Intellectual Property and Machine Learning: An exploratory study
MetadataShow full item record
Our research makes a contribution by exemplifying what controls the freedom-to-operate for a company operating in the area of machine learning. Through interviews we demonstrate the industry’s alternating viewpoints to whether copyrighted data used as input to machine learning systems should be viewed differently than copying the data for storage or reproduction. In addition we show that unauthorized use of copyrighted data in machine learning systems is hard to detect with the burden of proof on the copyright owner. We also demonstrate how existing products, both physical and software, becomes more vulnerable to reverse engineering with the rapid progress in machine learning, and how this challenges and reduces a company’s freedom-to-operate and the way they appropriate innovation. Another contribution is made from demonstrating how machine learning systems can create new valuable content from those patterns and structures found through parsing databases of texts, images, music or arts, and how this challenge the existing intellectual property regulations. We claim that with huge amounts of data used as input to a machine learning system, giving all intellectual property rights of output to the input data owners may be the wrong thing to do. We also think that intellectual property regulation should start discussing when a user of a machine learning system can be seen as having made a creative effort in the generation of the new content.