Artificial Ingenuity Is Here

Exploring the legal challenges posed by machine inventors and authors

BY LERON VANDSBURGER (A HUMAN)

In recent years, artificial creative systems have crossed from rudimentary creativity, typified by the traceable recombination of known concepts or predictable rules-based outputs, to a creative domain that—if practiced by a human—would be worthy of interpretation, analysis, examination, or critique. Enabled by large sets of training data and computing resources, generative machine-learning models can produce truly original works of visual art, music, literature, and technical innovation. When weighed against the standards used to judge originality, novelty, inventiveness, or misappropriation, works of artificial ingenuity would merit protection under intellectual property law.

Efforts at getting substantive review of artificially created works in the patent context have been blocked in the United States, Europe, and the United Kingdom for lack of a natural person to name as an inventor. Australia and South Africa reached a different answer to the same question, bringing the technical state of the art into a confrontation with legal systems established to protect and encourage the arts and sciences—legal systems that will only grow as artificial ingenuity and creativity become more sophisticated and accessible. Already, generative models and their conditional variants are essential elements in machine translation and human-computer interface technology, and will soon permit software developers to “write code” by speaking in their native languages to a trained AI model.

TECH SUMMARY

Description of generative models

Two examples of generative models are the sequence-to-sequence model and the generative adversarial network (GAN). While quite different in structure, both models are trained to generate “original” outputs according to a learned distribution. Synthetic natural language and sound are often generated using sequence-to-sequence models, while images and other visual media are often generated using GANs, but combinations of these models are also common. In some cases, images can be described as sequences of pixels, making image generation possible with sequence-to-sequence models.

Very briefly, a sequence-to-sequence model includes an encoder model and a decoder model that are trained in tandem to convert an input sequence to an output sequence and vice versa. The “sequence” in the model’s name describes a variable length sequence of values that can represent letters, numbers, pixels, or other values. A familiar example of a sequence-to-sequence model is a translator that takes in a phrase in one language and generates the corresponding phrase in another language. Sequence-to-sequence models are also useful as predictors, where the sequence is a vector of time-series data.

Training the sequence-to-sequence model includes comparing the output of the decoder to ground truth training data, as part of supervised learning. Training can also include bi-directional operation so that the model can also generate the input sequence from the output sequence. In the example of the translator, training a model to translate from English to Mandarin would rely on a large set of training data including English and corresponding Mandarin phrases. Bi-directional training would, in effect, train a model to be an English-to-Mandarin translator and a Mandarin-to-English translator.

A GAN also includes two models, but instead of working together the GAN pairs a generator model with a discriminator model, each with a different role. The generator is trained to create an arbitrary output without an explicit input that could be an image, sound, text, or video, or many other forms. The discriminator is trained to judge the output of the generator against the training data to decide if it is from the training set or not. Training is done in an adversarial way, with each model being trained to “defeat” the other. After many rounds of training, the output of the generator can be quite convincing to humans, giving the impression that the output comes from the same ground truth dataset used to train the GAN. For this reason, GANs form a basis for deceptive technologies used to create deep fakes and voice/affect simulators,¹ but also have been developed to generate art.²

Real world examples of artificial ingenuity

Two notable examples of artificial ingenuity include DABUS,³ an “artificial inventor” created by Imagination Engines, Inc. and OpenAI’s GPT-3 model,⁴ but many other examples exist. DABUS and GPT-3 highlight developing legal and popular recognition of machines as creative agents.

DABUS: The Artificial Inventor Project (AIP) is an international group of IP attorneys and computer scientists that is working with Imagination Engines to secure patents for artificially generated inventions around the world.⁵ The DABUS system generates outputs that are used to apply for patents naming DABUS as the sole inventor, which cannot be traced to a human’s inventive concept, either through training, selection of training data, or model architecture.

DABUS is “a swarm of many disconnected neural nets, each containing interrelated memories, perhaps of a linguistic, visual, or auditory nature.”⁶ As a machine-learning model, DABUS simulates associational creativity to generate outputs relatively free of external constraint. In technical terms, “through cumulative cycles of learning and unlearning, a fraction of these nets interconnect into structures representing concepts, using relatively simple learning rules. Thereafter, such ephemeral structures fade, as others take their place, in a manner reminiscent of what humans consider [a] stream of consciousness.” In this way, DABUS forms a new architecture for each problem without human intervention, such that the system structure and output are not traceable to a human inventor.

