It is decisive that decisions by algorithms can be incomplete. It is also important that human decisions can be quite incomplete. But the defects of these two types of decisions are probably not the same. How should society use decision -making by and when and when? Cass R. Sunstein provides careful outline “Use algorithms in society” (Austrian economy reviewDecember 2024, 37: 399–420).
Sunstein emphasizes two significant advantages of decision -making by algorithms. You can avoid prejudice and noisy. Prejudice occurs when a specific judge or a doctor is too influenced by a particular factor, such as seeing a similar event. The noisy is only a human attribute of the inconsistency, and if the weather is good, bad, or lunch, the decision rule can be applied in different ways. (In relation to the possible roles of prejudice, noisy, and algorithms, the outline of decision making of jewelry cases is the useful starting point as follows. Crystal S. Yang and Will Dobbie’s “American Trial System: Balance of Personal Rights and Public Interest In the fall of 2021 Economic perspective journal, Where I work as an editor -in -chief.)
Here, I would like to focus on some reservations that Sunstein summarizes the algorithm decision -making. Some of them can be overcome with a better system over time.
1) Even if the algorithm performs better than most humans, some humans do better than the algorithm. Sunstein is used as follows:
Some important studies suggest that in the context of jewelry decisions, algorithms surpass 90%of human judges, but the top 10%of the judges are better than the algorithm. The reason is that the best judge uses and uses personal information to make better decisions. They consider factors without algorithms. They seem to have local knowledge -a situation that lacks an understanding or algorithm of the defendant. We could easily imagine a similar discovery with doctors. The best doctor may know someone who will test heart disease because it sees something that the algorithm does not consider or looks intuitive.
2) In many situations, profits using algorithms are in proportion, but the benefits of the ratio that apply to many people can be significantly meaningful. Sunstein is used as follows:
As I mentioned, the algorithm is better than people, but not magnificent. In terms of welfare benefits, the impressive total figures come from the fact that a very large population is involved. If the algorithm shows an increase in the ratio of less accurate compared to humans, we can find a big improvement. If the algorithm can increase the accuracy of the screening test for heart disease slightly, death can be greatly reduced. But slightly increasing of accuracy remains. For example, the formula is better than the clinic. However, in intermediate studies, the formula is 73%of the time and 68%of the time of clinical trials.
3) A person cannot forgive the algorithm error more than human errors. Sunstein:
In short, people are less forgiven than the algorithm.
Human. … Is that reasonable? If people want to make the right decision, it’s not. If their goal is to make money or improve health, you must rely on a better determinant. But once more: If people like to make decisions, prefer to make their own decisions can be perfectly reasonable. Perhaps people can have fun with related decisions. Perhaps they like to learn. Perhaps decision -making is a kind of game. Perhaps they like responsibility. Perhaps they will love the real castle of responsibility. If so, hate is not a mistake at all.
4) If the complexity increases, the benefits of the algorithm are limited. For example, the algorithm behind the date app can provide recommendations that may be slightly successful than those who are not recommended, but they do not guarantee true love. Sunstein tells an interesting story about the predictive competition called The Fragile Families Challenge. This teaches humility about the prediction of the algorithm.
There is a data set of fragile families and children’s well -being research. My mother collected data for thousands of unmarried parents who had a child in 2000.This studies have collected many data for this family at the time of birth and then traced 1, 3, 5, and 9 years old. As a result, many predictions were similar to each other, but literally either was very accurate. Sunstein is used as follows:
The central question was simple. Which of the 160 teams will make a good prediction? The answer is as follows. In fact, machine learning algorithms are better than random. They were not terrible. But they are not much better than random, and the results of single events, such as the first caregiver are fired or are being trained, are slightly better than random. The researchers concluded that “predictive accuracy with low predictive accuracy could not be due to the limitations of certain researchers or approaches, hundreds of researchers attempted this task and no one could predict exactly.” The 160 teams produced close predictions in spite of various ways. As the researchers said, “Submission was much better to predict each other than to predict the truth.” A reasonable lesson is that we do not actually understand the relationship between a year’s position and a few years. … You can now learn a lot about where someone is in life, but you may still be unable to say at all.
For the future certain results.
5) Algorithms are not good and sometimes cannot be called path -dependent events. When a particular political movement occurs or falls, or whether certain music behaviors or films will be popularized, they will depend on the development of a series of events. Algorithms depend on the patterns of past events and may not be good at predicting the timing or probability of the future event chain.
The task is to consider the benefits and trade offs of algorithms in various settings. If I do not know who my loved one is facing jewelry decisions and who the judge is, I personally prefer the algorithm to make a decision. Regarding the algorithm that manages autonomous cars, many forgive the errors and thinking of the algorithms than the human driver’s errors and accidents. In many contexts, from love to future possibilities, the guidelines for following algorithms can be positive but very small.
For me, the algorithm is always interesting because it specifies the reason for the basic decision. Sometimes the reason will reveal prejudice and noise in human decisions. Sometimes the algorithm itself can display bias. But if you know why, you can make the decision more clearly.