We’re all seeing those insistences around and about that artificial intelligence must be free from bias. We can only build that glorious new world of equity if this is so. Sadly, the actual truth is entirely the other way around. AI must be biased.
No, really.
We do need to distinguish a little here. As with that division into taste discrimination and rational discrimination. It is rational to rationally discriminate, obviously - someone numerate to do the accounts is an aid to getting the accounts done. But insisting that the numerate person is male or female - or other variation - or not is an irrelevance.
However, with AI the distinction has to be a little different. Not between rational and irrational beliefs at all - but between those held and those not held. We wish to use AI to make real world decisions. Therefore the model AI incorporates must describe the real world.
We don’t have to go far to find claims that Britain is institutionally racist, or grossly sexist, or biased in favour of property owners and many other such insistences. To the extent that any of those claims are true then our AI models have to be as racist, gendered and propertarian as the population and system is itself. Otherwise we’ll not be describing Britain. An AI that did not describe us in all our irrationality would be like the GOSPLAN executive pencilling in ten million pairs of leather shoes when only three cows have been planted to provide the materials for next year.
Say, for example, that the population believed all capitalists were top hatted cigar chompers (not true, some gnaw their cigars) and also that all socialists were wastrels about to run out of other peoples’ money (true, obviously). It’s not the truth or not of either statement which means that the AI must believe the same thing. It’s that Britons believe those things therefore an AI describing or prescribing for Britons must also do so.
The only useful AI is one as biased as the people it is being an AI for. Demands to examine the AI entrails to make sure that it does not incorporate population biases are therefore futile. For that very insistence upon not describing people means that the AI is worthless in making decisions for people.
It is possible to insist that AI not be racist, gender biased, but only if the insistence is equally that society itself is neither of those things. Which isn’t something that those shouting about making sure AI is not racist nor gender biased are going to agree to, are they? For the usual insistence is that we’ve got to make sure that AI does not incorporate these biases of society at large. Exactly the thing which will make AI useless.
We’re sorry about this but it is all true. The only useful AI is one that accurately describes the people it is being applied to - intellectual warts and all. Anything else is just a repeat of designing a system that requires New Soviet Man to turn up to make it work.