That's not necessarily true. General-purpose 3rd party models (chatgpt, llama3-70b, etc) perform surprisingly good in very specific tasks. While training or finetuning your specialized model should indeed give you better results, the crazy amount of computational resources and specialized manpower needed to accomplish it makes it unfeasible and unpractical in many applications. If you can get away with an occational "as an AI model...", you are better off using existing models.

Here is a better idea: have some academic integrity and actually do the work instead of using incompetent machine learning to flood the industry with inaccurate trash papers whose only real impact is getting in the way of real research.