Teachable Machine

Teachable Machine is a generative art tool. A free Google web tool for training image, sound and pose classifiers in the browser with no code. You group examples into classes, train with one click, test live, then export a TensorFlow.js model for sites, apps, Arduino or Coral hardware.

Teachable Machine is a web tool from Google that lets anyone train a small machine learning model without writing code. A project starts by choosing one of three input types: images, audio clips, or body poses. You then create classes (the categories you want the computer to tell apart) and feed each one examples, either by uploading files or by capturing them live through a webcam or microphone. A single click trains the model, and a preview panel immediately shows how it classifies new samples, so the loop from idea to result takes minutes rather than days. It is built for people who are curious about machine learning rather than people who ship production models. Teachers use it to demonstrate classification and bias in the classroom, designers and artists use it for interactive installations, and developers use it to prototype a gesture or sound trigger before committing to a larger build. The site points to teaching material from MIT Media Lab and Ready AI, plus example projects such as a cardboard sorting machine, a paper-based game controller, and an assistive communication setup. The training interface is deliberately shallow: gather examples, train, test, export. What makes it useful beyond a demo is the export path. Models are real TensorFlow.js models that run anywhere JavaScript runs, and they can be downloaded for offline use or hosted online with a shareable link. The tool also supports export to formats suited to hardware such as Coral and Arduino. Raw samples can stay on the machine: the webcam and microphone streams do not need to leave the computer, which matters for classrooms and privacy-sensitive settings. Against no-code vision platforms, Teachable Machine is narrower but freer and faster to start, with no account needed for basic training and no per-model billing. Against a full ML framework it is not a substitute: there are no options for custom architectures, large datasets, or fine-tuning. Treat it as an on-ramp and a prototyping surface, not a training pipeline.