AI code editors get smarter by learning from code they co-wrote with humans

Researchers created a new dataset for training AI models that write and edit software. This data, which includes code co-written by humans and AI, makes AI code editors perform better than models trained only on human work.

AI models learned to code by studying only human-written code. This paper shows the best way to teach AI to code is by letting it learn from code it helped write, as long as humans still approve the final changes. This means future AI coding assistants will likely become much more reliable and useful for software developers.