Scientists talked about the vulnerability of AI in the search for extraterrestrial life

Artificial intelligence (AI) is prone to serious errors and false conclusions when identifying signs of extraterrestrial life. Researchers from the University of Michigan have experimentally proved that modern machine learning algorithms can be deceived in 100% of cases. This was reported by Popular Science magazine on July 7.

Computer engineer Ankit Gupta and his colleague Christophe Adami conducted an experiment using the Avida program, which simulates evolutionary processes. Scientists have trained a neural network on tens of thousands of such digital organisms so that it can distinguish self-replicating chains from the rest.

Despite the high accuracy at the training stage, the algorithm stopped coping when faced with new examples. A total of 150 minor changes in the program code of organisms led to the fact that AI began to massively mistake false signals for signs of life.

According to the authors of the study, the problem goes beyond space missions. False alarms can affect facial recognition systems, self-driving cars, and medical scanners. Scientists emphasize that the results of the work are not a verdict for technology, but they remind us of the need for human control.

In connection with the experimental results, Adami pointed out the need for an independent way to verify the operation of the algorithm. At the same time, he stressed that a person must participate in this cycle.

The Code of the Universe: How AI searches for life outside the Solar System

Neural networks help astronomers analyze huge amounts of data in search of objects similar to Earth.

On February 1, Bloomberg reported on SpaceX's plans to launch 1 million satellites for AI training. According to the publication, the request outlines the issue of deploying data centers in space to perform complex computing tasks in the field of AI.