I don't see why this would be a copyright violation anymore than somebody learning something from multiple sources and reformulating what they learned into an answer to a question. As long as it isn't explicitly reciting its training data, there shouldn't be an issue of copyright.