For programming languages we’re not just bringing in more information, we can run translated code to see if the translation is correct. You can’t do that in many fields at all. With synthetic data, who’s to say whether a reconstruction is accurate or just noise?

With a programming language you can run the resulting code and see that it works exactly the same as the code it’s translated from. You can’t do that in very many cases at all. Who’s to say a synthetic reconstruction is accurate, or if you’re just adding noise?