In 2011, a Celebes crested macaque took a shot that was heard ’round the world. In a jungle in Indonesia, it depressed the remote trigger button of a photographer’s camera, effectively taking a selfie. The “monkey selfie” has ignited a great deal of commentary musing on the nature of copyright ownership. The human photographer claimed that he was the author of the photograph because he had “engineered” the shot and that “it was my artistry and idea to leave them to play with the camera and it was all in my eyesight. I knew the monkeys were very likely to do this and I predicted it. I knew there was a chance of a photo being taken.” Wikimedia Commons, among others, disagreed and posted the photograph on its website claiming that the photograph was in the public domain because its true author was a monkey. The U.S. Copyright Office has implicitly agreed with Wikimedia by including “a photograph taken by a monkey” in a list of examples of unprotectable “Works that Lack Human Authorship” in section 313.2 of the most recent edition of the Compendium of the U.S. Copyright Office Practices, Third Edition.
The U.S. Copyright Office notes in the Compendium that “to qualify as a work of ‘authorship’ a work must be created by a human being.” Works “produced by nature, animals, or plants” are not copyrightable and will not be registered by the Copyright Office. The Copyright Office’s position on works created by animals has received a lot of press, but its impact on the real world is dubious (that is, unless one is concerned that all selfies could be deemed to be the product of animals of questionable intellectual agency …). Considerably more important is the remainder of section 313.2 – “Similarly, the [Copyright] Office 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.”
Where Are the Humans?
The Copyright Office’s requirement that a human have “creative input or intervention” appears to have its roots in the first court decision that struggled with the interplay between man and machine – Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53 (U.S. 1884). There the Supreme Court decided for the first time whether copyright could extend to a photograph – the one in question being of the author Oscar Wilde. Defendants argued that photography required very little artistic intervention – “It is simply the manual operation, by the use of these instruments and preparations, of transferring to the plate the visible representation of some existing object, the accuracy of this representation being its highest merit.” Id., at 59. However, the plaintiff claimed that he set out to create a “graceful photograph,” and that he achieved this:
by posing the said Oscar Wilde in front of the camera, selecting and arranging the costume, draperies, and other various accessories in said photograph, arranging the subject so as to present graceful outlines, arranging and disposing the light and shade, suggesting and evoking the desired expression, and from such disposition, arrangement, or representation, made entirely by plaintiff, he produced the picture in suit.
Id., at 60. The Court found that the photograph was accordingly the plaintiff’s “intellectual invention,” qualifying the work for copyright protection. Although no one would question today that photographs are generally subject to copyright protection, a photograph whose “accuracy of  representation [is] its highest merit,” i.e., one that merely attempts to faithfully re-create something from the public domain, can still lack originality and thus be ineligible for copyright protection. See, generally, Bridgeman Art Library, Ltd. v. Corel Corp., 36 F.Supp.2d 191 195 (S.D.N.Y. 1999).
The key to copyright protection is the role of human expression in using the machine, but what happens when human intervention is minimized or eliminated? An interesting convergence of technologies is occurring beneath our noses – the rise of both “machine learning” and of the Internet of Things (IoT). Machine learning results in at least two outputs that could be subject to copyright protection – content (such as Google’s DeepDream or the automated journalism created by companies such as Narrative Science) and program modifications (such as by collecting, structuring, and analyzing data and creating outputs based on that analysis). It has even been suggested that the time will come soon when computers will self-generate code.
IoT, on the other hand, refers to the “network of physical objects, devices, vehicles, buildings, and other items [that] are embedded with electronics, software, sensors, and network connectivity, which enables these objects to collect and exchange data.” This sharing of data and analysis between machines is producing everything from smart cars to smart homes, and transforming healthcare and heavy industry. It has been suggested that artificial intelligence (of which machine learning is a subset) and IoT are “inseparable.” IoT networks will likely become more and more complex and rely increasingly on artificial intelligence and machine learning. The sheer volume of data that will be generated by IoT will necessitate autonomous processing of data, including identification of patterns and deviations from those patterns to help establish user preferences, responses to environmental conditions, and other adaptations to external stimuli. So in this sort of echo chamber of machine learning from machine, what happens to the copyrightable subject matter that gets generated?
