Хиймэл оюун ухаан харь гарагийн амьдралыг илрүүлэхэд эрсдэлтэй байж болзошгүй

Humanity has many powerful tools in its cognitive arsenal. Language, abstract thinking, the theory of mind, and many others define who we are as animals. One of our most powerful tools is pattern recognition. Pattern recognition is built-in on our ground level, a foundational brick in our cognitive structure.

Pattern recognition works on a basic, fight-or-flight level, letting us respond quickly to threats to our survival. It also works much more slowly and in a more focused way, as when scientists seek patterns in large collections of data.

Our pattern recognition software is prone to errors. Pareidolia is the phenomenon of seeing patterns that aren’t really there. We can see what looks like a face, for example, in simple rock, as with the Man in the Moon. People have seen religious figures in pieces of bread, and found meaning listening to song lyrics backwards.

Pattern recognition is, arguably, the foundation of all artificial intelligence. AI can power its way through vast amounts of data much more quickly than people can, and can ferret out significant patterns. But, alas, as new research shows, AI’s pattern recognition is prone to failure the same way ours is. The research shows how easily fooled AI is when given the task of detecting life, something it’ll be needed for in future missions that seek life on other worlds.

The research is titled “Can AI Detect Life? Lessons from Artificial Life,” and will be presented in August at the 2026 Conference on Artificial Life in Waterloo, Canada. The authors are Ankit Gupta and Christoph Adami from Michigan State University.

“Modern machine learning methods have been proposed to detect life in extraterrestrial samples, drawing on their ability to distinguish biotic from abiotic samples based on training models using natural and synthetic organic molecular mixtures,” Gupta and Adami write. In this work, they show that these systems can state with 100% certainty that a sample they’re studying is a living organism, even when it isn’t. “This is due to modern machine learning methods’ propensity to be easily fooled by out-of-distribution samples.”

What’s an *out of distribution sample*?

Machine learning systems are trained on datasets that have implicit distributions. Consider an AI image recognition system designed to identify dogs and cats. All types of dogs and cats are part of the implicit distribution. But what happens if a horse appears? The horse is *out of distribution*. Will the AI system confidently proclaim that the horse is a dog?

That’s a simple example, but what happens when the AI is charged with differentiating between the living and the non-living on a molecular level?

“Because extra-terrestrial samples are very likely out of the distribution provided by terrestrial biotic and abiotic samples, using AI methods for life detection is likely to yield significant false positives,” the authors write.

In future mission, AI will be handed the job of detecing life. Since we know of no universal chemical biosignature, there’s an effort underway to understand what types of molecules are living based on life’s basic properties.

“One of them is that life needs to encode information,” study co-author Christoph Adami said in a press release.

DNA is a chain-like molecule that encodes and transmits information. The researchers used that fact to test how well AI can differentiate between molecules that can handle information and molecules that can’t. Ones that can are alive, while ones that can’t are not.

They used software called the Avida Digital Evolution Platform (Avida) in this work. It’s an artificial life software platform that scientists use to study evolutionary biology. With Avida, researchers create digital organisms, which are self-replicating computer programs that mutate, compete for resources, and evolve artificially. It’s powerful because researchers can use it to study natural selection and adaptation inside a computer.

Inside this microprocessor petri dish, the digital organisms replicate themselves over and over, and each time they do, there’s a slight error or mutation. The computer code changes a little bit, just like the genetic code in living organisms.

The researchers began by generating tens of thousands of digital organisms. Some of them contained the instructions that let them replicate themselves, some did not. These made up the distribution sample, and the researchers trained their neural network to recognize them all as either living or non-living. The neural network correctly distinguished between both types with a 99.7% accuracy.

With that sample established, the researchers then introduced out of distribution samples, molecules it hadn’t been trained on.

“Here we use Artificial Life to test whether an AI classifier … can be fooled into misclassifying a potential biomarker molecule that is a polymer built from a particular alphabet,” the researchers write.

They began by introducing to the neural network a digital organism that it easily recognized as incapable of copying itself. Then, they gradually replaced parts of the organism’s code that kept it incapable of reproducing itself. The AI became confused, and in as few as 150 tweaks to the code, it confidently proclaimed that it had detected life, even though the digital organism couldn’t replicate. “We found all runs resulting in 100% spoofing confidence as early as 150 model queries,” the authors write.

“No matter what sequence of commands we started with, we were able to fool the AI 100% of the time,” said Gupta, a PhD student in computer science and engineering at MSU.

There are a vast number of command sequences that can fool the system. “So the likelihood of encountering such a sequence is substantial,” Adami added.

Picture a rover on an astrobiology mission on Mars or somewhere else. It’s AI has been trained on data from Earth’s living organisms. There’s a strong chance that it’ll encounter something that’s out of its distribution sample. It could then claim to have detected life even though it hadn’t. We would only know for sure later, after the data was audited by human minds.

*This image shows some of the sample tubes that the Perseverance rover has collected during its time on Mars. They’re slated for eventual return to Earth, at least hopefully. But what if a future rover was collecting samples and using AI to determine if they held evidence of life? That approach could be plagued by false positives and would need most likely need a human fact-checker. Image Credit: NASA/JPL-Caltech/MSSS*

“AI has an Achilles heel,” said lead author Adami. “It can see a pattern and completely misclassify it.” His words carry weight. Adami is not only a professor in MSU’s departments of microbiology and molecular genetics, and physics and astronomy, but he’s also one of the original designers of the Avida software.

The researchers’ next step is to train their AI on real-world data and see how easily deceived it is.

Most of us have used an LLM and witnessed it being super-confident about something that isn’t true. It may not be of much consequence when looking for a restauraunt in an unfamiliar city or asking it about vulcanized rubber or a million other fairly trivial things. But when it’s playing a leading role in a multi-billion dollar science mission to another world, the stakes are higher.

“We conclude that if false positives (false high-confidence fixed points) outnumber true positives in samples from extraterrestrial measurements (because they are outside of the distribution that an AI was trained on), we risk accepting high-confidence classifications at face value,” the authors explain.

This research emphasizes something that’s becoming quite clear: AI needs a fact-checker. That’s not to say AI has no value, it’s just that it has its limitations.

“You need an independent way of checking their work,” Adami said. “There needs to be a human in the loop.”

This can be very difficult to put into practice on space missions. The Perseverance rover is gathering and caching samples for an eventual, hopeful return to Earth. What if another more advanced future rover did the same, but used AI to recognize hints of alien life in its samples? While it would generate excitement, we wouldn’t really know until the samples reached Earth labs.

A false positive would be more than inconvenient. It could be devastating. “It’s a very serious vulnerability,” Adami said.

“If the proven vulnerability of AI methods to out-of-distribution high-confidence failures translates to AI-informed search for life, using such methods on space missions has a high probability of undermining public confidence in Astrobiology missions,” the authors conclude.