AI Can Be Easily Fooled When Searching for Alien Life

The neural network identified signs of life with 99.97 percent accuracy, until researchers set out to deliberately fool it. It took only about 150 small changes to a digital organism's code to make the algorithm classify something as alive with complete certainty, even though it wasn't. Researchers at Michigan State University warn that the same mechanism could fail in space missions searching for life beyond Earth.

The team used Avida software, which biologists have used for years to simulate the evolution of digital organisms, snippets of code capable of self-replication, mutation and selection, the basic mechanisms of life. The researchers generated tens of thousands of such organisms, some able to copy themselves and some lacking that ability, and used them to train a neural network to tell the two apart.

How the Experiment Worked

On data the network had seen during training, the model performed almost flawlessly, with 99.97 percent accuracy. The real test began when researchers started feeding the algorithm examples outside the training set, gradually swapping single instructions in the code of an organism that had no capacity for self-replication. After about 150 such swaps, the network began to claim with total certainty that it was looking at a living organism, even though none of the changes had actually given it a real ability to copy itself.

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

AI has its Achilles' heel. It can spot a pattern and completely misinterpret it. That's a very serious vulnerability - Christoph Adami, professor of microbiology, molecular genetics and astronomy, Michigan State University

Why It Matters for Space Missions

The problem isn't academic. Many current and planned NASA missions, from rovers drilling into Martian soil to probes analyzing the atmospheres of distant exoplanets, include searching for biosignatures, traces indicating the existence of life, on their agenda. Increasingly, AI is meant to perform the initial analysis of the vast amounts of data these instruments collect before the results reach human scientists. The Michigan State study shows that samples from beyond Earth will, by definition, fall outside the distribution of such systems' training data, making them fundamentally more prone to misclassification.

Adami notes that life is defined in part by the ability to encode and pass on information, but a machine learning model recognizing that trait is not the same as understanding why a given pattern actually indicates life. A neural network learns statistical relationships in data, not biological mechanisms, which is why a small, deliberate change in input data can completely mislead its reasoning, despite seemingly perfect performance on known examples.

What's Next

The authors' conclusions are cautious but concrete: current AI systems used to analyze potential biosignatures need an independent method of verifying results, rather than making final decisions on their own. Adami stresses that a human must remain in charge of overseeing the process in this kind of application, at least until methods are developed to check whether an AI's classification rests on real biological mechanisms rather than chance correlations in training data. The work will be presented to a wider audience of artificial life researchers in August at the conference in Waterloo, which could speed up discussion of AI validation standards in astrobiology ahead of the next major missions searching for life beyond Earth.

Sources: It's disturbingly easy to trick AI into seeing aliens (eeb.msu.edu), AI may misidentify life beyond Earth with high confidence (devdiscourse.com), Sztuczna inteligencja i poszukiwanie obcych. To nie takie proste (tech.wp.pl)