Cognimates Review 2026: Is MIT's Kids AI Platform Still Alive? (And 5 Alternatives That Work Today)

Cognimates is an open-source platform from MIT Media Lab that lets kids aged 7-14 train their own AI models, program smart devices, and build games using Scratch-style block coding. It was groundbreaking in 2017-2019 — the first real attempt to teach children not how to use AI, but how to teach it. In 2026, the picture is more complicated: the original site at cognimates.me is inconsistently reachable, the founder has moved on from MIT, and most of the momentum has shifted to community forks and newer alternatives. This review covers what Cognimates is, whether it's still usable today, what kids can actually build, and the five working alternatives we recommend if you can't get Cognimates running.

Quick Verdict

Category	Rating	Details
Educational value	4.5/5	Unique "teach AI" angle — still unmatched in philosophy
Current reliability	2/5	Original site spotty, project appears stalled since ~2020
Ease of setup	2.5/5	Requires multiple accounts and works best on desktop Chrome
Supported devices	4/5	Alexa, Cozmo, Hue Lights, Wemo — broad but some integrations are dated
Safety	5/5	No chat, no social features, research-grade design
Value for money	5/5	Free, open source, runs in browser
Overall	3.5/5	Brilliant idea that needs a caretaker. Try the community fork first; fall back to alternatives.

What Cognimates Actually Is

Cognimates was created by Stefania Druga during her graduate research at the MIT Media Lab's Personal Robots Group. Her thesis — Growing up with AI: Cognimates, from coding to teaching machines (MIT, 2018) — argued that children shouldn't just be users of AI systems; they should learn to train, critique, and collaborate with them. Cognimates was the software she built to make that possible.

Under the hood, Cognimates is a customized fork of Scratch 3, the block-based programming language built by MIT's Lifelong Kindergarten group. On top of Scratch's familiar drag-and-drop blocks, Cognimates adds:

• "Teach the Machine" models — kids can train their own image classifier or text classifier from scratch, then use the trained model inside a Scratch project
• Extensions for smart devices — Alexa, Cozmo (the Anki robot), Philips Hue lights, Belkin Wemo plugs
• Generative text and image blocks — early experiments with AI-generated content
• Vision and speech blocks — send a webcam frame or microphone audio to an ML service and get labels back

The philosophy is summarized in one of Druga's frequently-cited lines: "We want children to grow up with AI the way they grow up with reading — critically, creatively, and with agency."

A 2018 MIT study conducted with 107 children aged 7-14 across four countries found that after one month of using Cognimates, participants showed measurable improvements in their understanding of how AI systems work and were less likely to attribute human-like understanding to chatbots and voice assistants. That study is the strongest evidence base the platform has to date, and it still holds up — but it's also the last major published research from the project.

Is Cognimates Still Alive in 2026? (The Elephant in the Room)

This is the question that matters, and the honest answer is: sort of.

Here's what we found when we tried to reach the platform in April 2026:

• cognimates.me (the original site): Loads intermittently. Some days the landing page renders, other days the connection hangs. The last visible copyright is 2020.
• cognimates.mit.edu: No longer resolves to any server. The subdomain appears to have been retired.
• MIT Media Lab project page (media.mit.edu/projects/cognimates/): Still up, still describes the project, but "Last updated" stamps on the group's page indicate the core research wrapped up around 2019-2020.
• GitHub repositories: The three main repos — mitmedialab/cognimates-vm, hackidemia/cognimates-training, hackidemia/cognimates-website — are all still public. The VM repo has had only small maintenance commits since 2020. The hackidemia fork is where most of the community work has moved.
• Founder status: Stefania Druga's public profile shows she's now at the University of Chicago Booth School of Business. She continues to work on children's AI education but is no longer shepherding the original Cognimates platform day-to-day.
• Cognimates Copilot (cognimates-copilot.replit.app): A newer experiment — a ChatGPT-style assistant that helps kids write Scratch code. Separate project, different scope.

Our read: Cognimates as an MIT-hosted, actively maintained product effectively ended around 2020. What exists today is (a) the MIT research legacy, (b) an open-source codebase anyone can self-host, and (c) a community fork on GitHub Pages. It's not dead — you can still get it running — but nobody is shipping new features, fixing broken device integrations, or responding to support requests.

If that sounds harsh, remember: this is the norm for research-lab software. Cognimates did its job. It proved a thesis, inspired an entire generation of "kids teaching AI" products (including our own), and left a working open-source reference implementation behind.

What Kids Can Actually Build (When It's Working)

When Cognimates does load, here's what's genuinely still worth doing with it:

1. Train a Rock-Paper-Scissors AI that gets better over time

Kids take webcam photos of their hand signs, label them "rock," "paper," "scissors," and train a classifier directly in the browser. Then they build a Scratch game that uses the trained model to play against them. This is one of the best 45-minute intros to supervised learning we've ever seen for this age group.

