MANAGING CORPORATIONS' RISK IN ADOPTING ARTIFICIAL INTELLIGENCE: A CORPORATE RESPONSIBILITY PARADIGM.

• Author: Chiu, Iris H.-Y.

INTRODUCTION 349 I. CORPORATIONS' ADOPTION OF ML 352 II. MAPPING THE LANDSCAPE OF RISKS FOR CORPORATIONS ADOPTING ML SYSTEMS 359 A. LEGAL RISKS 360 B. REGULATORY RISKS 365 C. REPUTATIONAL RISKS 368 D. OPERATIONAL AND FINANCIAL LOSSES 371 III. FRAMING CORPORATIONS' ML RISKS WITHIN THE CORPORATE RESPONSIBILITY PARADIGM 373 A. THICK AND BROAD CONCEPTION OF CORPORATE RESPONSIBILITY 376 IV. THE APPLICATION OF THE CORPORATE RESPONSIBILITY PARADIGM TO MANAGING ML RISKS 380 A. CORPORATE GOVERNANCE 380 B. ENTERPRISE-WIDE APPROACH 382 C. STAKEHOLDERS AND GATEKEEPERS 384 D. PROACTIVE MANAGEMENT 385 E. PRUDENTIAL PROVISION 386 F. TRANSPARENCY 387 V. CONCLUSION 389 INTRODUCTION

Accelerating developments are being observed in machine learning (ML) technology as the capacities for data capture and ever-increasing computer processing power have significantly improved. This is a branch of artificial intelligence technology that is not 'deterministic,' but rather one that programs the machine to 'learn' from patterns and data (1) in order to arrive at outcomes, such as in predictive analytics. (2) It is observed that companies are increasingly exploring the adoption of various ML technologies in various aspects of their business models, (3) as successful adopters have seen marked revenue growth. (4)

ML raises issues of risk for corporate and commercial use that are distinct from the legal risks involved in deploying robots that may be more deterministic in nature. (5) Such issues of risk relate to what data is being input for the learning processes for ML, the risks of bias, and hidden, sub-optimal assumptions; (6) how such data is processed by ML to reach its 'outcome,' leading sometimes to perverse results such as unexpected errors, (7) harm, (8) difficult choices, (9) and even sub-optimal behavioural phenomena; (10) and who should be accountable for such risks." While extant literature provides rich discussion of these issues, there are only emerging regulatory frameworks (12) and soft law in the form of ethical principles (13) to guide corporations navigating this area of innovation. This article intentionally focuses on corporations that deploy ML, rather than on producers of ML innovations, in order to chart a framework for guiding strategic corporate decisions in adopting ML. We argue that such a framework necessarily integrates corporations' legal risks and their broader accountability to society. The navigation of ML innovations is not carried out within a 'compliance landscape' for corporations, given that the laws and regulations governing corporations' use of ML are yet emerging. Corporations' deployment of ML is being scrutinised by the industry, stakeholders, and broader society as governance initiatives are being developed in a number of bottom-up quarters. We argue that corporations should frame their strategic deployment of ML innovations within a 'thick and broad' paradigm of corporate responsibility that is inextricably connected to business-society relations.

Section 1 defines the scope of ML that we are concerned about and distinguishes this from automated systems. We argue that the key risk that ML poses to corporations is unpredictability of results, (14) even if ML systems may perform efficiently and flawlessly most of the time. (15) Such unpredictability poses four categories of legal and non-legal risks for corporations, which we will unpack in Section 2: (a) risks of external harms and liability; (b) risks of regulatory liability; (c) reputational risks; and (d) risks of an operational nature and significant financial losses. These risks do not insularly affect corporations and their shareholders, as both often interact with a broader narrative in relation to business-society relations. Indeed, these risks pose broader consequences for business-society relations.

Section 3 anchors the risks depicted above in the narratives of business-society relations by first examining their impact on the social, economic, and moral realms and, secondly, arguing that corporations should navigate these narratives in a 'thick and broad' paradigm of corporate responsibility. (16) This Section explains that the 'thick and broad' paradigm of corporate responsibility is based on the perspective that integrates corporations into citizenship within the broader social fabric. The location of corporate management of ML risks in this paradigm compels corporations to internalise this socially conscious perspective and to shape their strategic and risk management approaches to ML risks accordingly.

