Thesis: Copilot was almost certainly trained on a large body of code under GPL and very likely some under AGPL, making all the code it generates "derivative works" under the terms of those licenses (after all, every weight in the network has been updated using code under those licenses, so what else could the code the network generates b…
Thesis: Copilot was almost certainly trained on a large body of code under GPL and very likely some under AGPL, making all the code it generates "derivative works" under the terms of those licenses (after all, every weight in the network has been updated using code under those licenses, so what else could the code the network generates be but derivative of the inputs?) Ergo: all code generated by Copilot is covered by the GPL (almost certainly) and the AGPL (very probably.)
Whether or not this thesis is correct will have to be decided by the courts, but personally I'd love to see it litigated, and think it very likely to stand up to legal scrutiny. I don't use Copilot, and if I still ran a dev team I'd institute a rule against using it unless the company was prepared to take the risk of being required to open-source their code.
Thesis: Copilot was almost certainly trained on a large body of code under GPL and very likely some under AGPL, making all the code it generates "derivative works" under the terms of those licenses (after all, every weight in the network has been updated using code under those licenses, so what else could the code the network generates be but derivative of the inputs?) Ergo: all code generated by Copilot is covered by the GPL (almost certainly) and the AGPL (very probably.)
Whether or not this thesis is correct will have to be decided by the courts, but personally I'd love to see it litigated, and think it very likely to stand up to legal scrutiny. I don't use Copilot, and if I still ran a dev team I'd institute a rule against using it unless the company was prepared to take the risk of being required to open-source their code.