INTRODUCTION 62 PART I: WHY NOT USE AUTONOMY TO DISTINGUISH PRODUCTS FROM THINKING ALGORITHMS? 70 A. On Automation and Autonomy 70 B. The Many Faces of Autonomy 72 PART II: PRODUCTS LIABILITY RATIONALES 77 PART III: IS IT A PRODUCT? IS THERE A DEFECT? 82 A. On "Products" 83 B. On "Defects" 84 PART IV: DISTINGUISHING PRODUCTS FROM THINKING ALGORITHMS 87 A. Examples of Products Versus Thinking Algorithms 88 B. Which Products Liability Rationales Apply to Which Algorithms 90 1. Promoting Safety 90 a. The OODA Loop--General 94 b. "Observe" ("Information Acquisition") 95 c. "Orient" ("Information Analysis") 97 d. "Decide" ("Decision Selection") 99 e. "Act" ("Action Implementation") 101 f. Interim Summary and Measurability of Success Rates 101 2. Avoiding a Chilling Effect 102 3. Victims' Compensation 104 C. Keep It Simple, System 107 CONCLUSIONS 113 INTRODUCTION
Culminating in the Industrial Revolution, machines and tools replaced or assisted humans in performing physical tasks. (1) Drills, engravers, weaving machines and the like employed machines' physical advantages to free human beings from repetitive physical labor. In general, damage caused by such tools has been governed, since the middle of the 20th century, by the legal framework of products liability. (2) Under this system the seller, manufacturer, distributer, or any other party in the distribution chain of a defective product is liable for the physical harm caused to the user or her property. (3) Thus, victims of tools that have caught fire or come apart, (4) of motor vehicles that have crashed, (5) of food containing external undesired objects, (6) etc., could all have brought successful products liability claims against the products' manufacturers or sellers.
Over time, and with the advance of technology, the products used by humans became more sophisticated. Machines were no longer used merely to replace humans in performing physical tasks. Instead their superior computational abilities were utilized to assist or replace humans in processing data. The electronic calculator, for example, allowed engineers, merchants, accountants and other professionals to provide better and quicker outputs; (7) autopilots were installed in airplanes to improve flight safety through an automated system capable of processing huge amounts of information in split seconds, (8) while cruise-control and auto-parking devices were similarly installed in cars. (9) Despite the increasing level of "sophistication" of these machines and devices, manufacturers and sellers of these products were generally held to the traditional products liability legal framework, (10) akin to the liability applied to "simpler" or "less sophisticated" products. (11)
Technology is ever advancing, and in addition to relinquishing physical and computational tasks to machines, algorithms' self-learning abilities now allow them to reach their own conclusions based on databases of previous cases. (12) This in turn enables humans to both entrust machines with making complex decisions that until lately required human discretion and even replace professional human judgment in matters of expertise where there is no clear right or wrong answer. (13)
In the field of law, for example, virtual attorneys--such as ROSS Intelligence's cognitive computing platform that works with IBM's Watson--are utilized by law firms in conducting legal research, (14) algorithmic ODR mechanisms solve disputes online (often without any human facilitator), (15) and bail algorithms determine whether defendants awaiting trial may post bail to be released. (16) Physicians, too rely more and more on algorithms in order to diagnose medical conditions and select optimal treatment. (17) Meanwhile, algorithms significantly assist, or sometimes replace, tax advisors, company directors, and even priests. (18)
Yet advanced as such machines and algorithms may be, occasionally they are still bound to cause damage. (19) A futuristic transition to a world of robo-doctors, for example, could never achieve perfect rates of patient recovery, and some patients' condition would inevitably deteriorate as a result of decisions made by algorithms. Litigants relying on strategic advice provided by ROSS Intelligence or the like, are also bound to lose cases or negotiations from time to time, while some of the individuals filing tax reports prepared by a virtual tax advisor would certainly still be subject to tax investigations and sanctions due to the algorithm's choices. Indeed, where such decision-makers have already been put to test--in the field of driving--we have witnessed that advanced algorithms are not immune to reaching damaging decisions. (20)
Should the traditional products liability framework continue to apply to the new generation of decision-making tools, (21) those that replace human discretion and enjoy rising levels of "autonomy" and self-learning abilities? (22)
Several scholars have argued that certain sophisticated or autonomous decision-makers require treatment different from their traditional predecessors.
