Humanising machine intelligence

Rather than being an unalloyed and unstoppable threat, the ethics of Machine Learning (ML) offer the world an opportunity to make progress on challenging ethical problems.

Below I will elaborate this opinion and argue why I believe that the ethical challenges arising from ML can not only be met, but can also improve the way we approach a range of ethical problems.

ML is a technology that involves the mathematical analysis of patterns in data, implemented via computer programs. Now deployed at scale, it underpins the modern web, and is becoming increasingly important to all areas of human activity.

Like all technologies, ML affects people, and this is where ethical issues arise. No technology removes human choice. The ethical consequences of ML are primarily a consequence of how the technology is used, and are not intrinsic to the technology. I illustrate with two examples.

Suppose an ML system was used to predict (very accurately) who would do well at university on the basis of a wide range of information. This could be used to allocate places (this is a choice in itself, not an inevitability).

But if that same system allocated 70 per cent of places to women and 30 per cent to men (or vice versa), many would decry its ‘unfairness’. But what is meant by ‘unfairness’ in this case?

My recent work on this has led to some interesting conclusions regarding this problem. There is no single notion of fairness – just think of equality of opportunity versus equality of outcome, for example. The latter can be readily measured. But the former has arguably greater moral legitimacy.

ML offers the opportunity to precisely quantify the trade-offs implied by particular choices of fairness. That there is a trade-off is no surprise to engineers or to philosophers – trade-offs are pervasive in the art of design, and to any plausible moral philosophy.By insisting on an exact mathematical formulation, one finds there is a best-possible trade-off, and furthermore that this is independent of any choice of algorithm or method, depending instead only on the underlying data, which is of course merely a reflection of the world itself (including all the choices made earlier about what data was collected).