In studies LLM-aided coding is [not] [having] a [great] [eff...

In studies LLM-aided coding is not having a great effect on the quality or safety of the code. (Setback.) Sure, they’re more “productive” in that they’re writing more lines of code, but the industry discovered decades ago that measuring coding productivity this way is a very bad idea as it inherently promotes poor coding. (Setback.)

Defenders say that in actual practice nobody actually commits LLM-generated code directly so this is all just scare-mongering, except – because we, as an industry, is so utterly foolish as to commit completely to a Microsoft-owned development environment through and through (Setback) – Microsoft can track the code generated by Copilot, through Visual Studio Code, all the way to being committed in a GitHub repository, and it turns out that at least 40% of the code generated is committed unmodified.

Developers copy more than they understand.