believe. How many characters are on the famous album cover Sgt Paper’s Lonely Hearts Club Band Of the Beatles? You can give an estimate with the naked eye, say, between 40 and 60. Now, imagine that a computer vision algorithm makes its decision and declares 70. Chances are you will fall for specialization. From the program. Three researchers in information systems management at the University of Georgia in the United States conducted a series of such experiments to show that people actually followed a response provided by an algorithm rather than a third party, which Want to help you His article appeared in the Journal’s Scientific Reports on April 13, 2021 Nature.
A computer program is not tired, never shows sick, has no bias and remains rational under all circumstances. It is, in any case, more or less conscious on these arguments, that this belief rests. All the more so as to count as neutral when it comes to an operation.
A hypothetical algorithm trained on 5,000 images
The team designed three experiments based on the same principle: asking participants recruited from the online platform Amazon Mechanical Turk to evaluate the number of people present in the ten photographs, ranging from the simplest (15 people) to the most difficult (5,000 People) practice till. Participants give a number. Researchers then present two estimates, one considered from the algorithm, the other being an average of estimates given by 5,000 others. And researchers ask participants whether they stick to their initial response or rely on one of the two suggestions.
Participants know only one thing about the algorithm: it was trained based on 5,000 images. But they do not know what this database is, who the developers are, how efficient the algorithm is. “We didn’t really use an algorithm to generate suggestions, we just pretended, the same For estimates from other people “, Specifies Science and future Eric Bogert, teacher-researcher and co-author of this research.
For the first experiment, a group of 530 people was divided into two parts. One benefited from algorithm evaluation, the other from human estimation. The result is clear: in the first half, they are 11% higher than the second to revise their estimate. And when the exercise gets complicated, with an increase in the number of people in photos, people who rely on algorithms are 14.6% more likely than those who rely on human suggestions.
The second experience is for 514 other participants. This time everyone has received two estimates, an algorithm for 5 photos and a human for 5 others. “Results reinforce those of first experience”, Write the researchers in your article. “Algorithm appreciation has a clear effect But this effect is entirely due to dependence on algorithms when it comes to difficult-to-solve problems. “
The quality of the algorithm advice put to the test
“For these two Prior experience, the suggestions given always correspond to the correct number of people in the image, calculated by the database editor, Eric Bogert This bias Allowed us not to depend on the performance of the algorithm. “
On the other hand, it is a variable of the quality of the suggestion that is tested in the third experiment. The researchers again divided the participants into two groups, which received algorithm-generated feedback and those that received one from others. Within each group, half of the suggestions were correct, with the other half having a 100% exaggerated rating (exactly twice the number of people in a picture). In other words: the error of evaluation was almost obvious.
However, once again, the belief in the algorithm is confirmed. However, with more nuances. The participants who choose to rely on the algorithm are higher when the program gives the correct answer than when it returns incorrect. More interesting: the difference between the number of people who rely on the wrong proposition and those who follow the correct answers is greater when it comes to the algorithm’s suggestion. “Topics over-penalize algorithms for bad advice”, Prepares the team.
As theoretical, this work refers to the use of algorithms in real cases. For example, if I plan to buy a house, I will compare it to my friends, who tell me about the places where I should live. [un site d’immobilier], Which uses an algorithm, will tell me that such and such a house is “for me”. Similarly, Spotify recommends new music for me to discover, and so do my friends ”, Eric Bogert Description.
This research shows an extreme confidence in these tools, a belief in their effectiveness when involved in a complex task, even though they may impact on the survival of some people. It is also not certain that an algorithm will be able to distinguish that there are 71 characters on an abundant cover. Sgt Paper’s Lonely Hearts Club Band.