Yhyjyk

Doing some research for an entrepreneurship class. We are building an email classifier for a logistics company to automate part of their customer service process. We are having some discussion on how to properly license this technology, and whether it should be hosted on the cloud or on their internal servers.

Specifically, my questions are:

1) Who actually owns the rights to the trained machine learning classifier output by the machine learning algorithm?

2) Since the company itself owns the data, can they claim rights to the output?

3) If the service is hosted on their servers, is there any way to protect against them hacking the classifier and re-deploying it themselves?

Your title asks the right question first. Before discussing who owns Intellectual Property rights, we should first establish which rights that would be.

There are a few different types of Intellectual Property. Chiefly, those are patents, design patents, trademarks, and copyright. We can quickly dismiss patents and design patents - they exist only after registration, and you would know about that.

It's also clear that the model itself is not a trademark. You may invent a trademark for the model and do business using that trademark, but that trademark would only name the model.

Finally, there's copyright. That requires creative activity at least. But is the ML model the result of creative activity, or is the model created by a mathematical procedure (Stochastic Gradient Descent)? This is not obvious. You probably have some input in pre-selecting the data, but it's likely that the training algorithm had the final decision in weighing all inputs. Reasonable people may have different opinions here. My personal view is that the process is almost fully automated, and therefore not creative.

The Wikipedia article on Threshold of Originality has a section on stuff that machines produce, and it appears that, at least in the US, this would probably not be original enough to be copyrightable. Consult a local lawyer of the appropriate specialty if you want a more definite answer. The only other category of IP that's applicable would be trade secrets, and for that to apply you'd have to keep control of the output and not show it to others.

In the US, anyway, data usually belongs to whoever gathered it, but owning data doesn't mean owning something generated from it.

I assume that there's a generic engine and that the classifier output is what's important here. If it's not copyrightable, then there's nothing stopping anyone with access to it from copying, using, or redistributing it. If you show it to the company without an NDA in place, it won't be a trade secret. If you want to make a profit from it, consider charging for it up front.