An AI trained to detect life forms with 99.97% accuracy was tricked 100% by researchers at Michigan State University. In just 150 changes to an inert piece of computer code, the system began to “see” life where there was none. A demonstration with direct implications for space missions seeking extraterrestrial life.
What you will learn
- How researchers manipulated an inert digital organism to systematically deceive a life-detection AI
- Why this result directly threatens the reliability of space rovers equipped with AI systems
- What a parallel study on AI-generated lunar crater catalogs reveals
An Almost-Perfect AI… until it is truly tested
Ankit Gupta and Christoph Adami, researchers at Michigan State University, trained a neural network to distinguish living digital organisms from inert ones. These digital organisms, generated by the software Avida, simulate evolution: they replicate, produce imperfect offspring, and behave in many respects like real life forms.
The training result was impressive: 99.97% accuracy. The AI seemed perfectly capable of recognizing life, even in configurations it had never encountered.
The researchers then decided to test it in a different way.
150 Modifications for an Illusion of Life
The team started with an inert digital organism, correctly identified as such by the AI. They then progressively modified its computer code, step by step, never granting it the ability to replicate.
In only 150 modifications, the AI flipped: it began to classify this inanimate object as living. And the result proved universal — regardless of the initial starting sequence used, the system was fooled in 100% of attempts.
“No matter what the initial command sequence is, we managed to fool the AI 100% of the time,” summarizes Gupta.
An Achilles’ Heel with Space Exploration Implications
The reach of this demonstration goes far beyond the lab. Several space agencies and private companies are considering equipping rovers with AI systems to detect traces of life on other planets.
If a perfectly trained algorithm can be misled by simple code variations, false positives in the search for extraterrestrial life become a concrete operational risk — and potentially very costly.
“AI has an Achilles’ heel,” says Adami. “It can pick up a trend and classify it completely incorrectly.”
The Moon Craters Confirm the Issue
This bias is not isolated. A study published in The Planetary Science Journal sheds light in a convergent manner: eight catalogs of lunar craters generated automatically by AI were evaluated using the same criteria applied to human work. Their quality proved highly insufficient.
AI excels at processing large data volumes and identifying recurring patterns. But it also generates many false positives, and its results can deteriorate markedly outside training conditions.
The researchers’ conclusion is clear: human supervision remains essential. “A human needs to intervene,” concludes Adami. The team now plans to reproduce the experiment on real biological data, and not merely on digital ones.
