Achieving ethical AI by narrow, technical means alone is an illusory goal. Broader societal measures are necessary to ensure that AI does not cause harm.
Assumption: Ethical principles can be encoded into AI.
Counterpoint: Achieving ‘ethical AI’ requires expansive measures that extend far beyond strictly technical fixes, including – potentially – uncomfortable organizational and societal reform.
States rightly recognize that it is impossible to implement AI successfully if they cannot set and enforce ethical principles. However, this assumes that ethical AI, as commonly envisioned, is feasible. In fact, the ‘translation to actions’ of ethical AI principles, as the authors of one report put it, ‘is often not obvious’. Certain ethical AI principles likely require measures to address not only issues with the technology itself but also with those who develop, govern and use it. This means that organizations and states that do not act equitably, transparently or accountably are unlikely to be equipped to embed those principles in the design, development, regulation and use of AI. True ethical AI might therefore require lengthy, systemic reform. All of which raises uncomfortable questions about whether ethical principles can be reconciled with the precepts of AI supremacy. A failure to engage with these questions risks derailing the enterprise of AI ethics entirely.
The technical challenges of AI ethics
The common principles of AI ethics are inarguably noble. Few would claim that it would be bad to have a high level of ‘fairness’, ‘safety’, ‘accountability’ and ‘transparency’ in AI. However, as is clear from the many instances of complex AI failure in recent years, as well as what we know to be inherently true of complex algorithmic systems, it is still far from certain that the technology itself can ever meet most states’ definition of ethical AI.
For example, it is common for policies to call for critical AI to be ‘explainable,’ meaning that they can be understood by those who interact with them (be they users or subjects of the system). And yet, the creation of explainable high-performance AI, particularly deep-learning models, remains an open research challenge – perhaps even a mathematical impossibility. Much AI that is marketed today as being ‘explainable’ does not actually meet this criterion in deployment; if it does, that explainability may very well be ineffective at enabling users to understand the system to an extent required by law, even in cases where those users are experts. Explainability may also often come at the expense of system performance. Meanwhile, it would be difficult to set broad standards for system understandability, given that every user’s capacity for understanding these systems will be unique. (None of which is a reason to abandon explainability – but rather a good reason to caveat any mention of explainability as a fix-all for AI understandability issues.)
The creation of explainable high-performance AI, particularly deep-learning models, remains an open research challenge – perhaps even a mathematical impossibility
Similarly, strategies that call for ‘fair’ or ‘unbiased’ AI assume that datasets and models can be expunged of bias. However, in any AI system where the data and the model do not perfectly represent the characteristics and dynamics of the environment in which they are used, bias can only be reduced. The non-expungeable biases that persist cannot always be easily quantified. This is especially true when the population with which an AI system interacts is constantly shifting. A dataset that originally exhibited no evident harmful biases may become much less representative if its target population experiences demographic or economic changes. This poses a challenge for drawing a measurable and enforceable regulatory line between acceptable and unacceptable levels of bias, not to mention for developing measures to offset these biases.
Meanwhile, ethical principles such as reliability and predictability cannot be technologically assured. No matter how large they are, datasets and models can only capture historical trends, patterns, phenomena and statistical distributions. Therefore it is rarely possible to guarantee that an AI system operating in an open, complex environment will not encounter inputs for which it is ill equipped to respond reliably or predictably. Even an approximated litmus test for the performance of a machine-learning tool would depend on validation datasets that are, as the EU’s proposed AI bill put it, ‘relevant, representative, free of errors and complete’. Such datasets for testing and validation could be beset by the same challenges that face the datasets on which systems are developed.
Nor is it clear that any amount of finite testing could identify all the ways that a system will experience either unforced errors or failures that are the result of intentional misuse by users. Here, too, trying to gauge a system’s unreliability or unpredictability (to determine whether it is above or below an acceptable threshold) is a challenge, since it is hard to develop concrete metrics for the degree of uncertainty regarding that system’s future behaviour. Rigorous AI regulations would likely hinge on a novel scheme of recursive review that goes beyond the existing processes that are currently in place for, say, new weapon systems.
Meanwhile, although third-party auditing is gaining traction as a potentially vital instrument for verifying whether an AI system is ethically aligned, the architecture for such audits remains skeletal. Gaps also remain in the processes for ensuring that audits result in organizations reforming their algorithms or the manner in which they are used. In the absence of major breakthroughs in audit practice and implementation, these will be, as the scholar Mona Sloane has written, ‘toothless’. (As with explainability, this is not a reason to abandon audits as a potential solution – but rather a cause to caveat any mention of audits as a fix-all.)
More fundamentally, much of the policy-level language of AI ethics assumes that a computer can replicate the ethical parameters that guide human decision-making. In reality, human ethical parameters cease to be ethical parameters, as such, when they are translated into the mathematically defined parameters that guide the outputs of the computerized system. There is a significant difference between the mathematically defined processes by which AI systems achieve their goals, for example, and the human capacity to address grey-zone cases, account for uncertainty and engage productively with ambiguity in decision-making. Concepts such as ‘fairness’ or system ‘trustworthiness’ therefore cannot be codified with concrete, testable numerical metrics that rate the degree to which an AI system is ethical. At best, machines can only offer an illusion of ethical behaviour – a computational mimicry of ethical decision-making that could fail at the first contact with something unexpected.