Protecting AI systems

In what ways can an AI model be protected?

AI models are a form of software and therefore can be protected in similar ways. First, there may be a copyright on the model's code. The owner of that copyright thus has the right to exploit the model by making it public or reproducing it. Although others are not allowed to analyse software and view its internal workings without permission of the copyright holder, publishing is still a risk when disclosing a software program. Therefore, a second option is to make the code a sort of trade secret and only allow access to the model without disclosing the underlying software. This is what happens, for example, at Google Translate. You can still try to figure out the underlying model by guessing a model structure and type and then collecting lots of outcomes, but this is (almost) impossible for complicated models and also easy for the Google to detect.