GitHub Copilot: what owners actually say

Multiple owners cite ongoing litigation against Microsoft, GitHub, and OpenAI. One user notes verbatim code from training data appears ~0.1% of the time, raising GPL and derivative-work concerns. A Stack Exchange answer explicitly says they couldn't recommend AI code generation in a professional context due to unresolved legal questions.

Owners report it generating code with security issues like unencoded query parameters, and note that if you glance at suggestions without careful review you can overlook edge cases. One commenter says 'if you need to go through the suggested code to ensure it's correct, you may as well write it yourself.'

An owner points out that for less-represented combinations like a Rust game using the hecs ECS library, there's almost no training data, so Copilot has nothing useful to offer. It works best where many similar examples exist in its training set.

Numerous Stack Exchange answers walk through disabling auto-suggestions, with settings changing across versions and frequently deprecating — at least four different settings keys have been used over time. Owners want on-demand-only suggestions but find the defaults pushy.

Owners observed Copilot inserting TODO comments with specific developers' names (e.g., 'TODO (Linus Torvalds)') scraped from open source projects, demonstrating it surfaces identifiable fragments from training data rather than truly synthesizing original code.

Owners value it for reducing boilerplate and repetitive patterns, particularly where context is explicit — type annotations, variable naming, data structure scaffolding.

One owner reports successfully using it to generate solutions to errors and simple unit tests to validate fixes, calling it a useful iteration tool.

An owner compares it favorably to an apprenticeship model — it absorbs patterns from vast codebases and can apply them, even if occasional inaccuracies require vigilance.