There is no doubt that the new technological breakthroughs will change the face of scientific research. A new article points out three failures to watch out for
Uri Fogel, Davidson Institute for Science Education, the educational arm of the Weizmann Institute
In June of this year, the film "The Matrix" celebrated its half-anniversary of its release. At the very beginning of the film, his hero, Neo, discovers that the world he lives in is controlled by machines with artificial intelligence, which rebelled against their creators and enslaved the human race. When the film first came out, viewers flocked to the cinemas in droves, as they have done in countless other films about artificial intelligences that have gone out of control. But today, as artificial intelligence deepens its penetration into our lives, the understanding sharpens that the dangers involved in the use of artificial intelligence may be less dramatic and much more elusive.
An example of these dangers can be seen in the field of scientific research. Similar to many other fields in the economy and society, the use of artificial intelligence tools in scientific research is increasing. In response to the rapid developments in the field, the psychologist Molly Crockett and the anthropologist Lisa Messeri from the United States recently published A joint opinion piece in the journal Nature, where these failures were mapped.
In their article, the two examined the vision presented by other scientists in relation to the research work supported by artificial intelligence, in order to assess what the scientific work might look like in the near future. They depicted a picture of a future in which the classical roles of the scientist will completely pass into the hands of artificial intelligence. The artificial intelligence will process extensive scientific literature, and present the most interesting research questions. Instead of collecting data from nature or man, we can generate artificial data using simulations created by artificial intelligence. The conclusions we draw from the data will all be based on artificial intelligence assessments. And the icing on the cake - the artificial intelligence will check and criticize new articles that are proposed for publication.
"My immediate reaction to these predictions was, 'Did they slip up?' Crockett admitted in an interview For the online magazine Ars Technica. "But we didn't just choose the articles out of a whim. These are the things that senior scientists are saying about the future of artificial intelligence."
Benefit with harm on the side
The advantages offered by artificial intelligence for scientific research are clear: among them is a tremendous saving in money and working hours, and opening a gate to new ideas that are not limited to the way of thinking of the human brain. However, Crockett and Masri claim that this benefit comes with a difficult price tag: damage to our understanding of reality.
The article is based on studies that examined the degree of trust of researchers in artificial intelligence tools. Based on them, the authors defined three types of illusions that scientists tend to follow in their work with such tools. The first type is the illusion that we can explain the findings obtained. The artificial intelligence tools are based on complicated calculations and finding incomprehensible relationships between parameters. Researchers who use these tools are indeed experts in their field of research, but this does not mean that they are necessarily experts in the algorithm itself - the collection of decision-making rules of the artificial intelligence software. As soon as a researcher holds findings in his hands, he will tend to adopt them as fact and forget that the way in which they were obtained is not transparent to him.
The second illusion is that the artificial intelligence examines all the available options. In practice, she learns limited data, which is derived from the scope of the database on which we trained her and the nature of the information. Therefore, its conclusions will also necessarily be limited. For example, behavioral researchers who are looking for insights into human nature An artificial intelligence model may be trained based on materials from the Internet. Therefore, their model will not include characteristics that are not reflected on the Internet. And artificial intelligence trained on a medical database will not conclude anything about indicators that do not appear in this database.
The third illusion is that AI tools are objective, while humans have biased assumptions. As man is the template of his native landscape, so the artificial intelligence is the template of the data landscape it used for its learning. Amazon company, for example, noticed that its algorithm for scanning resumes favored male applicants. The algorithm simply gave a higher score to resumes in which appeared words that are more widely used by men than by women, and it did so based on the database with which it was trained - a database that included the choices made by humans before it.
Under the influence of these illusions, we may find ourselves with a degenerate and less diverse science, without realizing that there is a problem with our research. One of the strategies that Crockett and Masri suggest is working in teams that combine diverse research disciplines. This way the danger will be reduced that they will focus on only one dominant line of thought. It should be emphasized that they really do not rule out work with artificial intelligence, since its benefit to science is obvious to everyone. "We must teach ourselves the ways in which artificial intelligence may endanger the process of creating scientific knowledge," Crockett concludes. "Scientists working alone will not be able to help us reduce these dangers."
More of the topic in Hayadan:
