The discovery of extraterrestrial life is one of the most ambitious goals of modern science. However, as researchers from Michigan State University have found, the main tool for future searches may prove unreliable. Scientists have discovered a critical vulnerability in artificial intelligence algorithms: neural networks tend to incorrectly classify inanimate objects as living organisms.
This problem jeopardizes the reliability of future space missions. If robotic rovers or orbital probes rely on current versions of AI, humanity risks spending years and billions of dollars studying "dead" rocks, mistaking them for signs of biology.
The Problem of "Out-of-Distribution" Data
The main difficulty lies in how modern models are trained. Neural networks work effectively only with data similar to that used during their training. However, space is full of the unknown. Scientists call such objects "out-of-distribution" samples — they differ radically from the terrestrial biology on which the algorithms were trained.
Christoph Adami, a computational biologist and one of the authors of the study, emphasizes that AI is very easy to mislead. In the context of the search for life on other planets, where biology may be structured completely differently than on Earth, spacecraft are forced to rely on universal markers. This is where the risk of a fatal error lies.
Experiment with Digital Organisms
To test the reliability of the algorithms, Christoph Adami and his colleague Ankur Gupta conducted extensive testing. As a testing ground, they used the Avida platform — a computer program simulating the evolution of digital organisms. In this environment, code can mutate, self-replicate, and compete for processor time, mimicking the fundamental properties of life.
The experiment lasted three months and required colossal computing power: 1,000 parallel computers were used to conduct it. The goal was to check whether AI could distinguish "living" code from "dead" code with minimal changes.
Successful Deception of the Algorithm
The results of the experiment were alarming. Through targeted code adjustments, the scientists managed to completely fool the artificial intelligence. The algorithm, designed to search for signs of life, recognized the modified inanimate program as a fully-fledged living organism.
This experiment clearly demonstrates that current models do not possess sufficient reliability for autonomous operation in the extreme conditions of space. Incorrect identification of biomarkers can lead to false-positive results.
The consequences of such errors go far beyond scientific interest. A false alarm entails the wasteful expenditure of resources and time on the detailed analysis of objects that actually have nothing to do with extraterrestrial life. Before launching new missions, engineers must find a way to make AI more resistant to such manipulations.