Merely a Tool?
Some commentators have suggested that machine learning programs should be treated just like cameras were by the Court in Burrows-Giles – as mere tools deployed by humans that simply enable human expression. Under this conception, the machine simply enables the artistic process between author and expression, but does not negate the human’s expression. By selecting which picture to feed into the DeepBlue program, for example, the human actor would become the author of whatever the algorithm outputs. However, this conception does not withstand much scrutiny – “Human creativity is necessary for the production of the work, but the human creative agent is not the author of the work in the traditional sense. Nor is generative software an author’s tool in the traditional sense; unlike a pen or a paintbrush, or even a camera, generative software has a verbal or visual vocabulary of its own and the ability to compose a range of distinct works from that vocabulary by independently applying a system of rules.” Annemarie Bridy, “Coding Creativity: Copyright and the Artificially Intelligent Author,” 2012 STAN. TECH. L. REV. 5, 20 (2012).
Put differently, machine learning programs, especially ones that will be learning from other machines, do not leave much room for human expression. Given the amount of control that the program exerts over the output, it is hard to see how the user of the program could be deemed an author. Imagine that a circus performer trained a monkey to paint images that looked remarkably like Starry Night. Could the trainer claim that he was the author of the painting, as opposed to the monkey, and rescue the created images from the public domain? As seems clear from the Compendium, certainly not.
Could the programmers claim that what the artificially intelligent program produces is “based upon” the preexisting code, such that the new output is merely a recasting, transformation, or adaptation of the preexisting code (as paraphrased from §101 of the Copyright Act)? Perhaps. But the problem would persist that the “author” of the derivative work would be the computer, thus placing the derivative work in the public domain.
A “joint work” is defined in §101 of the Copyright Act as “a work prepared by two or more authors with the intention that their contributions be merged into inseparable or interdependent parts of a unitary whole.” Perhaps an argument could be made that the human author intended that his or her contribution be merged with the contributions of the artificially intelligent author to create a unitary work (this would beg the question of whether the human author should be the programmer or the end user feeding inputs into the program, but we will leave that discussion for another day). It would seem that the “joint author” would be the computer program, which would thus make the joint author the public domain, likely making the work uncopyrightable.
There are, however, several examples where courts have treated as a legal nullity what amounts to a nonhuman element in a work that arguably had a “joint author,” and vested full authorship in the human joint author. This occurs in “psychography” cases, where the author claims that he or she was merely a vessel through which some divine spirit dictated the work in question. See, e.g., Cummins v. Bond, 1 Ch. 167 (1927); Penguin Books U.S.A., Inc. v. New Christian Church of Full Endeavor, Ltd., No. 96 Civ. 4126, 2000 U.S. Dist. LEXIS 10394 (S.D.N.Y. July 21, 2000). While these cases focused on the degree to which the human involved him- or herself in arranging, editing, and presenting the allegedly dictated material, and did not address the consequences of having a “non-physical” joint author, they are examples of courts emphasizing the contributions of the human authors over the entities that cannot own copyright.
Work Made for Hire?
Perhaps the best solution to the problem has been suggested by Professor Annemarie Bridy in her article “Coding Creativity: Copyright and the Artificially Intelligent Author.” There Professor Bridy makes the compelling case that a slight change in the law is needed to recognize that providing a computer with a set of instructions and a desired outcome, configuring that program in a fashion so that it has the means and methods of performing that work, and then simply letting it do its thing, is very similar to what takes place under the work-for-hire doctrine. Unfortunately, of course, such legislative changes to the law occur at a glacial pace. The fact that amendments to the Copyright Act have made their way through congressional hearings for years now with no definitive proposed legislation in place, and the fact that this computer authorship issue was not included in those hearings, does not bode well for a legislative fix to occur in the next decade or longer.
In the meantime, it appears that ownership rights will have to be fixed by contract. This will be no small feat in the complex relationships that define IoT. Authors of computer programs need to be thinking now, in the early days of IoT, about establishing strategies for claiming ownership and policing that ownership. However, to the extent they can, producers of the code that create generative software will have to decide whether to lock up outputs through licensing arrangements or, as Google has done with DeepDream, donate those outputs to the public domain.