Michigan State scientists warn AI could misidentify non-life as alien life

Full text

Artificial intelligence could easily fall into the trap of identifying non-life as life on other worlds, claim two researchers from Michigan State University who have tested AI on simulated life in a computer program.

"We had previously seen that AI has a big Achilles heel when it is trying to classify things that are unlike the things in its training examples," Michigan's Christoph Adami told Space.com. "We call these 'out-of-distribution' samples and it is just incredibly easy to get AI to misclassify."

Adami is a computational biologist who uses computers to apply information theory to the study of evolution and biology. One of his leading tenets is that life can be defined by its ability to encode information and replicate it. To this end, he devised the Avida computer program in 1993. It runs digital organisms written as code that can replicate by copying themselves and competing for resources — in this case, CPU time — just like real life organisms. Although the use of digital life in evolutionary studies remains controversial, what it does provide Adami with is a huge dataset of simulated lifeforms that AI can be tested on.

Spending three months of computer analysis on a thousand parallel machines, Adami and his student Ankit Gupta asked AI to determine which programs in Avida had the properties of life, and which didn't. The life and non-life programs have very similar coding, so the difference is not obvious — and this might very well be the case on another planet where life could have differences in biology to Earth life.

Adami and Gupta started out with programs representing a random sequence of molecules and asked the AI to classify them. They then set about tweaking that sequence of molecules to try and fool the AI into thinking it was seeing life, one change at a time, each time checking whether there had been a change in how confident the AI was that it was life or non-life.

"Within about 15 changes or so we can get AI to be perfectly confident of a life classification when in fact not a single time when it was being 100% confident was it actually life," said Adami.

Furthermore, no matter the sequence they started with, the AI was constantly fooled.

Many in the scientific community stand by AI as an invaluable tool because it can process huge amounts of data and search for patterns in that data. Adami himself believes that AI has an important role to play, but as we see in everyday life, AI is prone to making things up and identifying patterns that don't exist.

Digital organisms replicating and mutating over time in the Avida program. Each colored dot represents an organism of a particular genotype, and each time they mutate into a different type, they change color. (Image credit: Adami et al/Michigan State University) This really comes down to what the AI has been trained on, said Adami. If you ask an AI about data it has been trained on, it usually provides a correct answer. For example, if you train AI to identify pictures of apples and then ask it to pick out the fruit from a dataset that also includes pictures of non-food items, it will answer correctly virtually all the time. However, if you replace the apple pictures with photos of bananas and ask it to identify the fruit, it will struggle and start to misidentify things.

That's because the bananas represent "out of distribution" data. The AI wasn't trained on bananas, which are a very different shape to apples, and therefore the AI doesn't know what to make of them.

Similarly, we don't know what alien microbes will look like, and they could be quite different to the terrestrial microbes that the AI has been trained on. In other words, the alien life would be out of distribution, and the AI would have no context for saying whether any particular collection of molecules is life or not.

"You need to know your training data, and if you know that your testing data is part of the same distribution as the training data, then you'll be fine," said Adami. "But you can't guarantee that with extraterrestrial life."

This could pose a problem for missions designed to look for life.

If a Mars rover goes to the red planet and cuts open a rock and directly sees something that looks like microbial life as we know it, then the answer will be clear-cut and can be tested using traditional methods. However, many attempts to detect life will probably not be so hands-on, relying on mass spectrometry data to identify molecules and processes related to life. This could range from attempts to detect life in the atmosphere of Venus , or in the ocean of Europa , or on exoplanets via the Habitable Worlds Observatory , which NASA aims to launch in the 2040s to directly image exoplanets in the habitable zone of stars .

"This means that if there were an AI on a mission looking at particular mass spectrometry samples, then there's a very great chance that while it has been trained on the ground on a number of biotic and abiotic samples, it could still return a positive verdict when it has absolutely nothing to do with life," says Adami.

The next step, said Adami, is to move out of the digital world and run the same test with real-world data.

Adami and Gupta will be presenting their findings in August at the 2026 Conference on Artificial Life, which is being held in Waterloo, Canada.