Inventions created by DABUS include a fractal food container and a “neural flame,” described as a light source designed to attract attention and stimulate mental activity using transient light signals that incorporate a fractal dimension.⁷ The apparent tendency of DABUS to add a fractal dimension to objects or methods could be a result of its creativity paradigm (e.g., the simple learning rules mentioned above), but it could also be a result of simulating natural systems that present fractal geometries.

***A drawing of a fractal* food container submitted as part of a patent application naming DABUS as the sole inventor.*** *The fractal shape of the container purportedly can be used to attach multiple containers together without using external components or adhesive. The high surface area also assists in heat transfer when heating or cooling the container.

GPT-3: The General Purpose Transformer (GPT) model is a text generator developed by OpenAI. The GPT-3 is an example of a sequence-to-sequence model that accepts natural language as inputs and generates natural language or code outputs. In particular, GPT-3 is a deep neural network designed to translate text, answer questions, summarize natural language, and create original written text including software code. The outputs of GPT-3 can be indistinguishable from text created by humans and have been used in popular media with the model being credited with authorship.⁸ GPT-3 has also been incorporated into products as a tool to translate natural language into code.⁹ A noteworthy aspect of GPT-3 is that it conceptualizes natural language inputs and describes abstract conceptual outputs using an arbitrary “language” that can be natural language or structured language, such as code. In this way, GPT-3 generates outputs that satisfy the originality threshold of copyrights.

LEGAL TREATMENTS

The global nature of computer technology, and explicit goals of organizations like the AIP, have led to simultaneous test cases of artificial creativity in multiple IP jurisdictions, with inconsistent results. The United States and Europe have refused to recognize inventorship or authorship in machines. Australia, however, has granted patents that name DABUS as sole inventor, while refusing to credit machines with authorship. These divergences between jurisdictions and between IP domains can be traced to technical and philosophical issues that are difficult to reconcile with creative machines.

Copyright

In the United States, a copyright is only available to “register an original work of authorship, provided that the work was created by a human being.” Copyright law only protects “the fruits of intellectual labor” that “are founded in the creative powers of the mind.” See Trade-Mark Cases, 100 U.S. 82, 94 (1879). Because copyright law is limited to “original intellectual conceptions of the author,” the U.S. Copyright Office refuses to register a claim if it determines that a human being did not create the work. See Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53, 58 (1884).

In Europe, a consistent position has been that copyright only applies to original works, and that originality must reflect the “author’s own intellectual creation.” See C-5/08 Infopaq International A/S v Danske Dagbaldes Forening. Similarly, in Australia, a human author is required and a work generated with the intervention of a computer cannot be protected by copyright. See Acohs Pty Ltd v Ucorp Pty Ltd.

In contrast, jurisdictions such as India, Ireland, New Zealand, and the U.K. attribute authorship to the nearest human creator. The philosophical basis for this distinction lies in recognizing the creative effort dedicated to building a creative machine. For Americans, this approach parallels the treatment of photography in Burrow-Giles. As an example, the U.K. defines that for “a literary, dramatic, musical or artistic work which is computer-generated, the author shall be taken to be the person by whom the arrangements necessary for the creation of the work are undertaken.” (U.K. copyright law, section 9(3) of the Copyright, Designs and Patents Act (CDPA)).

Patent

Current patent systems trace their origins to enlightenment-era philosophies of rewarding individual genius, in contrast to the preceding system by which a patent could be bought or otherwise arranged politically. In the United States, the patent system was created to stimulate the economy through innovation and to discourage anti-competitive behavior, such as secrecy and destructive monopolies through mechanisms like mandatory publication and substantive examination. Similar systems were established in Europe, albeit with slightly different philosophical bases.

Unforeseeable at the time of the drafting of the U.S. Constitution, artificial ingenuity is precluded in the U.S. and in many other IP jurisdictions on a statutory basis by requirements for naming natural persons as inventors. AI inventors, while capable of generating patentable outputs, are excluded from the patent system for lack of legal status.

