AI's major flaw found that could derail search for alien life

AI's heel of Achilles

Researchers believe that the main weakness of modern neural network models lies in handling so-called "out-of-distribution" samples—objects that differ from the data used to train the algorithm.

Computational biologist Christoph Adami says AI is remarkably easy to fool. He defines life by its ability to encode and reproduce information.

The researcher explained that because biology on other planets could be radically different from life on Earth, spacecraft will have to rely on these universal markers.

Digital life experiment

To test AI's reliability, researchers Christoph Adami and Ankit Gupta conducted experiments using the Avida computer platform.

The platform simulates digital organisms in the form of computer code capable of mutating, self-replicating, and competing for processor time.

The experiment lasted three months and used 1,000 computers running in parallel:

• AI was tasked with distinguishing digital organisms that exhibited characteristics of life from non-living code
• Initially, the algorithm was presented with sequences designed to mimic random collections of molecules
• Researchers then gradually altered the structure of the code while monitoring the AI's confidence in its classifications.

After making targeted adjustments, the scientists were able to completely deceive the AI, causing it to identify a non-living program as a fully living organism.

A threat to future space missions?

The newly discovered vulnerability suggests that AI-powered robots and Mars rovers could produce false-positive results when searching for signs of life.

Incorrectly identifying biomarkers on other planets could waste valuable time and resources investigating objects that have nothing to do with extraterrestrial life.