Every new technology poses challenges for the concept of intellectual property rights ownership. When copyright was first introduced into legislation in the US following the adoption of the Constitution, the copyright of works extended to “maps, charts, and books” and only for a period of 14 years. Growth of the coverage of copyright over time moved forward in fits and starts, occasionally with the not-so-subtle hand of the Supreme Court in the United States. A perfect example of this extension of copyright was the inclusion of photography in the late 19th Century. When first developed, people viewed the output of a photograph to simply be a chemical process, lacking the originality of activities like painting, and therefore not copyrightable. This was upended in the 1884 Supreme Court decision Burrow-Giles Lithographic Co. v. Sarony, which held that in fact photographs “that are a representation of an author’s original intellectual conceptions” may be afforded protections under copyright.
We are now at another inflection point with a new technology, artificial intelligence (AI), and similar questions about the boundaries of intellectual property rights are coming to the forefront. There could be profound implications for the publishing and scientific communities, which are becoming key sources of training data for artificial intelligence systems, as well as for publishers themselves, potentially becoming reliant on artificial intelligence for creation, curation and engagement of new content.
Last week, the Copyright Office together with WIPO hosted a meeting on Copyright in the Age of Artificial Intelligence. The meeting drew a high-powered roster of speakers and panelists including Francis Gurry, Director General of WIPO; Maria Strong, Acting U.S. Register of Copyright; Andrei Iancu, Director of the U.S. Patent and Trademark Office; Mary Rasenberger, Authors Guild Executive Director; David Hughes, CTO at RIAA; and Ros Lynch, Director of Copyright & IP Enforcement at the U.K. Intellectual Property Office (UKIPO); among many others. This was the second meeting hosted by WIPO and the Copyright Office, with the first being held in Geneva last fall. The two meetings are being held in conjunction with a public consultation that WIPO is conducting to seek input on policy issues related to artificial intelligence and intellectual property rights. The draft paper is “designed to help define the most-pressing questions likely to face IP policy makers” at the convergence of AI and intellectual property issues. The deadline for input on the draft paper is on Friday, February 14, 2020.
There are many interesting threads in this discussion. The first comes in the form of the process of creating and training an artificial-intelligent-like system. The overwhelming majority of the current AI systems need to be trained on real-world information as input data to get the AI to recognize patterns and discern an approach to a problem. Whether, and if so how, should the content — much of which is under copyright — be recognized for its role in training the AI? Could the training of the AI be considered a transformative work in the world of intellectual property? Some people during the day’s discussions viewed the use of content by a machine more similar to reading by humans, while others thought of the process more similar to the type of sampling most common in hip-hop music. In that context, they asked if there were some way to track a novel creation back to some element of the original input file.
This seems to me to fundamentally misunderstand the technology of machine learning, and what makes these questions more fraught. In the most novel of cases, machine learning algorithms do not simply regurgitate the best compilation of the input, but instead generate a new outcome based on the training. Some have argued that by exposing the AI to the world of content to derive new approaches is most like how humans are exposed to art, reading, and cultural styles in developing their own voices. Some are concerned that machine reading is simply a way to ingest and utilize authors’ works without due compensation, particularly as the machine is generating new works based on original works. One example of this discussed during the forum was the work of the Next Rembrandt Project. This effort, led by Microsoft, ING, and TU Delft, Mauritshuis, and Museum Het Rembrandthuis sought to create a new three-dimensional painting based on a thorough AI analysis of dozens of Rembrandt’s original works. One speaker compared the work of the AI in this case to that of a forger. If so, is that a work we want to privilege with copyright protections, though it should be noted that forgeries are still works in their own right and do get the benefits of copyright.
Another interesting thread has precedent from an odd direction, with an argument that derived from the now famous monkey selfie. In 2011, a nature photographer left his camera on a tripod and an endangered Celebes crested macaques, intrigued by its reflection in the lens snapped perhaps hundreds of pictures of itself. One of those photos ended up promoted by the photographer and it ended up in the British press. Other sites, such as Wikipedia and Techdirt reproduced the photo on their sites, that the photographer and PETA eventually perused in court to seek compensation as violation of copyright. Whether the photographer could assert copyright in the photograph was eventually dismissed by the Ninth Circuit court of appeals in 2018.
In the Copyright Office’s Compendium of U.S. Copyright Office Practices, released on 22 December, 2014, the Office stated that, “only works created by a human can be copyrighted under United States law, which excludes photographs and artwork created by animals or by machines without human intervention” and furthermore, “Because copyright law is limited to ‘original intellectual conceptions of the author,’ the [copyright] office will refuse to register a claim if it determines that a human being did not create the work. The Office will not register works produced by nature, animals, or plants.” What is left open to wide interpretation is the role of human intervention. In the case of AI, so the argument goes, there is human intervention in the design of the algorithms, in their training, and in their post-algorithm curation.
