Scratch Add-ons AI

Scratch AI Extensions give the child complete authorship over their project. The child decides what AI will do, collects their own training data, trains the model, and programs how it behaves within their Scratch project. No prescribed path exists. Each project is the child's original creation with AI capabilities they designed.

Training ML models involves genuine difficulty. Models don't always classify correctly, training data needs iteration, and debugging requires sustained effort across multiple attempts. But difficulty is self-selected. Simple projects (Face Sensing with one sprite) can be completed quickly. Complex projects (multi-class image classifier integrated into a game) require real persistence.

ML model training is inherently metacognitive. When a model fails, the child must diagnose why and try a different approach: more training examples, different categories, different features. Each new project requires adapting prior knowledge to a new domain. The child constantly reflects on how the AI learns, not just what it produces.

Scratch AI Extensions create genuine "why did it do that?" moments. Training a model to recognize cats and watching it misclassify a dog creates an information gap the child wants to close. The exploration space is vast: image recognition, text classification, gesture detection, language processing. Depth is available for kids who pursue it.

AI extensions add a new creative dimension to Scratch's already-creative environment. The child imagines AI-powered projects and builds them from scratch: gesture-controlled games, emotion-detecting art, voice-activated stories. Creative risk is constant because model behavior is unpredictable. Revision is natural to the workflow as projects evolve.

Training ML models exercises a specific type of judgment: Is my training data representative? Is the model accurate enough? Am I introducing bias? MIT's RAISE curriculum includes an AI ethics component. But ethical reasoning and real-world tradeoff evaluation aren't central to the tool experience.

Solo coding activity. Scratch's community sharing platform exists, but social interaction during AI project creation depends on classroom or club context, not the tools themselves.

Coding frustration is real but not scaffolded. No emotion labeling, coping strategies, or regulation features.

AI projects can connect to genuine personal interests when a child builds something that matters to them. Scratch's sharing community lets children contribute projects others can learn from and remix. But purpose engagement is emergent, not designed into the tools.