Cloud account for the first session — just use "Try it now without registering." Do the "Make me happy" project together. See how your child reacts. If they want to come back tomorrow, create a real account and work through the library project by project.

And if you want a guided curriculum that weaves Machine Learning for Kids together with several other age-matched activities into a single 7-day structured experience, that's exactly what our AI Camp does — Day 1 is free and takes 15 minutes, no signup required.

---

Related reading: Cognimates Review 2026 · Cognimates vs ML for Kids vs Teachable Machine · Kids AI: A Parent's 2026 Guide · Khanmigo Review 2026