Algorithms are Gods

What's the worst way to begin a newsletter?

Invoking a Latin phrase is surely up there. And yet here we are, staring at three fusty words in a dead tongue:

Deus ex machina.

It's a phrase I have thought a lot about this week, "god in the machine." Originating in Ancient Greece, the term was first used to describe the work of the playwright Aeschylus, later that of Euripides. It described a dramatic device in which the main action of a play is resolved by calling on a deity. In Medea, for example, the titular character is rescued by Helios, the sun-god, who appears apropos of nothing to whisk Medea to safety. To give the entrance of these almighties their due, productions usually leveraged "machines" of some sort — cranes, or trapdoors. Hence the name, and more truthful translation, "god from the machine."

Over the past two decades, the deus ex machina mantra more accurately fits the tech industry rather than the dramatic arts. The ambition of iconic founders has been to embed the divine in lines of code, to put god in their machines. Uber, notoriously, built "God View," which gave viewers total visibility into the movement of passengers in a given city, while Sergey Brin, founder of Google, once said, "[T]he perfect search engine would be like the mind of God."

For twenty years or so, that may have been the goal. But an apotheosis has occurred. We once dreamt of building god in the machine. Today, the machine is god. Our lives are ruled by invisible algorithms that decide what we buy and read and watch, what healthcare we are given, what job offers we receive, whether or not we go to prison. Amazon's algorithm is said to influence a third of our decisions on the platform. Netflix drives 80% of our choices. These are minor examples, but as Francis Bacon noted "God hangs the greatest weights upon the smallest wires."

These gods are, of course, nothing like the one Bacon would have recognized. Unlike man, supposedly made in god's image, these deities have been made in ours. Which is to say that they are unjust and unfair, prejudiced against those most in need. Last week, we alighted on this subject to note that technology can be better used to correct racial imbalances: we should begin with algorithmic affirmative action.

Women and minorities endure systemic biases from technology. That may not surprise, but it should still sting.

COMPAS, a system used in courtrooms across the country to predict the likelihood of an individual committing further crimes, was shown to be biased against African-Americans. These "risk assessments" informed bond amounts and decisions about which individuals were set free.

An algorithm used by healthcare systems serving millions of patients graded black patients as lower risk than white patients with similar ailments. The result is that black patients received less than half the additional care they would have if they had been white.

A fintech platform offering mortgages charges African-Americans and Latinx users 5.3 basis points more than white applicants. Refinancing an existing mortgage costs these minorities 2.0 basis points more.

So: money, health, freedom. Each fundamentally altered by the decision-making of machine learning models. While in some rare cases, prejudiced algorithms may be the product of some rogue, race-baiting developer, the more common truth is less vicious, harder to solve. More often than not, algorithms suffer because of the data on which they rely. Or as the Yale Law Journal summarizes, "Bias In, Bias Out." The world is biased — how could algorithmic reflections of it be any different?

As with anything of sufficient complexity, solutions are not simple. Some have suggested unveiling the workings of each "black-box," laying bare the cogs and wheels by which decisions are made. That way lies madness. Not only would it require trade secrets to be revealed, it would also allow bad actors to game the system more effectively. If Google were to expose the inner workings of its page ranking algo, it would become much easier to spam, potentially exacerbating the problem.

A better place to start may be in reframing what we expect from these systems, what we consider fair. A change in mindset is required: while most algorithms are "procedurally" fair (which is to say, they operate fairly when the outcome is excepted), they are rarely "substantively" fair. That is, they do not produce just results, regardless of the methodology used.

How do we make this shift? How do we ensure that even in the digital dominion, the moral universe bends towards justice?

Above all, it may be a question of paying attention to the inputs. Just as a "Prince who is not himself wise cannot be well-advised" (Machiavelli), so too can an algorithm with incomplete data fail to be anything but flawed. Recent research from MIT has shown how increasing the diversity of data can lead to less biased outcomes. This new approach is able to determine how much more data is needed from a certain group to improve the accuracy of predictions related to that population. For example, the system might suggest adding data from Asian-Americans to better predict mortality rates for Asian-Americans in an ICU environment. Augmenting the data set does nothing to diminish the efficacy of predictions for other groups.

For all of its idealism and professions of tolerance, tech is unlikely to make such changes unbidden. Real regulation will need to be levied. Calls to form a sort of "FDA for algorithms," designed to evaluate the cost/benefit of different models is an interesting suggestion, as is the meat of the "Algorithmic Accountability Act," which is still under consideration by Congress. The bill requires companies to "study and fix flawed computer algorithms."

The challenge in all of this will be securing justice without neutering innovation. In that endeavor, the private sector may be invaluable. Will data-collection companies emerge to serve different ethnic and affinity groups? Could a data-munger create proportionate data sets by nation, city, county to ensure adequate representation? Can "prejudice-detection" systems be used to alert companies of algorithmic bias? Asking for more, asking for better does not need to stifle invention.

A phrase is inscribed on the gravestone of famed psychoanalyst, Carl Jung. Latin, of course, though we need not dwell on those old words. In English, it reads, "Summoned or unsummoned, god is present." Whether we like it or not, our world is governed by an invisible, nearly omnipotent power or set of powers. It is up to us to ensure that we create deities of which we can be proud, godly not only in their might but in their mercy.