Back to index
OtherMinutes to train a first model; a few hours to explore image, audio and pose projects properly·Free, no account needed

Teachable Machine (Google)

4.3

The quickest way I know to give someone a gut feel for how supervised learning works, including how it fails. It is a tool rather than a course, and it teaches almost nothing about what happens under the hood, but as a first hour with machine learning it is hard to beat.

What We Liked

  • Train a working classifier in a few minutes in the browser with no code and no sign-up
  • Training runs locally in the browser, so webcam samples are not sent to Google unless you choose to save the project
  • Makes the effect of bad or unbalanced training data obvious, which is the most important lesson for beginners
  • Exports to TensorFlow.js, TensorFlow Lite and Keras, so curious students can carry a model into a real project
  • Free, and works on school Chromebooks, which is why so many teachers build lessons around it

What Could Be Better

  • Not a course: there is no curriculum, explanation of the maths or assessment
  • Only image, audio and pose classification, so nothing on language models or generative AI
  • Built on transfer learning that it barely explains, which can leave learners thinking training is magic
  • Has seen little visible development in years, and older export paths can trip up current TensorFlow versions

Detailed review

Teachable Machine is a free web tool from Google Creative Lab. You make a few classes, record examples with your webcam or microphone or upload files, press train, and within seconds you have a model that recognises your hand gestures, your cat or a clap. Strictly speaking it is not a course, but it gets searched for and used like one, especially in classrooms, so it deserves a place here. What it does brilliantly is make the core idea of supervised learning physical.

You can show it twenty photos of yourself in a hat and twenty without, watch it work, then turn on a lamp and watch it fall apart. That five-minute experience teaches more about training data, lighting, class imbalance and overfitting than an hour of slides. When I have used it with non-technical groups, it is the moment the room stops thinking of AI as either magic or useless. The privacy design is good too.

Training happens in the browser, so webcam samples stay on the device unless you save the project to Google Drive, which matters a lot for schools. Exports to TensorFlow.js, TensorFlow Lite and Keras mean a motivated student can take a model into a website, a phone app or an Arduino project. The limits are just as clear. There is no explanation of what is going on underneath.

It uses transfer learning on a pretrained network, and if nobody tells learners that, they leave with a slightly wrong mental model of how much data real systems need. It covers classification only, so it says nothing about the language models that dominate what people mean by AI today. And it has the look of a project Google is keeping alive rather than investing in, with little visible change in a long time. My recommendation is to treat it as a hands-on demo inside a bigger lesson.

Pair it with Elements of AI, AI for Everyone or a school curriculum like Day of AI, and use the failure cases to talk about bias. Used that way it is one of the best free teaching tools in AI education.

[ final ]

The verdict.

Use it as the first hour of any intro to AI, especially with kids or non-technical teams, then move on to a real course. Pair it with a lesson on bias and data quality and it does more than many paid beginner courses.