2. Teach a smart light to understand new commands

Using the Hue Lights extension, kids can wire up a text classifier to control a real light bulb. "Make it cozy" → train the model to recognize this phrase → pipe the classifier output to a Hue scene. This exercise makes the abstract idea of "intent classification" tangible in a way no textbook does.

3. Build a story that reacts to drawings

The vision blocks let kids train a model to recognize objects they draw on paper, hold up to the webcam, and have the Scratch project branch the story based on what's drawn. This bridges art and ML beautifully.

4. Program Alexa to do something new

For families that already own an Echo, Cognimates had one of the few kid-friendly ways to write new Alexa skills. The Alexa integration is the most fragile part today — Amazon's developer APIs have changed since 2019 — but when it works, it still impresses.

5. Control Cozmo the robot

Anki stopped making Cozmo in 2019, so this is only relevant if you already own one or picked one up second-hand. That said, Cognimates' Cozmo blocks remain one of the richest educational integrations the robot ever had.

Getting Started (Realistic 2026 Instructions)

If you want to try Cognimates today, here's the path that has the best chance of working:

1. Start with the community fork, not cognimates.me. Visit hackidemia.github.io/cognimates-website/home/. It's hosted on GitHub Pages, so it'll load reliably.
2. Skip the account-based features initially. The login/save flows depend on backend services that are intermittently down. Plan for ephemeral sessions: kids build, screenshot, and come back another day.
3. Use desktop Chrome on a laptop with a webcam. Mobile and tablet support was never great, and has only gotten worse.
4. Lower expectations for device integrations. The Scratch-native blocks (Teach the Machine, vision, text) work locally in the browser and are reliable. Anything that talks to a cloud service (Alexa skills, some vision APIs) is a coin flip.
5. Save the kid's models as files. Cognimates lets you download trained models — do this at the end of every session, because the online save might not be there tomorrow.

Expected setup time: 20-40 minutes for the first successful session, assuming you hit one or two dead ends along the way.

Pricing

Item	Cost	Notes
Cognimates platform	$0	Free and open source, MIT license
Community fork	$0	Self-hostable from GitHub
Hardware (optional)	$25-$300	Hue bulb $25, Wemo plug $30, Cozmo ~$200 second-hand
Your time debugging	Priceless	Budget 1-2 hours for first setup

There has never been a paid tier of Cognimates. This is refreshingly honest in an era when "free for education" usually means "free until we pivot to a $19/month subscription."

Cognimates vs Modern Alternatives

Platform	Best for	Reliability (2026)	Coding required	Cost
Cognimates	Kids who want to train AI AND control real devices	Spotty	Block-based	Free
Machine Learning for Kids	The closest spiritual successor — Scratch + real ML	Excellent	Block-based	Free (paid IBM tier optional)
Google Teachable Machine	Fastest path to "I trained an AI!"	Excellent	None	Free
Scratch 3 (with AI extensions)	Kids already comfortable in Scratch	Excellent	Block-based	Free
MIT App Inventor (with ML blocks)	Kids who want to ship a real Android app	Excellent	Block-based	Free
Cognimates Copilot (Replit)	Getting unstuck in Scratch, assisted by an LLM	Good	Block-based	Free

5 Alternatives That Definitely Work Today

If you can't get Cognimates running — or you want something under active maintenance — here are the five platforms we recommend in order of how closely they match Cognimates' philosophy:

1. Machine Learning for Kids (mlforkids.org) — The Closest Successor

Created by IBM engineer Dale Lane, this is the project we recommend most often when parents ask "what should I try instead of Cognimates?" It combines Scratch-style block coding with real IBM Watson ML models. Kids can train text, image, number, and sound classifiers. The author ships updates regularly, the site is rock-solid, and there are dozens of lesson plans for teachers. Free for basic use, with a paid IBM Cloud tier for classrooms.

Best for: Kids aged 8-14, parents or teachers who want a structured curriculum, anyone who liked the "train your own classifier" part of Cognimates.

2. Google Teachable Machine (teachablemachine.withgoogle.com) — The No-Code Path

Want a 7-year-old to train their first image classifier in 10 minutes? This is the tool. Image, sound, and pose classification, all in the browser, no account required. The catch: it doesn't include Scratch or any programmable layer — the trained model is the endpoint. That's fine for a first lesson, limiting for a second one.

Best for: First ML experience, kids aged 6-10, classroom demos, pairing with Scratch via exported models.

3. Scratch 3 (scratch.mit.edu) + AI Extensions

Regular Scratch has gained several AI-flavored extensions since Cognimates was built: speech recognition, text-to-speech, translation, video sensing, and more. They're not as deep as training your own model, but they're immediately available, work on Chromebooks and tablets, and Scratch itself has 100+ million registered users and is never going away.

Best for: Kids already comfortable in Scratch who want AI-powered projects without the setup pain of a separate platform.