Most of the test cases in the patent regime result from explicit efforts by the AIP. In April 2020, the U.S. Patent and Trademark Office (USPTO) ruled that only “natural persons” could be credited as the inventor of a patent.¹⁰Federal District Court Judge Leonie Brinkema reinforced the USPTO interpretation by tying the term “individual inventor” to a “natural person.” See Thaler v. Hirshfeld, 20-903, 2021 WL 3934803 (E.D. Va.). Applications in Europe and the U.K. have met with similar treatment, being denied substantive examination for lack of a human inventor. For example, the international patent office (IPO)¹¹ and the European Patent Office (EPO)¹² have ruled that patent applications listing DABUS as an inventor are void.

This approach is not consistently applied everywhere, even in jurisdictions that trace their origins to the British legal tradition. Notably, South Africa and Australia have each granted patents to DABUS. See Thaler v Commissioner of Patents [2021] FCA 879. In Australia, the judge found that while only a human or other legal person can be a patentee, it is a fallacy to extend ownership to inventorship, to say that an inventor can only be a human.¹³ As such, an inventor can be an artificial intelligence system, but cannot be the owner, controller, or patentee of the patentable invention. The judge extended his reasoning to include that denying a patent to an otherwise patentable invention that is eligible, novel, and inventive, is “antithetical” to the purpose of the patent system. In this way, Australia and South Africa represent a perspective on artificial ingenuity that focuses on the value of generative models to the stimulation of the arts and sciences, rather than as a reward for individual inventive effort.

CONCLUSION

Artificial creative systems including DABUS and GPT-3 interrogate the purpose of the IP system. Is a patent meant to reward individual genius or to stimulate the marketplace of ideas? Or is intellectual property meant to discourage anti-competitive behavior like secrecy, destructive monopolism, and political corruption?

Generative models are complex, being created and maintained by large teams of experts that can be entirely uninvolved in the models’ creative end-use. For example, GPT-3 is accessible to anyone to use, so would copyright vest in its creators or its users? Artificial ingenuity is within the authority of the government to protect, because intellectual property as a concept is itself a creation of the human mind. IP systems can be reshaped to best serve society by balancing the benefits of artificial ingenuity against the potential risks of artificial competition with human genius and the concentration of innovation in the hands of resourceful entities.

About the author

Leron Vandsburger, Ph.D., is a member of the board of Washington Lawyers for the Arts and an associate at Christensen O’Connor Johnson Kindness PLLC. He focuses his practice on patent prosecution and counseling in the areas of materials, engineering, and electronics. Representative technology experience includes AI/machine learning, materials chemistry, optics, computer vision, and AR/VR systems. He can be reached at:

Leron.Vandsburger@cojk.com

NOTES

1. “Designed to Deceive: Do These People Look Real to You?” The New York Times, available at www.nytimes.com/interactive/2020/11/21/science/artificial-intelligence-fake-people-faces.html.

2. Xue, A. “End-to-End Chinese Landscape Painting Creation Using Generative Adversarial Networks,” 2021 IEEE proceedings. Available at https://openaccess.thecvf.com/content/WACV2021/papers/Xue_End-to-End_Chinese_Landscape_Painting_Creation_Using_Generative_Adversarial_Networks_WACV_2021_paper.pdf.

3. Thaler, S. “DABUS in a nutshell.” APA Newsletter on Philosophy and Computers, Vol. 19, no. 2 (Spring 2020). 2019. Available at https://cdn.ymaws.com/www.apaonline.org/resource/collection/EADE8D52-8D02-4136-9A2A-729368501E43/ComputersV19n2.pdf.

4. Tom Scott: “I asked an AI for video ideas, and they were actually good,” YouTube. Available at https://youtu.be/TfVYxnhuEdU?t=1.

5. The Artificial Inventor Project: https://artificialinventor.com/.

6. See note 3.

7. PCT Application serial number WO2020079499A1.

8. GPT-3. “A robot wrote this entire article. Are you scared yet, human?” The Guardian. Available at www.theguardian.com/commentisfree/2020/sep/08/robot-wrote-this-article-gpt-3.

9. “From conversation to code: Microsoft introduces its first product features powered by GPT-3,” The AI Blog, available at https://blogs.microsoft.com/ai/from-conversation-to-code-microsoft-introduces-its-first-product-features-powered-by-gpt-3/.

10. USPTO ruling that only “natural persons” can be credited as the inventor of a patent: www.theverge.com/2020/4/29/21241251/artificial-intelligence-inventor-united-states-patent-trademark-office-intellectual-property.