On March 16, 2023, the United States Copyright Office (USCO) published Copyright Registration Guidance (Guidance) on generative AI. In the Guidance, the USCO reminded us that it “will not register works produced by a machine or mere mechanical process that operates randomly or automatically without any creative input or intervention from a human author.” This statement curiously conjures the notion of a machine creating copyrightable works autonomously.
While the operation of a machine, or specifically the execution of the underlying AI technology, may be largely mechanical with little human involvement, the design of the AI technology can take significant human effort. If we look at protecting human works that power machines as intellectual property in the broad context where AI has been applied, just like authorship has been an issue when an AI technology is used in creating copyrightable subject matter, inventorship has been an issue when an AI technology is used in generating an idea that may be eligible for patent protection. Unlike the evaluation of authorship, though, the assessment of inventorship puts human contribution to the AI technology front and center. Without getting into the reasons for this difference in treatment, let’s consider the question of whether an AI technology used in creating copyrightable subject matter, or specifically the human contribution to such an AI technology, generally does or does not provide any “creative input.”
It may be helpful to first review some additional statements made in the Guidance: “For example, if a user instructs a text generating technology to ‘write a poem about copyright law in the style of William Shakespeare,’ she can expect the system to generate text that is recognizable as a poem, mentions copyright, and resembles Shakespeare’s style. But the technology will decide the rhyming pattern, the words in each line, and the structure of the text. When an AI technology determines the expressive elements of its output, the generated material is not the product of human authorship.” Well, doesn’t it follow that the person who designed the AI technology determines the expressive elements to some extent and thus could be considered as a co-author if not the sole author of the generated material? Or does it depend on how the person determines the expressive elements?
If we look inside the black box of an AI technology, we often see a machine learning model (e.g., an artificial neural network) with a long list of parameters that were obtained through training on a large set of samples. While the basic structure of such a machine learning model may reflect a generic cognitive process not specific to the task to be performed, a human would decide at least what to include in the set of samples so that the machine learning model generates the intended output. For example, the training samples can be literary works together with their keyword descriptors, and the machine learning model can be intended for capturing relationships between keywords and literary features. Then, a human may decide to include in training samples “Shakespearean poems” as keywords associated with writings that demonstrate a rhyming pattern, a number of syllables in the words of each line, or some other characteristic associated with a Shakespeare poem. The human may decide to include as additional training samples writings that discuss copyright, together with “copyright” as a keyword. The result is a machine learning model that has learned from a vast volume of experiences what literary features should be produced when given possibly unrelated keywords. Essentially, a human identifies and simulates a learning process by creating the machine learning model, and the simulated learning process is applied to generate new output. Then perhaps the question becomes whether learning from experience can incorporate any “creative input.”
The USCO has actually likened a machine to a commissioned artist in the Guidance, which is not a bad analogy. But doesn’t a commissioned artist own the copyright of their creations? A commissioned artist would have learned to create literary works from their own combination of experiences. Can a unique combination of experiences intrinsically be a creative input?
In addition, imagining how a commissioned artist may purposefully add their personal touch in their creation process, one could easily envision a case where a human decides to explicitly instill “expressive elements” into the machine. The human may do so by specifically directing the training of the machine learning model or by simply customizing the output of the machine learning model to generate the final output. For example, the human may hardcode assigning large weights to their favored works of literature in the training samples for the machine learning model or program to replace certain text in the writings produced by the machine learning model with their chosen synonyms. The result is still a computer model that can be readily executed by any individual, even if the final output could conceivably range from individualized to simply wild given even the simplest prompt of keywords. Would it be right for the final output to be deemed not copyrightable?
Today, the owner of an AI technology may have a number of reasons for not retaining any copyright over the output generated by the AI technology, even if the owner might have indicated an intent to retain copyright before relinquishing any such claim and merely requiring a license to use or sell. Such a course of action, however, does not necessarily follow from the owner’s belief and should not lead to our conclusion that the product of an AI technology or any computer model is not generated with the aid of creative input built into the AI technology. Moving forward, it should be interesting to see if the boundary between artificial intelligence and artificial creativity can be drawn, and whether “artificial” should matter at all.
 A human may also determine additional factors, such as how many parameters to use or how to refine the basic structure of the machine learning model, to further improve the performance of the machine learning model in performing the task at hand.