This post is the eighth installment of my passing on to both MindBlog readers and my future self my idiosyncratic selection of clips of text from O’Gieblyn’s book ‘God, Human, Animal, Machine’ that I have found particularly interesting. Here are fragments of Chapters 11 and 12 from the sixth section of her book, titled "Algorithm."
In the year 2001 alone, the amount of information generated doubled that of all information produced in human history. In 2002 it doubled again, and this trend has continued every year since. As Anderson noted, researchers in virtually every field have so much information that it is difficult to find relationships between things or make predictions.
What companies like Google discovered is that when you have data on this scale, you no longer need a theory at all. You can simply feed the numbers into algorithms and let them make predictions based on the patterns and relationships they notice…
“Google Translate “learned” to translate English to French simply by scanning Canadian documents that contained both languages, even though the algorithm has no model that understands either language.
These mathematical tools can predict and understand the world more adequately than any theory could. Petabytes allow us to say: ‘Correlation is enough,’…We can stop looking for models. We can analyze the data without hypotheses about what it might show. We can let statistical algorithms find patterns where science cannot. Of course, data alone can’t tell us why something happens—the variables on that scale are too legion—but maybe our need to know why was misguided. Maybe we should stop trying to understand the world and instead trust the wisdom of algorithms…technologies that have emerged .. have not only affirmed the uselessness of our models but revealed that machines are able to generate their own models of the world…this approach makes a return to a premodern epistemology..If we are no longer permitted to ask why…we will be forced to accept the decisions of our algorithms blindly, like Job accepting his punishment...
Deep learning, an especially powerful brand of machine learning has become the preferred means of drawing predictions from our era’s deluge of raw data. Credit auditors use it to decide whether or not to grant a loan. The CIA uses it to anticipate social unrest. The systems can be found in airport security software…many people now find themselves in a position much like Job’s, denied the right to know why they were refused a loan or fired from a job or given a likelihood of developing cancer. It’s difficult, in fact, to avoid the comparison to divine justice, given that our justice system has become a veritable laboratory of machine-learning experiments…In his book Homo Deus, Yuval Noah Harari makes virtually the same analogy: “Just as according to Christianity we humans cannot understand God and His plan, so Dataism declares that the human brain cannot fathom the new master algorithms.”
Hans Blumenberg, the postwar German philosopher, notes in his 1966 book The Legitimacy of the Modern Age—one of the major disenchantment texts—that theologians began to doubt around the thirteenth century that the world could have been created for man’s benefit…Blumenberg believed that it was impossible to understand ourselves as modern subjects without taking into account the crisis that spawned us. To this day many “new” ideas are merely attempts to answer questions that we have inherited from earlier periods of history, questions that have lost their specific context in medieval Christianity as they’ve made the leap from one century to the next, traveling from theology to philosophy to science and technology. In many cases, he argued, the historical questions lurking in modern projects are not so much stated but implied. We are continually returning to the site of the crime, though we do so blindly, unable to recognize or identify problems that seem only vaguely familiar to us. Failing to understand this history, we are bound to repeat the solutions and conclusions that proved unsatisfying in the past.
Perhaps this is why the crisis of subjectivity that one finds in Calvin, in Descartes, and in Kant continues to haunt our debates about how to interpret quantum physics, which continually returns to the chasm that exists between the subject and the world, and our theories of mind, which still cannot prove that our most immediate sensory experiences are real . The echoes of this doubt ring most loudly and persistently in conversations about emerging technologies, instruments that are designed to extend beyond our earthbound reason and restore our broken connection to transcendent truth. AI began with the desire to forge a god. It is not coincidental that the deity we have created resembles, uncannily, the one who got us into this problem in the first place.
Here are a smaller number of clips from the last section of Chapter 12, on the errors of algorithms.
It’s not difficult to find examples these days of technologies that contain ourselves “in a different disguise.” Although the most impressive machine-learning technologies are often described as “alien” and unlike us, they are prone to errors that are all too human. Because these algorithms rely on historical data—using information about the past to make predictions about the future—their decisions often reflect the biases and prejudices that have long colored our social and political life. Google’s algorithms show more ads for low-paying jobs to women than to men. Amazon’s same-day delivery algorithms were found to bypass black neighborhoods. A ProPublica report found that the COMPAS sentencing assessment was far more likely to assign higher recidivism rates to black defendants than to white defendants. These systems do not target specific races or genders, or even take these factors into account. But they often zero in on other information—zip codes, income, previous encounters with police—that are freighted with historic inequality. These machine-made decisions, then, end up reinforcing existing social inequalities, creating a feedback loop that makes it even more difficult to transcend our culture’s long history of structural racism and human prejudice.
It is much easier…to blame injustice on faulty algorithms than it is to contend in more meaningful ways with what they reveal about us and our society. In many cases the reflections of us that these machines produce are deeply unflattering. To take a particularly publicized example, one might recall Tay, the AI chatbot that Microsoft released in 2016, which was designed to engage with people on Twitter and learn from her actions with users. Within sixteen hours she began spewing racist and sexist vitriol, denied the Holocaust, and declared support for Hitler.
For Arendt, the problem was not that we kept creating things in our image; it was that we imbued these artifacts with a kind of transcendent power. Rather than focusing on how to use science and technology to improve the human condition, we had come to believe that our instruments could connect us to higher truths. “The desire to send humans to space was for her a metaphor for this dream of scientific transcendence. She tried to imagine what the earth and terrestrial human activity must look like from so far beyond its surface:
“If we look down from this point upon what is going on on earth and upon the various activities of men, that is, if we apply the Archimedean point to ourselves, then these activities will indeed appear to ourselves as no more than “overt behavior,” which we can study with the same methods we use to study the behavior of rats. Seen from a sufficient distance, the cars in which we travel and which we know we built ourselves will look as though they were, as Heisenberg once put it, “as inescapable a part of ourselves as the snail’s shell is “to its occupant.” All our pride in what we can do will disappear into some kind of mutation of the human race; the whole of technology, seen from this point, in fact no longer appears “as the result of a conscious human effort to extend man’s material powers, but rather as a large-scale biological process.” Under these circumstances, speech and everyday language would indeed be no longer a meaningful utterance that transcends behavior even if it only expresses it, and it would much better be replaced by the extreme and in itself meaningless formalism of mathematical signs.”
The problem is that a vantage so far removed from human nature cannot account for human agency. The view of earth from the Archimedean point compels us to regard our inventions not as historical choices but as part of an inexorable evolutionary process that is entirely deterministic and teleological, much like Kurzweil’s narrative about the Singularity. We ourselves inevitably become mere cogs in this machine, unable to account for our actions in any meaningful way, as the only valid language is the language of quantification, which machines understand far better than we do.
This is more or less what Jaron Lanier“warned about in his response to Chris Anderson’s proposal that we should abandon the scientific method and turn to algorithms for answers. “The point of a scientific theory is not that an angel will appreciate it,” Lanier wrote. “Its purpose is human comprehension. Science without a quest for theories means science without humans.” What we are abdicating, in the end, is our duty to create meaning from our empirical observations—to define for ourselves what constitutes justice, and morality, and quality of life—a task we forfeit each time we forget that meaning is an implicitly human category that cannot be reduced to quantification. To forget this truth is to use our tools to thwart our own interests, to build machines in our image that do nothing but dehumanize us.
Post a Comment