Will Fair Use Protect Automated Creativity?
When computers develop the potential to produce works that rivals human expression, the difference between human-generated content and computer-generated content may be indistinguishable. Emerging applications of machine learning pose a challenge for copyright law where works have minimal, or even non-existent human interaction.
Artificial intelligence has been used in conjunction with creative work for decades.[1] Traditionally, most computer-generated works of art were limited by the creative input of the programmer and machines operated mainly as instruments or tools.[2] To understand how modern artificial intelligence developed, the popular proverb “if you teach a man to fish, you feed him for a lifetime” comes to mind. Instead of teaching computers how to perform certain pre-determined tasks, engineers contemplated teaching them how to learn to solve tasks on their own.[3] With the evolution of the Internet and the exponential increase of digital information, it has become more efficient for engineers to code programs that access available online data and learn from works made by human beings.[4] This is where machine learning has become a reality.
Machine learning is a subset of artificial intelligence that allows computers to learn on their own without explicit programming.[5] Computer programs intended for machine learning have built-in algorithms that are used to identify patterns from observed data to automate decision-making.[6] Machine learning is distinct from broad artificial intelligence because of its independent learning mechanism. While programmers can set parameters, the work is generated by the computer’s “neural network” system, mimicking the thought process of humans.[7] By classifying information in the way the human brain does, the neural network can be taught to recognize elements and even utilize its input to make predictions.[8] Using a feedback loop allows the program to “learn” whether its decisions are right or wrong and to modify its approach as it goes forward.[9] Learning from past experiences allows the computer program to independently and consistently improve.
Today’s AI relies on large datasets for input called “training data” that often include unauthorized copies of works.[10] The training process involves reduplicating and modifying copyrighted works multiple times.[11] While large-scale computerized reproduction of copyrighted works has been protected by the fair use doctrine,[12] widespread use of this data in the context of machine learning poses a dilemma: Are emerging applications of machine learning “transformative” enough to fall under the fair use doctrine?
If computer programs have the potential to rival human expression, it creates legal obstacles that the current fair use doctrine may not be equipped to handle. First, machine learning challenges non-expressive fair use as it can derive valuable information from the way that authors express their ideas instead of merely using facts about their works.[13] As machine learning becomes more advanced, its use of input appears more expressive and raises questions of possible infringement.[14] Second, machine learning may pose a threat of market substitution to relevant “potential market” for copyrighted works by offering cheaper and more efficient alternatives.[15] Should fair use be expansive enough to incorporate a copyright defense for machine learning, it would “deprive contributors to the store of knowledge of a fair return for their labors.”[16] Conversely, if courts rule that machine learning infringed with its use of copyrighted data, it could potentially halt the progress of valuable technological innovation.[17] For example, machine learning algorithms trained on limited datasets has been seen to lead to skewed results that exacerbate societal bias.[18]
While fair use, in its current state, presents significant challenges when applied to machine learning applications, there is potential for reconciliation. The fair use doctrine allows permissive use of copyrighted works in the interest of advancing knowledge and artistic progress. This aligns with the goal of machine learning and artificial intelligence to understand the nature of learning and promote technological innovation. A viable solution may utilize fair use to create “fairer” AI systems,[19] allowing the opportunity for both legal and scientific discourse to take place.
Footnotes
