Steve Gimbel
In 1903, W.E.B. Dubois published “The Talented Tenth”, arguing that the top 10 per cent of African Americans – he later generalized this to all groups—should dedicate their lives to learning becoming the teachers, leaders, and decision-makers. Their superior intelligence should be used to improve life for everyone.
This view’s elitism diminishes the majority. Egalitarianism asserts that everyone should be a decision-maker. Those outside the top tenth know what is important to them and should have the right to choose, even if they so do badly. Your mistakes are yours to make.
Cass R. Sunstein’s latest book, Imperfect Oracle: What AI Can and Cannot Do, revives the ideas of a talented tenth and the right to make your own mistakes. Some think AI is an emotionless, rational saviour delivering truths our brains would never discover. Others see it as a mortal threat, robbing us of autonomy. Sunstein avoids both extremes. He shows what the technology does well, contrasting it with what we do not, carefully setting out the costs and benefits of surrendering decisions to the black box.
There are benefits. With complex judgments, such as deciding which defendants should be remanded into custody, artificial intelligence is better than 90 per cent of human judges. When diagnosing patients, AI is more accurate than nine out of ten physicians. The machine’s decisions decrease crime, save lives, and make life better. There remains a talented tenth, the best in each field who still outpace the technology, but AI itself becomes a part of the new talented tenth. We are again confronted with Dubois’s question of whether to turn over a culturally superior position to it.
Sunstein points out that there are cases in which we need to strictly limit its use, where errors are beneficial. Consider education. As a philosophy professor, I can tell you that Chat GPT does undergraduate level work. The initial model produced B-essays, but the latest version has improved to B+ work.
Many students would be delighted with a B+ and see what comes out of the LLM as better than what they would have written. But the point of the assignment is not to give the right answer; rather it is to wrestle with the problem, see the sticky issues, and have an insight. Even if we err, we can learn from our mistakes. It is the process, not the product, that is important.
But when we want the best answer – say, when picking a mutual fund from among the options offered by your employer – then it seems we should turn to the machine. Yet we remain hesitant. Seeking advice from a friend, we can ask them to explain their thinking. But all we get from AI is output. This makes us uneasy, especially when the result affects a person’s life. The computer does not have empathy, cannot show mercy, and the coldness is worrisome.
Imperfect Oracle: What AI Can and
Cannot Do, by Cass R. Sunstein
(Philadelphia, PA: American
Philosophical Society Press, 2025,
ISBN 978-1606181379), 232 pp.,
US Dollar : 26.95.
Yet, Sunstein contends, our feelings are affected by biases that lead to unjust results. Yes, AI also can be biased. Sunstein explains how it appears, from being trained on biased information to not being concerned with equality of results. But human bias can result from the noisiness of our decisions, which is not a problem for AI.
My optometrist has a new machine. Before being ushered back to the big chair; I put my chin in the cup, see a blurry image, hear a whirring and suddenly see in perfect focus. The machine reads the curvature of my eye, correcting for it. I would love to see the way I do in the machine.
But I am sent back for the ritual of reading the lowest clear line and determining if it is better “now” or “now.” At my last visit, I stopped the process, moving from patient to philosopher. I asked my optometrist if he repeated options, asking the same two choices in random order. He said he always does, I asked if patients ever prefer A to B one time and then B to A the next. He confirmed that happens almost every single time.
That is noise: same in, different out. Humans are noisy, but not the machine. It gives the same result every time. Noise arises for multiple reasons, including cognitive biases from our psychology. In their book Nudge, Sunstein and Richard Thaler argued that because of these cognitive biases, we need to take charge of how options are presented to us. Make it so that we have to, say, opt out of saving a portion of our paycheck instead of actively having to put money away. Because we are not perfectly rational, we need to manipulate ourselves to do the smart thing.
Imperfect Oracle is a sequel to Nudge in which we now have a technology that can be an e-nudge. Instead of arranging how we choose to more likely get the rational result, should we use the rational choice making machine to make the decision for us? Is this surrendering our autonomy?
Yes and No. We would not be the decision-maker. Sunstein points out that there is both a first-order decision—the choice at hand-and a second order decision—who makes the choice. Allowing AI to decide may harm your first order autonomy, but if you decide if it does, then you are embracing a second-order freedom. The question of using artificial intelligence, Sunstein argues, is a second order issue. This transforms the conversation, and the end of the book is a thoughtful discussion of the issues to be considered in making a second-order decision.
As a philosopher, I live in my head, removed from the material world. I’m not good at judging according to my senses. When a friend took me wine tasting, my answer to his “Well?” was, “Yes, this tastes like wine, too.” I don’t appreciate subtle differences.
That applies visually. I sit flustered in the optometrist’s chair, unsure if the letters are clearer before or after. I am certainly not in the talented tenth for this skill. I want to use my second-order freedom to decide but rather to let the machine determine my prescription. That would make me happy.
Sunstein argues that this is what the discussion around artificial intelligence ought to look like. This is the sort of focus his conceptual lenses lend to this debate. Imperfect Oracle brings into view the operative elements needed to approach this topic clearly. It is a book worth reading and I plan to reread it as soon as I can figure out what I did with my reading glasses.
Courtesy: Free Inquiry
