Artificial intelligence is increasingly proposed as a tool for astrobiology. A future Mars rover, an instrument analysing material from an icy moon or a telescope examining an exoplanet atmosphere could generate so much data that AI systems may help decide which samples contain possible signs of life.
A new experiment shows why that could go spectacularly wrong.
Programs able to copy themselves reproduce, their code can mutate, and populations evolve — allowing researchers to study fundamental characteristics of life without using biological cells.
“Michigan State University researchers Ankit Gupta and Christoph Adami used an artificial-life environment called Avida, in which small computer programs behave like digital organisms.”
EuroAsia.News
The researchers generated tens of thousands of digital sequences, some capable of self-replication and others incapable of reproducing. They then trained a neural network to distinguish the two.
Its performance appeared extraordinary: 99.97% accuracy.
If that were the end of the experiment, it would look like a convincing demonstration that machine learning can identify life-like systems.
Then researchers deliberately moved outside the type of data on which the network had been trained.
Starting with a digital program the network correctly classified as non-living, they progressively changed individual operations in its computer code while keeping the resulting program incapable of self-replication. In some cases, after only about 150 attempts, the neural network changed its judgement and classified the non-replicating program as living.
The researchers could fool the model regardless of the starting sequence they tested — effectively achieving a 100% success rate at finding examples that triggered a false-positive classification.
The AI had not learned the abstract meaning of “life”. It had learned statistical patterns associated with the examples labelled as living.
This is an example of the out-of-distribution problem. Machine-learning models can perform extremely well when new inputs resemble their training data, yet behave unpredictably when confronted with inputs from a fundamentally different statistical distribution.
And extraterrestrial biology represents perhaps the ultimate out-of-distribution problem.
Every confirmed life form used to train a biological AI comes from one planet and one evolutionary history: Earth. Terrestrial organisms share DNA or RNA chemistry, related proteins, similar cellular machinery and billions of years of common ancestry.
Life somewhere else need not follow the same details. Conversely, unfamiliar non-biological chemistry on another planet could accidentally resemble statistical features that an Earth-trained model associates with life.
There is no single universally accepted biosignature. Scientists instead look for combinations — complex organic chemistry, isotopic patterns, atmospheric disequilibrium, structures, replication or information storage — while trying to eliminate geological and chemical alternatives.
The researchers therefore do not argue that AI should be removed from astrobiology. Its ability to sift enormous datasets remains valuable. Their warning is about confidence.
A model can say “99.9% probability of life” without possessing anything resembling a scientist's understanding of why the signal is biological.
The work is being presented at the 2026 Conference on Artificial Life in Waterloo, Canada, held August 17–21, following an earlier preprint of the study.
Its lesson extends far beyond Mars.
The same weakness applies when AI evaluates an unusual tumour scan, controls an autonomous vehicle or assesses a battlefield image: the most dangerous input may be precisely the one unlike anything the machine has seen before.
In the search for extraterrestrial life, the first rule may therefore be surprisingly terrestrial: never allow an AI's confidence score to become the discovery itself.