Section 4 explores the applicational implications for corporations in addressing ML risks within a thick and broad corporate responsibility paradigm. We argue that the deployment of ML provides corporations with both the opportunity and the social obligation to carry this out with social discourse and expectations in mind. ML technologies can potentially usher in major institutional change, (17) and corporate behaviour and leadership in adopting ML should be more holistically interrogated. (18) Section 5 concludes.

1. CORPORATIONS' ADOPTION OF ML

   Businesses increasingly deploy artificial intelligence (AI) systems in finance, (19) healthcare, (20) taxation, (21) sales and marketing, (22) production and manufacturing, (23) and risk management. (24) While there are different definitions for what constitutes "AI", at its core, AI are systems designed to reason and act like intelligent and rational human beings for the purpose of attaining specified objectives. (25) The deployment of AI has evolved from the business adoption of automation, which has been ongoing since the 1940s. (26) Automation is deterministic in that machines complete tasks in a self-governing manner 'by means of programmed commands combined with automatic feedback control to ensure proper execution of the instructions'. (27)

   There is a relentless movement from 'automation' to 'autonomy' as machine development is steered towards ML. Machines would be elevated from slavishly performing pre-programmed commands to working out the most optimal and efficient routes to achieving performance. Such machines are programmed to process volumes of data within frameworks such as: 'natural language processing,' (2) " which allows human language expressions to be directly engaged with instead of translation into code; 'decision trees' (29) that allow pathways to information analysis and processing to be organised with statistical and consequential logic; or 'artificial neural networks,' (30) which simulate the human brain's associations and organise data in statistical but non-linear manners. ML processes data and recognises patterns within its learning frameworks in order to achieve certain outcomes and decisions. However, ML is still far from attaining 'super intelligence,' (31) the term used to describe AI able to replicate human intelligence. The development of AI is often discussed in three stages: narrow, general, and super.

   Narrow AI refers to the ability of computers to undertake specific tasks, such as learning the rules of chess in order to play it. (32) The machine is trained with the rules of the game and voluminous data relating to previous plays and moves, in order to work out the pathways needed for it to play or compete. (33) ML is able to devise more than one manner of pattern recognition in order to achieve outcomes, surpassing the programmed robot that operates on a precise, pre-determined sets of rules. (34)

   General AI is more ambitious, as it relates to machines with more 'holistic' or integrated capacity, simulating human reasoning that is more multi-faceted in nature. (35) Such a machine would not only be a chess player, Roomba vacuum, or facial recognition software, but more of an all-around android. Recent research given in conference proceedings shows that there has been only incremental development towards building general AI. (36) As the developments in communications robotics show, (37) general AI today remains rudimentary. An area of much-hyped development in general AI is that of self-driving cars, (38) as self-driving encompasses a number of different functions that, taken together, constitute the complex act of driving. General AI may attain greater human resemblance. However, in developing such general AI, a plethora of errors and hazards would have to be dealt with, such as the fatalities that have been caused by self-driving cars. (39)

   Super AI refers to AI that is indistinguishable from human sentience and capacity. Fiction provides us with a glimpse of the heights super AI may one day attain, such as the fiercely independent AI character Ava in Ex Machina (40) or a more benign AI personality as in the Japanese animation Time of Eve. (41) Super AI and humans would live side by side and would be almost indistinguishable except for the laws of robotics that govern android behaviour, such as laws safeguarding the superiority of humans. (42) As fiction uncannily shows, developments towards super AI would necessarily be underpinned by policy choices involving law, governance, ethics, and social considerations such as inclusion and cohesion.

   With scientific developments in the realm of narrow and possibly general AI, the corporate sector has been attracted to adopting the new capacities offered by such technology. This adoption has been incremental and focused on areas where there is strategic perception of a natural fit between ML and efficiency, revenue expansion, and cost reduction. (43) We provide a brief survey of corporate adoption of ML below.

   First, corporations are attracted to ML's potential ability to manage increasing data volume and overload, including for compliance or risk management purposes. Human management of voluminous amounts of data can result in errors caused by fatigue or negligence, while ML may offer more consistent performance. The question, however, is whether the performance of ML is comparable to that of...