The World Economic Forum, in 2016, published its list of the top 10 ethical issues in AI, which is still an excellent summation of the most hotly debated topics about the impact of AI and machine learning today in modern society. Some of the ethical issues on the list are grounded in everyday realities (e.g., unemployment and racism); others will seem futuristic to many, and may never even materialize (e.g., evil genies or the Singularity). Nevertheless, there’s no doubt the adoption of AI is spreading across a range of industries, and there are many issues, both ethical and legal, that researchers and practitioners are dealing with today as a consequence. The following are a few broad themes in the legal and ethical spaces.

Legal Issues

There are very few legal precedents for AI applications, and that makes the legal dimension of this technology particularly challenging for those working in the field. AI shares some common elements with other technologies, but there are several unique aspects that add more complexity to traditional legal frameworks.

Since data is the foundation of any AI model, proper laws that deal with the ownership of that data, and data privacy, are extremely important. Current regulatory regimes are mainly concerned with what data is collected—and how it’s collected, used, and retained. Recent media attention around data security and the misuse of personally identifiable information has certainly increased public and industry awareness around these issues, and recent regulatory initiatives, such as the General Data Protection Regulation in Europe, are definitely steps in the right direction. However, end users’ awareness of (and attention to) these laws is lacking. Quickly scrolling to the end of a licence agreement, and blithely accepting its privacy policies without reading them, when signing up for a service or installing an app is the typical user action.

Data breaches also expose firms to liability. (Interestingly, the SEC considers safeguards and privacy protection of client records and information to be within the scope of a fiduciary’s responsibilities.) Moving beyond the data itself toward the output of AI models, determining liability becomes more difficult. Traditionally, the determination of liability depends “on attribution of fault by application of causation principles.”¹ In other words, liability depends on assigning fault to factors based on the legal principles of causation. This is somewhat straightforward in the context of hardware and software products.

The problem with AI is that the algorithm’s decisions could be far removed from any direct programming as the model adapts to new data and continuously optimizes. Take autonomous vehicles, for example: when there’s a fatal accident, and the car model has been significantly changed from its original version, who, exactly, is liable? While model interpretability and explainability—two distinct but closely related terms—may help to determine liability, they’re only part of the solution.

Another legal consideration is that of intellectual property (IP). This is a standard concept in traditional products (e.g., software and drugs), but it’s more nuanced when it comes to AI. The typical lifecycle of an AI model involves the data, the algorithm, and the optimized model that trains the algorithm on a specific dataset. Consider the following: an AI platform vendor collaborates with a company that owns the data to solve a specific problem—say, a system to provide personalized customer recommendations based on customers’ usage patterns. Who would own the IP rights to a prediction model that has trained a proprietary algorithm on the clients’ dataset? Given the lack of clear legal guidelines, and variations in different jurisdictions that could be applied uniformly, companies need to think ahead by clearly specifying who owns the IP and the work product in any contract they enter into.

Given these unique challenges and new conundrums on the applications of AI, there’s potential for AI to spur the creation of new, specialized practices for law firms, with experts who focus exclusively on AI legal issues. As AI becomes even more prevalent, we hope that policy-makers will also take up these issues and establish laws that would provide guideposts for practitioners.

However, there’s a parallel dimension of AI—the ethical—that may never get due attention by regulators. And each industry or profession, as a consequence, would be responsible to create and adhere to an ethical framework or code of conduct.

Ethical Issues

Ethical issues are stacked similarly to legal ones, with data at the core. Statisticians and quantitative analysts have always encountered biases in data, which are also a common occurrence in AI. All of the traditional biases—like look-ahead, sample selection, overfitting, and underfitting—apply when training AI models. After all, the only advantage AI has over its human counterparts is the sheer strength of its computing power; data preparation still requires human judgment.

Another bias AI models are prone to is that of an unbalanced dataset. Think of this pitfall as a different flavour of the sample selection bias. Consider fraud detection in credit card transactions, for example: the complete population of the dataset would inherently be biased toward non-fraudulent transactions, with a very small proportion of fraudulent transactions. This results in an unbalanced dataset and, by extension, a model biased toward non-fraudulent transactions. Without balancing the underlying dataset, the AI models created would have low accuracy and be completely unreliable.

Some of these biases are somewhat technical and may require the expertise of a data scientist. There are others that only require sound judgment and common sense to avoid, such as AI models that could be biased toward factors such as race, ethnicity, age, and gender. Making credit decisions based on default predictions by looking at a person’s race, for example, would result in a discriminatory prediction model, holding other factors constant. For the AI model, race simply becomes a factor that gets a higher weight for reliable predictions, but it’s morally binding on anyone relying on these predictions to ensure such biases are weeded out from the underlying dataset.

Data privacy and ethical use of data collected from end users also pose similar challenges that go beyond legal issues. There’s no surprise that profit-maximizing firms operating in different jurisdictions would exploit all opportunities to monetize their data to the fullest—whether it’s through targeted advertising based on listening in on users’ conversations, or extracting leads from social networks or digital address books. But, with no legal framework or ethical guidelines in place, we should never abandon our moral high ground.

These issues are by no means exhaustive, and others are likely to emerge as AI technology matures. Some of them may be completely unimaginable right now, and those that have been conjured up may never come to fruition. But, as with other technological innovations, we can be certain the adoption of AI will be accompanied by ambiguity, uncertainty, and rapid growth. And, as in anything we do, we must let our moral compasses guide us.

¹ https://www.osborneclarke.com/insights/artificial-intelligence-the-real-legalissues-an-article-by-john-c-buyers-osborne-clarke-llp/