If New York City mayor Bill De Blasio allows a new bill passed by the city council to go into effect, employers who use artificial-intelligence (AI) systems to evaluate potential hires will be obliged to conduct a yearly audit of their systems to show they are not discriminating with regard to race or gender.  Human resource departments have turned to AI as a quick and apparently effective way to sift through the mountains of applications that internet-based job searches often generate.  AI isn’t limited to hiring, though, as increasing numbers of organisations are using it for job evaluations and other personnel-related functions. 

Another thing the bill would do is to give candidates an option to choose an alternative process by which to be evaluated.  So if you don’t want a computer evaluating you, you can ask for another opinion, although it isn’t clear what form this alternative might take.  And it’s also not clear what would keep every job applicant from asking for the alternative at the outset, but maybe you have to be rejected by the AI system first to request it.

In any case, New York City’s proposed bill is one of the first pieces of legislation designed to address an increasingly prominent issue:  the question of unfair discrimination by AI systems. 

Anyone who has been paying attention to the progress of AI technology has heard some horror stories about things as seemingly basic as facial recognition.  An article in the December issue of Scientific American mentions that MIT’s Media Lab found poorer accuracy in facial-recognition technologies when non-white faces were being viewed than otherwise. 

Those who defend AI can cite the old saying among software engineers:  “garbage in, garbage out.”  The performance of an AI system is only as good as the set of training data that it uses to “learn” how to do its job.  If the software’s designers select a training database that is short on non-white faces, for example, or women, or other groups that have historically been discriminated against unfairly, then its performance will probably be inferior when it deals with people from those groups in reality.  So one answer to discriminatory outcomes from AI is to improve the training data pools with special attention being paid to minority groups.

In implementing the proposed New York City legislation, someone is going to have to set standards for the non-discrimination audits.  Out of a pool of 100 women and 100 men who are otherwise equally qualified, on average, what will the AI system have to do in order to be judged non-discriminatory?  Picking 10 men and no women would be ruled out of bounds, I’m pretty sure.  But what about four women and six men?  Or six women and four men?  At what point will it be viewed as discriminating against men?  Or do the people enforcing the law have ideological biases that make them consider discriminating against men to be impossible?  So far, none of these questions have been answered.

Perhaps the best feature of the proposed law is not the annual-audit provision, but the conferral of the right to request an alternative evaluation process.  There is a trend in business these days to weed out any function or operation that up to now has been done by people, and replace the people with software.  There are huge sectors of business operations where this transition is well-nigh complete. 

Credit ratings, for example, are accepted by nearly everyone, lendors and borrowers alike, and are generated almost entirely by algorithms.  The difference between this process and AI systems is that, in principle at least, one can ask to see the equations that make up one’s credit rating, although I suspect hardly anyone does.  The point is that if you ask how your credit rating was arrived at, someone should be able to tell you how it was done.

But AI is a different breed of cat.  For the newest and most effective kinds (so-called “deep neural networks”) even the software developers can’t tell you how the system arrives at a given decision.  If it’s opaque to its developers, the rest of us can give up any hope of understanding how it works. 

Being considered for a job isn’t the same as being tried for a crime, but there are useful parallels.  In both cases, one’s past is being judged in a way that will affect one’s future.  One of the most beneficial traditions of English common law is the custom of a trial by a jury of one’s peers.  Although trial by jury itself has fallen on hard times because the legal system has gone in for the same efficiency measures that the business world goes for (some judges are even using AI to help them decide sentence terms), the principle that a human being should ultimately be judged not by a machine, but by other human beings, is one that we abandon at our peril.

Theologians recognise that many heresies are not so much the stating of something that is false, as they are the overemphasis of one true idea at the expense of other true ideas.  If we make efficiency a goal rather than simply a means to more important goals, we are going to run roughshod over other more important principles and practices that have given rise to modern Western civilisation — the right to be judged by one’s peers, for example, instead of by an opaque and all-powerful algorithm. 

New York’s city council is right to recognise that AI personnel evaluation can be unfair.  Whether they have found the best way to deal with the problem is an open question.  But at least they acknowledge that all is not well with an AI-dominated future, and that something must be done before we get so used to it that it’s too late to recover what we’ve lost.

This article has been republished with permission from the Engineering Ethics blog.

Karl D. Stephan received the B. S. in Engineering from the California Institute of Technology in 1976. Following a year of graduate study at Cornell, he received the Master of Engineering degree in 1977...