Offshoring’s quick, easy, and guaranteed by overwhelming economic incentives---it zeros solvency
Matthew U. Scherer 16, Senior Policy Counsel for Worker Privacy at the Center for Democracy & Technology, J.D. from Georgetown University Law Center, Former Editor-in-Chief of The Georgetown Journal of Legal Ethics, M.S. in Educational Policy from the University of Pennsylvania’s Graduate School of Education, Attorney at Buchanan Angeli Altschul & Sullivan LLP, “Regulating Artificial Intelligence Systems: Risks, Challenges, Competencies, and Strategies”, Harvard Journal of Law and Technology, 29 Harv. J. Law & Tec 353, Volume 29, Number 2, Spring 2016, Lexis
The sources of public risk that characterized the twentieth century -- such as nuclear technology, mass-produced consumer goods, industrial-scale pollution, and the production of large quantities of toxic substances -- required substantial infrastructure investments. This simplified the regulatory process. The high cost of building the necessary facilities, purchasing the necessary equipment, and hiring the necessary labor meant that large corporations were the only non-governmental entities capable of generating most sources of public risk. Moreover, the individuals responsible for installing, operating, and maintaining the infrastructure typically had to be at the physical site where the infrastructure was located. The physical visibility of the infrastructure -- and of the people needed to operate it -- made it extremely unlikely that public risks could be generated clandestinely. 61 Regulators thus had little difficulty determining the "who" and "where" of potential sources of public risk.
By contrast, AI research and development can be performed relatively discreetly, a feature that AI shares with many other Information Age technologies. In 2009, Professor John McGinnis wrote that "[a]rtificial intelligence research is done by institutions no richer than colleges and perhaps would require even less substantial resources." 62 This actually overstated the resources necessary to participate in AI development, particularly with the rise of open-source programming. Simply put, a person does not need the resources and facilities of a large corporation to write computer code. Anyone with a reasonably modern personal computer (or even a smartphone) and an Internet connection can now contribute to AI-related projects. Individuals thus can participate in AI development from a garage, a dorm room, or the lobby of a train station. This potential for discreetness provides the most jarring difference between AI and earlier sources of public risk.
The participants in an AI-related venture may also be remarkably diffuse by public risk standards. Participants in an AI-related project need not be part of the same organization -- or, indeed, any organization at all. Already, there are a number of open-source machine-learning libraries; widely dispersed individuals can make dozens of modifications to such libraries on a daily basis. 63 Those modifications may even be made anonymously, in the sense that the identity in the physical world of individuals making the modifications is not readily discernible. 64
The AI program itself may have software components taken from multiple such libraries, each of which is built and developed discretely from the others. 65 An individual who participates in the building of an open-source library often has no way of knowing beforehand what other individuals or entities might use the library in the future. Components taken from such libraries can then be incorporated into the programming of an AI system that is being developed by an entity that did not participate in assembling the underlying machine-learning library.
These characteristics are not limited to open-source projects or freely available material. Many modern computer systems use commercial off-the-shelf ("COTS") hardware and software components, most of which are proprietary. 66 The ease with which such components can be acquired makes it tempting to maximize use of COTS components to control costs, despite the potential security issues associated with using software components developed wholly outside the system developer's control. 67 Modern AI programming is no exception; few, if any, AI systems are built from the ground up, using components and code that are wholly the creation of the AI developers themselves. Moreover, if past is prologue, the physical components of an AI system will be manufactured by yet other entities separate from those that developed the AI system's programming. While separately developed components are present in all complex machinery to a certain extent, the level of discreteness and the scale of interactivity between software and hardware components in modern computer systems already rivals or exceeds that of prior technologies, and that complexity seems likely to increase further with the development of stronger forms of AI. 68
In all likelihood, there will be considerable variation in the discreteness of the components of AI projects. Some AI systems likely will be built primarily with COTS or freely available hardware and software components, while others will mostly utilize programming and physical components designed and developed specifically for the AI project in question. Because of the cost advantages inherent in maximizing the use of COTS and freely available components, however, it seems all but certain that some AI systems will operate using a mishmash of hardware and software components harvested from many different companies. The interaction between numerous components and the disparate geographic locations of the companies involved will greatly complicate any regime designed to manage the risks associated with AI. 69
Finally, the inner workings of and the interactions between the components of an AI system may be far more opaque than with earlier technologies. COTS software components may be easy to acquire, but their coding often is proprietary. Critical features underlying an AI system's operation thus may not be immediately apparent or readily susceptible to reverse engineering. Contrast this with automobiles -- one of the twentieth century's great sources of public risk. Automobiles consist of approximately 30,000 individual physical parts, 70 but the ways in which those physical components interact is well understood -- not only by the designers and manufacturers of the vehicle itself, but also by the makers of parts for the vehicle and mechanics responsible for repairing the vehicles after they reach consumers. It seems unlikely that AI systems will demonstrate similar transparency if their development follows now-prevailing trends in information technology. Defects in the design of a complex AI system might be undetectable not only to consumers, but also to downstream manufacturers and distributors. 71
Taken together, these characteristics confront regulators with fundamental logistical difficulties that were not present in earlier sources of public risk. Participants in AI projects may be located in multiple countries and have no legal or formal contractual relationship with one another. Attempts by any one country to regulate their citizens' participation in such projects may not greatly impact the projects' development. Even for projects involving large firms, the relatively low cost of infrastructure and the small physical footprint required for AI development means that firms could simply move AI development work offshore if regulations in their country of origin prove too intrusive. Many would likely do so given the competitive advantages that accompany advances in AI. 72 [FOOTNOTE] 72 See, e.g., Vernor Vinge, The Coming Technological Singularity: How to Survive in the Post-Human Era, 10129 NASA CONF. PUBLICATION 11, 15 (1992), http://ntrs.nasa.gov/archive/nasa/casi.ntrs.nasa.gov/19940022855.pdf [https://perma.cc/J2SU-UK5E] ("In fact, the competitive advantage . . . of every advance in automation is so compelling that passing laws, or having customs, that forbid [human-level AI] merely assures that someone else will get them first."). [END FOOTNOTE]