Researchers from Michigan State University have discovered that modern artificial intelligence models can be highly unreliable when it comes to detecting signs of extraterrestrial life.
In their experiment, the scientists managed to deceive a specially trained neural network in every instance by making such minor alterations to the data that the algorithm began mistaking them for evidence of life.
The authors of the study tested how well AI could recognize biomarkers—indicators that might suggest the presence of living organisms. For their experiments, they used the university-developed Avida system, which simulates digital evolution. In this system, virtual organisms are represented by code that copies itself, gradually accumulating random changes—an analog of biological evolution.
The researchers trained the neural network on tens of thousands of such digital organisms, some of which contained a self-replication command, while others did not. On familiar data, the algorithm demonstrated nearly perfect classification accuracy.
However, after this, the scientists began to gradually alter the organisms’ code, presenting the neural network with variants it had never encountered before. It turned out that sometimes as few as 150 minor changes in the program code were enough for the algorithm to confidently mistake false signals for signs of life. According to one of the study’s authors, Ankit Gupta, regardless of the chosen sequence of changes, the researchers managed to deceive the model in 100% of cases.
“Artificial intelligence has an Achilles’ heel. It can identify a pattern and completely misclassify it,” noted co-author Christoph Adami. He emphasized that the issue lies not in the technology itself, but in its vulnerability to data that differs from what it was trained on.
The authors point out that these findings are significant not only for future space missions, where AI might analyze data from rovers or interplanetary probes without constant human oversight. Similar errors could also occur in facial recognition systems, autonomous vehicles, medical diagnostics, and other fields where decisions are made by a model.
Researchers believe that artificial intelligence should remain a supporting tool, and its conclusions must be independently verified. According to Christoph Adami, in scientific research tasks, especially those involving the potential discovery of extraterrestrial life, humans must remain an essential part of the decision-making process.
About the Author
