Michigan State Researchers Fool AI Designed to Detect Signs of Life

A team at Michigan State University has shown that artificial intelligence trained to distinguish living from dead forms can be fooled every single time samples stray from its training data. The finding is directly relevant to NASA, which wants to use similar systems to analyze samples collected by the Perseverance rover on Mars.

Digital organisms instead of DNA

Instead of real cells or soil samples, the researchers used Avida, a program that has simulated evolution for decades. In its virtual environment, organisms written as strings of computer instructions copy themselves in a digital petri dish, with each copy introducing small, random errors, much like mutations in real genetic code.

From tens of thousands of these digital organisms, the researchers trained a neural network to classify them as alive, meaning capable of self-replication, or dead. On data drawn from the training distribution, the model performed nearly flawlessly, reaching 99.97 percent accuracy.

How the algorithm was fooled

The trouble began when researchers started feeding the AI samples outside its training distribution. They gradually swapped individual instructions in a dead organism's code for others, nudging it visually closer to the pattern of a living one, even though the organism still could not replicate.

An average of just 150 such modifications was enough to make the model classify the dead code as alive with full confidence. Crucially, the method worked every time, regardless of which command sequence the researchers started from.

No matter what sequence of commands we started with, we were able to fool the AI 100% of the time - Ankit Gupta, PhD student in computer science and engineering, Michigan State University

AI has an Achilles heel. It can see a pattern and completely misclassify it - Christoph Adami, professor of microbiology, molecular genetics, physics and astronomy, Michigan State University

Stakes for NASA's missions

The results are not just an academic curiosity. NASA is developing AI systems meant to help sift through the enormous volumes of data gathered by missions such as the Perseverance rover, which has been exploring Jezero Crater since 2021 in search of traces of ancient microbial life. Automation makes sense because no human team could physically review every sample and measurement by hand.

Adami stresses that this kind of vulnerability to misclassification is a serious gap, especially for sensors deployed on planetary hardware, where there is no way for a human to verify results on the spot in real time. A false alarm about the discovery of extraterrestrial life could spread through the media before scientists had a chance to check the result carefully.

The researchers argue the answer is not to abandon AI but to keep a human in the decision loop as an independent check that verifies the algorithm's conclusions before they are made public.

you need an independent way of checking their work. There needs to be a human in the loop - Christoph Adami, professor, Michigan State University

Limits of pattern recognition

The key takeaway from the study concerns what the neural network is actually learning. The model does not understand the mechanism of life in a biological sense, it recognizes statistical patterns in the data it was trained on. When a sample departs from that distribution, the way real extraterrestrial samples might differ from their Earth counterparts, the risk of misclassification rises sharply.

Gupta said the team's next step will be to retrain the model on real biological data to see whether the same weaknesses show up outside Avida's simulated environment. The results are meant to help design more robust systems before they are deployed on actual missions searching for life in space.