This is even true for e.g. MIT licenses as while they are very permissive they still require a form of attributions which Copilot doesn't provide.
Also for anyone arguing that such ML models "lern", at least the moment they are not exatcly perfect sized or sized below that they are basically guaranteed to verbatime encode partial copies of code, it's fundamental consequence of their design. And while this encoding is "encoded" in some way it's not transformative in the definition of copyright law AFIK. I.e. any big model net is guaranteed to commit hard to trace copyright infringement.