Professor Jane Bambauer, for example, suggested that certain medical applications should be regulated similarly to how human doctors are regulated. (23) Attorney Jessica S. Allain too compared IBM's Watson units used in the medical field to a "consulting physician," (24) while attorneys Jason Chung and Amanda Zink likened it to a "medical student," explaining that Watson was not a typical medical device and that the products liability regime would not suit it. (25) Professor of Law and Health Science Ryan Abbott distinguished a conventional automobile from a driverless car, noting that it might warrant a separate treatment of scrutinizing its actions as compared with those of a reasonable human driver. (26) In a previous paper I too argued that, in general, algorithms replacing a human's professional judgment should be subject to the reasonableness standard that currently applies to humans, rather than being treated as a product. (27)
The European Parliament, to give another example, has issued a report to the E.U. Commission on Civil Law Rules on Robotics explaining that ordinary liability rules are insufficient for autonomous robots, since they can no longer be considered tools in the hands of other actors. (28) The report suggested granting autonomous robots an independent legal status of "electronic persons," which might even allow these robots themselves to pay damages for the harm they cause (for instance, through a compulsory insurance scheme developed for specific categories of robots). (29) The Committees on Transport and Tourism, on Employment and Social Affairs, and on the Environment, Public Health and Food Safety have all agreed that new liability rules ought to be developed to account for the new character of robotic decision makers. (30)
Indeed, several regimes of liability rules have been offered in the context of such autonomous robots or systems. Many of them focus on the similarities between these systems and human beings, and propose to apply similar legal treatment to both, either by subjecting systems to the standards of reasonableness as discussed above, (31) or by treating such systems as agents, (32) employees, (33) or children (34) for the purposes of applying tort liability on their manufacturers or users (similar to the tort liability that would apply to a principal, an employer or a parent, respectively). Other propositions focused on developing an insurance scheme adapted to the capabilities and potential danger posed by "sophisticated" or "autonomous" systems. (35)
In any event, a preliminary question not yet discussed in depth is when the system becomes different than a 'traditional product' such that products liability is no longer a sufficient framework to treat damages caused by it.
Automated machines of different kinds were described in detail as early as 800 years ago. (36) Algorithms themselves date back more than 2,000 years. (37) What is it that separates these and other traditional algorithms and machines that have been, and may continue to be, subject to products liability rules from what I will generally refer to as "thinking algorithms" that seem to warrant their own custom-made treatment? Why have auto-pilots, for example, been traditionally treated as products (38) while autonomous vehicles are suddenly seen as a more human-like system that requires different treatment? Where is the fine line drawn between products and decision-makers?
While several scholars have touched on distinguishing traditional from sophisticated technologies for the purpose of applying products liability, no in-depth discussion specifically on this question has yet been investigated. Rather, the discussions focused on different tort frameworks that ought to be developed for "sophisticated" or "autonomous" technologies, with only anecdotal references to what it is that renders such technologies "sophisticated" or "autonomous." Moreover, the potential parameters that have been mentioned to classify such technologies indeed related to the system's level of autonomy (while some referred to autonomy expressly, (39) other scholars discussed whether the system is able to wholly replace humans (40) or whether it outperforms humans (41)--all aspects are associated with the system's level of autonomy, as will be discussed below). This Article, however, demonstrates that autonomy is not a desirable differentiator between products and thinking algorithms, given its excessive complexity, the likelihood said differentiator will yield absurd or inconsistent results, and the non-practical nature of the results obtained using autonomy as a differentiator. A Roomba vacuum robot, for example, does replace a person in its action of cleaning; it (arguably) does so in a manner that outperforms the human cleaner, (42) and it possesses sufficient levels of autonomy that enable it to decide on its own which direction to...