Artificial intelligence is a buzz tool among scientists today. Researchers in many fields, from medicine to sociology, rush to use methods they don’t always understand. This causes a wave of false results, a crisis of reproducibility of results and makes science look like “magic”.
One example of such a crisis is given by Princeton professor Arvind Narayanan and his graduate student Sayash Kapur. Last year, they uncovered political science research that claimed AI could predict when civil war would break out with more than 90 percent accuracy. History shows that civil wars are among the messiest and most horrific human affairs.
A series of articles describes the surprising results of the use of artificial intelligence. The AI’s analysis of data such as a country’s gross domestic product and unemployment rate outperformed conventional statistical methods in predicting the onset of civil war by nearly 20 percentage points.
But when the Princeton researchers took a closer look, many of the results turned out to be mirages. Machine learning involves artificial intelligence learning to predict the future using data from the past.
However, in several works, researchers have not been able to properly separate the datasets used to train and test the performance of the code. This error is called a “data leak,” which causes the system to test against data it has seen before. Like a student taking an exam already having the answers.
When researchers corrected these errors, they found that artificial intelligence offered virtually no benefits.
The experience prompted the Princeton pair to investigate whether the misuse of machine learning distorts results in other fields. He concluded that the misuse of artificial intelligence is a widespread problem of modern science.
Artificial intelligence is considered a potentially revolutionary tool for science because of its ability to analyze large amounts of information and reveal patterns that are difficult to discern through traditional data analysis. Researchers have used artificial intelligence to make breakthroughs in predicting protein structures, controlling fusion reactors, and exploring space.
However, Kapur and Narayanan caution that the impact of artificial intelligence on scientific research has in many cases been small. Other researchers found errors in 329 studies that relied on machine learning in various fields.
Late last month, Kapur and Narayanan organized a workshop to draw attention to what they call the “crisis of reproducibility” in science that uses machine learning. They had hoped for 30 or so attendees, but received more than 1,500 registrations, a surprise they say shows that the problems with machine learning in science are widespread.
Momin Malik, a data scientist at the Mayo Clinic, points to a prime example of machine learning producing misleading results: Google Flu Trends. It’s a tool developed by a search company in 2008 that aimed to use machine learning to more quickly identify flu outbreaks based on the search queries of ordinary users.
Google received positive publicity for the project, but it failed to predict the course of the 2013 flu season. An independent study later concluded that the model was locked into seasonal conditions that had nothing to do with flu prevalence.