Automatic Test Generation is Like Fighting with Dragons

Having analyzed tests in Java applications, we see that most of them obey one of the templates, whose manual implementation requires a lot of routine work. If you generate a template programmatically, all that remains is to fill it with data and tweak it. What if we go further and automate data selection so that we don't forget about non-obvious scenarios? You can — there are symbolic execution and fuzzing techniques for this purpose. What if you already have some tests in your project? You can use them as samples for the language model when generating new tests. To what extent will it affect the quality of the result? We suggest to figure it out together.