I disagree. I think that GitHub is training Copilot on only repos currently on GitHub, because they can easily do so due to the Terms of Service [1] which allow them to "parse [user content] into a search index or otherwise analyze [user content] on our servers". They can reasonably show that an ML model fits the definition of "otherwise analyze content", regardless of the license.
If code that was never hosted on GitHub to begin with starts showing up verbatim in Copilot suggestions, it might be a more legally challenging position for them. It is unfortunate grey area if people decide to upload copies of free software repos to GitHub anyway, but perhaps future versions of the GPL or other free/libre licenses could be explicit about usage in training ML models.
I have no idea what Amazon or Salesforce are doing, but it would be interesting to hear what they are using for training data and how they justify compliance with the software licenses.
----------------------------------------
[1] https://docs.github.com/en/site-policy/github-terms/github-t...