Many tasks previously reserved for humans are now done by machine learning. Passport control at international airports uses machine learning for passport recognition. An experienced immigration officer who examines one passport per minute might have seen four million faces by the end of their career. The machines were trained on fifty million faces before they were put into service. No wonder they do well.
Extraordinary benefits are being seen in health care. There is now a growing number of diagnostic studies in which the machines outperform humans, for example, in screening ultrasound scans or radiographs. Which would you rather be diagnosed by? An established human radiologist, or a machine with demonstrated superior performance? To put it another way, would you want to be diagnosed by a machine that knew less than your doctor? Answer: ‘No!’ Well then, would you want to be diagnosed by a doctor who knew less than the machine? That’s more difficult. Perhaps the question needs to be changed. Would you prefer to be treated by a doctor without machine learning or by a doctor making wise use of machine learning?
If we want humans to be involved in decisions involving our health, how much more in decisions involving our liberty. But are humans completely reliable and consistent? A peer-reviewed study suggested that the probability of a favourable parole decision depended on whether the judges had had their lunch. The very fact that appeals are sometimes successful provides empirical evidence that law, like any other human endeavour, involves uncertainty and fallibility. When it became apparent that in the UK there was inconsistency in sentencing for similar offences, in what the press called a postcode lottery, the Sentencing Council for England and Wales was established to promote greater transparency and consistency in sentencing. The code sets out factors which judges must consider in passing sentence, and ranges of tariffs for different kinds of crimes. If you like, it is another step in algorithmic sentencing. Would you want a machine that is less consistent than a judge to pass sentence? See the sequence of questions above about a doctor.
We may consider that judicial sentencing has a special case for human involvement because it involves restricting an individual’s freedom. What about democracy? How should citizens decide how to vote when given the opportunity? Voter A may prioritise public services, and she may seek to identify the party (if the choices are between well identified parties) which will best promote education, health, law and order, and other services which she values. She may also have a concern for the poor and favour redistributive taxation. Voter B may have different priorities and seek simply to vote for the party which in his judgement will leave him best off. Other factors may come into play, such as the perceived trustworthiness of an individual candidate, or their ability to evoke empathy from fellow citizens.
This kind of dilemma is something machines can help with, because they are good at multi-objective optimisation. A semiconductor industry might want chips that are as small as possible, and as fast as possible, and consume as little power as possible, and are as reliable as possible, and as cheap to manufacture as possible, but these requirements are in tension with one another. Techniques are becoming available to enable machines to make optimal decisions in such situations, and they may be better at them than humans. Suppose that a machine came to know my preferences better than I can articulate them myself. The best professionals can already do this in their areas of expertise, and good friends sometimes seem to know us better than we know ourselves. Suppose also that the machine was better than me at analysing which candidate if elected would be more likely to deliver the optimal combination of my preferences. Might there be something to be said for benefitting from that guidance?