4. MIT App Inventor (appinventor.mit.edu) with ML Blocks

From the same lineage as Cognimates — MIT-built, block-based, research-backed — but with a different ambition: letting kids build real Android apps. Includes extensions for image classification and personal image classifiers. The learning curve is steeper, but the payoff is enormous: your kid's trained model runs on their phone.

Best for: Kids aged 11-15, anyone who wants a project they can actually show a friend on a real phone.

5. Cognimates Copilot (cognimates-copilot.replit.app) — The LLM Helper

This is a newer, separate project that shares the Cognimates brand. It's an LLM-powered assistant that helps kids debug Scratch code and learn programming concepts by explaining errors in kid-friendly language. Not a replacement for the original platform — it doesn't let kids train AI models — but a nice companion if your kid is already using Scratch.

Best for: Kids who get stuck on Scratch projects and want a patient explainer.

Who Should (Still) Try Cognimates in 2026

Try Cognimates if:

• You want a curated, research-backed curriculum around teaching AI (not just using it)
• You own Cozmo, Alexa, or Hue devices you want to integrate
• You're a teacher building a one-off "kids teaching AI" workshop and can tolerate some setup friction
• You want to fork the open source code and build something new on top of it

Skip Cognimates and use an alternative if:

• You need something that "just works" on the first try
• You're teaching multiple kids simultaneously and can't afford troubleshooting time
• Your child is under 9 (interface is dated for younger kids)
• You don't own any of the supported smart devices and don't want to buy one

The Bigger Lesson: Why Cognimates Mattered

Even if you never use Cognimates yourself, the project left a fingerprint on every serious kids-AI platform that came after. Before Cognimates, "AI for kids" mostly meant "let kids play with pre-trained voice assistants." After Cognimates, the standard became "let kids train the assistant themselves." That shift — from consumption to authorship — is the single most important idea in children's AI education, and Cognimates is where it got named and tested in a classroom for the first time.

So the honest way to talk about Cognimates in 2026 is: a fantastic piece of research, a moderately usable piece of software, and a cultural milestone. Try it, learn from it, then graduate to Machine Learning for Kids or Teachable Machine for daily use.

Frequently Asked Questions

Is Cognimates free?

Yes. Cognimates has always been free and is released under an open source license. There has never been a paid plan, and there are no ads or in-app purchases.

What age is Cognimates for?

Officially 7-14. In practice, kids under 9 struggle with the Scratch 3 interface and the amount of on-screen text. The sweet spot is 10-13, the same range that Scratch itself works best for.

Does Cognimates work on an iPad or Chromebook?

Chromebook: yes, works in Chrome. iPad: poorly — the webcam and some extensions behave inconsistently on Safari. We recommend a laptop with a webcam as the primary device.

Do I need to create an account?

You can use most of the "train the machine" features without an account, but to save projects between sessions you need to log in. Because the login service is intermittently down, we recommend downloading your trained models as files at the end of each session.

Is cognimates.me safe for kids?

Yes — from a content perspective, Cognimates is one of the safest platforms we've reviewed. There's no chat, no user-generated content feed, no social features, no ads, and no data collection beyond what's needed to save projects. From a reliability perspective, the site is inconsistent, but not unsafe.

What's the difference between Cognimates and Machine Learning for Kids?

Both let kids train AI models inside a Scratch-style editor. Machine Learning for Kids is actively maintained, has a structured curriculum, and uses IBM Watson under the hood — it's the better choice for daily use. Cognimates is more experimental, supports physical devices like Cozmo and Alexa, and has a philosophical edge around ethics and bias that Machine Learning for Kids touches more lightly.

Can my child build real AI projects with Cognimates?

Yes — kids can train real classifiers on real data, and those classifiers control real outputs (light bulbs, robots, game mechanics). It's not a toy simulation. The models are small and simple by 2026 standards, but the learning transfers directly to more advanced ML concepts.

Is there a Cognimates app?

No. Cognimates is web-only. The closest thing to a Cognimates mobile app would be MIT App Inventor, which comes from the same research lineage and lets kids build actual Android apps with AI features.

Our Recommendation

If you're a parent or teacher who specifically wants to use Cognimates because you read about it in a magazine or saw the MIT TED talk: try the community fork at hackidemia.github.io first, not cognimates.me. Spend one afternoon with it, do the Rock-Paper-Scissors project, and decide from there.

If you want a platform that captures Cognimates' philosophy but actually works reliably in 2026: use Machine Learning for Kids. It's the closest thing to "Cognimates, but maintained."

If you want the fastest possible first AI experience for a young kid: use Google Teachable Machine. Five minutes, zero setup, no account.

And if you want a guided path that weaves several of these tools together into a single curriculum for ages 6-15, that's exactly what we've built into the KidsAiTools 7-Day AI Camp — try Day 1 for free, no signup needed.

---

Related reading: Khanmigo Review 2026 · Google AI Tools for Kids · ChatGPT Prompts for Kids by Age