The House of Lords report on the implications of artificial intelligence is a thoughtful document which grasps one rather important point: this is not only something that computers do. Machine learning is the more precise term for the technology that allows computers to recognise patterns in enormous datasets and act on them. But even machine learning doesn’t happen only inside computer networks, because these machines are constantly tended and guided by humans. You can’t say that Google’s intelligence resides either in its machines or in its people: it depends on both and emerges from their interplay. Complex software is never written to a state of perfection and then left to run for ever. It is constantly being tweaked, increasingly often as part of an arms race with other software or networks that are being used to outwit it. And at every step of the way, human bias and human perspectives are involved. It couldn’t be otherwise. The dream of a computer system with godlike powers and the wisdom to use them well is a theological construct, not a technological possibility.
The question, then, is which forms of bias and which perspectives are desirable, and which we should guard against. It is easy to find chilling examples – the Google image recognition program that couldn’t distinguish between black people and gorillas, because it had been trained on a dataset where almost all the human faces were white or Asian; the program used by many American jurisdictions to make parole descriptions turns out to be four times as likely to recommend that white criminals be freed than black ones when all other things are equal. Without human judgment we are helpless against the errors introduced by earlier human judgments. This has been known for some time, but the report discusses these dangers very clearly.
One thing that has changed in recent years is that a lot of the underlying technology has been democratised. What had used to require the resources of huge corporations can now be done by private individuals, either by using the publicly available networks of Amazon, Google, and other giants, or simply by using cleverly designed software on private computers. Face recognition and voice recognition are both now possible in this way, and both will be used by malicious actors as well as benevolent ones. Most worries about the misuse of facial recognition software stem from their authoritarian use in places like China, where some policemen are already wearing facial recognition cameras, and concert-goers at large events are routinely scanned to see if they are of interest to the police. But the possibilities when they get into the hands of anarchists or apolitical bullies are also worrying.
We can’t step back into the past and we can only predict the future in the broadest terms. The committee is right to suggest principles, rather than detailed legislation. Since personal data can now be used for good and ill in ways that are impossible for the people from whom it has been gathered to predict, the benefits of this use need to be widely shared. The report is important and right in its warnings against the establishment of “data monopolies” where four or five giant companies have access to almost all the information about everyone, and no one else does. It is also prescient to identify “data poverty”, where people do not have enough of an online presence to identify them credibly as humans to other computer networks, as a threat for the future. But neither the problems, nor any solutions, are purely technological. They need political and social action to solve them.
