Yea, there are tons of public data to train on. Copilot's under fire right now: https://githubcopilotlitigation.com/, but we make sure we only train on permissively licenses because some companies are sensitive to the IP issues here.

Can you give a list of the permissive licenses you train on? MIT for example requires attribution. In fact, most permissive licenses have similar requirements.

MIT only needs attribution for redistributing "all copies or substantial portions of the Software".

Juat reading the code shouldn't be illegal, and producing a short snippet based on the code but not identical shouldn't count as a "substantial portion", but IANAL...