Artificial intelligence is an activity that makes machines intelligent.  The key to achieving intelligence lies in machine learning, that is, learning from data and "training data" by inputting and analyzing massive amounts of information to form a data model.  At this stage, information such as copyright works has been completely copied or even changed countless times as input data.  This may infringe the copy right or compilation right of the exploited work, etc.  Whether artificial intelligence "training data" necessarily constitutes copyright infringement, or whether it can be protected by the fair use system, this question is related to the future development of artificial intelligence technology.  Excessive copyright protection will hinder the research and development of artificial intelligence technology; and loose copyright protection may change the pattern of copyright interests and inhibit the creative motivation of human authors.

Artificial intelligence can be divided into "expressive" and "non-expressive" according to the different output content.  "Non-expressive" artificial intelligence does not involve the original expression of the work when it uses the work, but takes the factual information of the work as the object or treats the work as data, focusing on analyzing the physical characteristics of the original data text.  For example, in the "Authors VS Google" case, the purpose of Google search engine copying books was only to provide users with factual information about book publication, not book content.  The court ruled that the copying was fair use.  In this case, the purpose of using a work is to construct a data model by analyzing the "thoughts" behind the expression of the work, rather than generating "expressive" content that may substitute for the used work, which belongs to "non-expressive reasonable use". According to the rules of "convertible use", the more "non-expressive" the purpose of the use of a work is, the higher the degree of "conversion" and the more likely it is to be considered as fair use.  This kind of fair use has basically formed a consensus internationally.

Regarding the question of whether "expressive artificial intelligence" constitutes fair use, no conclusion has been reached yet.  In this type of artificial intelligence development, the purpose of inputting and analyzing works is to train the expression skills of artificial intelligence to create new works by analyzing the original expression of the work.  For example, Google's "Next Rembrandt Project", artificial intelligence inputs all Rembrandt's paintings into a data model for analysis, thereby "creating" more Rembrandt-style works.  Although style is thought, not expression, artificial intelligence products may have a market substitution effect on the works used.  Regardless of whether it is the "three-step method of fair use" or "conversionary use", it is required that the use of the work should not have an unreasonable impact on the market of the work.  The boundaries of intellectual property protection are constantly changing.  With the gradual relaxation of the originality standards of TV program models and computer fonts, the law of expression of works may also be original, and artificial intelligence products based on this law may be caused by and the use of works that have substantial similarity constitutes infringement.  It is conceivable that if Rembrandt is alive, he is likely to sue Google for infringement of its copyright.

However, if this denies fair use, the development of artificial intelligence technology will be hindered.  Because a large number of copied works means high copyright licensing fees, it is difficult for developers to abandon development at the cost, or they tend to use works whose copyright protection period has expired, but the technology developed based on outdated information may not be practical.  The one-sidedness of information may also form a biased data model, the so-called "data discrimination."

At the level of public policy, the fair use system has the function of redistributing public welfare.  It allows specific groups to receive subsidies by restricting the economic interests of right holders.  Through the redistribution of economic power and expression rights, the fair use system enables the public to freely engage in activities that are beneficial to society, while preventing the unlimited expansion of the "monopoly power" of right holders.

This dilemma of the fair use system is difficult to solve through its own system improvement, and the method of mainly statutory license, combined with public license is a more appropriate solution.  First of all, artificial intelligence developers can use the work without obtaining permission in advance, but they need to pay the copyright owner.  Secondly, since the products of expressive artificial intelligence may replace the used works, thereby affecting the authors' market revenue, it is necessary to allow the authors the right to withdraw from the legal license at any time. The "reservation of rights" in the statutory compensation system, that is, the "opt-out" mechanism of the right holder, can meet this requirement.  Finally, as non-professional creators, it is difficult for network users to form a stable transaction model with collective management organizations.  This shortcoming can be compensated by public licenses, that is, the release of some copyright rights in exchange for network.

From PEOPLE.CN

August 24th, 2020