Exoplanet K2-18b, located approximately 120 light-years from Earth, is a constant object of interest for astrobiology. A recent study utilized two of the world's most powerful radio telescopes in an attempt to 'listen' to this planet for artificial signals.
The analysis conducted did not detect any 'technosignatures' or signs of alien technology.
A study conducted by Michigan State University demonstrated that artificial intelligence systems are capable of classifying objects that do not correspond to real organisms as 'life'. This conclusion was reached after testing with simulated life forms in a computer environment.
The results of this study raise a serious warning for future space missions that plan to use AI in the search for signs of extraterrestrial life. Artificial intelligence has become widespread in science due to its ability to quickly analyze vast amounts of information and find patterns.
However, the new research indicates that these systems can fail when encountering situations significantly different from those they were trained on. Computational biologist Christoph Adami and researcher Ankit Gupta conducted experiments using the Avida program, developed in 1993. This platform allows for the modeling of code-written digital organisms that can self-reproduce, evolve, and compete for computational resources.
Over three months, the scientists simultaneously analyzed thousands of machines to test whether the technology could distinguish between digital programs exhibiting signs of life and those that do not. The results attracted attention: Adami explained that the system could be convinced that something was alive even if it was not.
He noted that after approximately fifteen changes, he managed to make the AI fully believe in the classification of life, although not once when it was 100% certain was it actual life.
The main difficulty lies in what is known as 'out-of-distribution data'. Simply put, AI models work most effectively when analyzing information similar to that used during training. Adami draws an analogy with training a system to recognize apples: when given images of fruit, the result is usually correct. But when bananas appear, the model may get confused because it never studied that pattern.
A similar problem arises in the search for extraterrestrial life. Organisms from other planets may possess characteristics entirely different from known terrestrial forms. Adami emphasized: 'You must know your training data, and if you know that the test data is part of the same distribution, everything will be fine. But this cannot be guaranteed in the case of extraterrestrial life.'
This problem potentially affects research dependent on the analysis of chemical signals, including the study of Venus, Europa, and exoplanets. Caleb Sharp from NASA Ames believes that the solution may lie in creating systems capable of recognizing unforeseen situations. He stated: 'Gupta and Adami are pointing out a critical problem in life detection that could undermine expected successes of AI and machine learning: what do we do with unknown unknowns?'
Currently, Adami and Gupta plan to transfer the testing to real-world data. The results will be presented in August at the Artificial Life Conference in 2026 in Waterloo, Canada.