Reflecting on the Google AlphaGo, which was described at the time as ‘mysterious’ and ‘beautiful’ in its move creations, it seems that human ingenuity wasn’t involved in the exact moves, nor the eventual strategy. I want to be careful though that I am not of attributing “creativity” in the traditional sense, to the machine or some higher intellect, which should be privileged in some way. Quite the contrary. There is a long-discussed thought experiment that stated if an infinite number of monkeys typed away randomly for long enough, they would recreate the complete works of William Shakespeare. The challenge today is that machines can virtually “type away” at a pace unmatched by any human and they do so in a way that is even less random then the monkeys in the thought experiment.
Could humans essentially be blocked out of content creation by the pace of AI text generation and the resulting claims of copyright for every possible meaningful text combination? With the expansion of tools, matched with the increasing speed of processing and available storage, such a world isn’t beyond comprehension. There are dozens of text-generating robots now that you can turn to produce your computer science paper, your news article, or even your book. Last winter, OpenAI, a nonprofit artificial intelligence research lab — perhaps in a pique of advertising bravado — said their GPT-2 software is so good they are worried it could be misused and therefore refused to release it. However, they overcame that ‘fear’ in the fall and released it anyway. Most of what these basic AI text generation tools produce is meaningless confusing drivel, but they are getting much better and the GPT-2 tool is proof of that. However, even within a simple AI, if you search long enough, you’ll find something that catches your fancy.
This is how the author Janelle Shane came up with the title of her book, You Look Like a Thing and I Love You: How Artificial Intelligence Works and Why It’s Making the World a Weirder Place. She was developing and testing an AI chat bot to see if it could come up with dating pick-up lines and that phrase drew her attention as the best of the lot. She’s obviously copyrighted it and the book, yet is drawing out the one thing derived by a machine that is vaguely appropriate copyrightable? Sure, it counts as “curation” in the strictest sense of the word, but simply selecting a single sea shell off the beach does not count as curation of shell collection and a single shell is not copyrightable, as the output of natural action. One might take a collection of shells and claim copyright over the collection, because the act of curation of a collection has some value (although even this is debatable, see Section 312.3 of the Compendium).
So if the infinite Mechanical Turks within machines crank out the next Shakespearean sonnet, will that be copyrightable? The lawyers in industry, WIPO, and the Copyright Office sure seemed to want it to be, gauging from the perspectives shared during the forum. This opinion was buoyed by a stunning legal victory for the owners of AI systems that are generating content in China. In January, a court in Shenzhen, China ruled in favor of the tech company Tencent, determining that financial reporting articles generated by AI are entitled to copyright protection. This followed quickly on another decision in the opposite direction in which two patents were rejected in the EU Patent office because they were created by machines. The team behind the DABUS software that created new patents attempted to argue that machines should be thought of not as tools, but as creators in their own right. While the initial claim was rejected, it seems obvious that those behind this initial case, as many others, will try again to assert intellectual property rights in the work of machines.
These cases indicate that there appears to be a growing divergence in our common law understanding of what a creator is. If the legal structures remain fixed in the presumption that only intellectual work of humans can be protected, the billions invested in the AI that is managing ever more of our digital interactions with the world might not be covered by intellectual property ownership. Here it is important to distinguish the difference between algorithmic creation and the outputs of those algorithms. The creation of the machine-learning tool is certainly patentable as processes, much like other forms of software. The distinction is in the output, which if truly a form of artificial intelligent machine learning, then the outputs are beyond the control of the originators of the algorithm.
My sense is that in a process driven by lawyers with a vested interest in the extension of intellectual property to the output of machines, and with the support of billions of dollars of technological investment, past precedent will be damned and the work of machines will be carved out for special treatment by our intellectual property rules. Regardless of the fact that as machine learning and applications grow they will be further outside of the realm of human understanding, it will still be claimed that the designers of the system somehow knew the machines would produce such-and-such outcome, and therefore they should be granted ownership of the outputs.
As AI digs deeper and deeper into more of our interactions, we want to reflect deeply on what sets human creativity and ingenuity apart from the work of the machines we are setting forth into the world. Intellectual property law was created to reward the creation of novel ideas and to pass them into the public domain after a LIMITED time frame. Within that context, to what extent do we want to reward machine creation of ideas or disfavor it, particularly in comparison with traditional human creation? I expect we will want to cherish the intellectual outputs of people more than the rote machinations of an algorithms, even if they are artfully rendered.
Editor’s Note: This article is republished with the author’s permission, with first publication on The Scholarly Kitchen.