AI Could Mistake Dead Rocks for Alien Life, New Research Warns

Key points

• Michigan State University researchers fooled an AI into confidently misidentifying non-life as life within roughly 15 small changes to a test sequence, every single time.
• The team ran their experiment across 1,000 parallel computers for three months, using a program called Avida that simulates digital organisms competing and evolving.
• The core problem is called "out-of-distribution" data: an AI trained on Earth life has no reliable frame of reference for alien biology, which could look completely different.
• NASA plans to launch the Habitable Worlds Observatory in the 2040s to search for signs of life on planets outside our solar system, a mission that would likely lean heavily on AI analysis.
• Adami and Gupta will present their findings at the 2026 Conference on Artificial Life in Waterloo, Canada, in August.

An AI that confidently identifies life where there is none sounds like a minor glitch. On a future Mars mission, it could be the difference between a scientific breakthrough and an embarrassing false alarm heard around the world.

Christoph Adami, a computational biologist at Michigan State University, and his student Ankit Gupta spent three months running a test that should worry anyone excited about AI-powered space exploration. Their tool was Avida, a computer program Adami built in 1993 that runs tiny digital organisms: snippets of code that copy themselves, compete for processing time, and even mutate, much like bacteria in a petri dish.

The organisms in Avida gave the team a large, clean dataset of things that are alive (by the program's rules) and things that are not. They asked an AI to sort them.

The AI did reasonably well at first. Then the researchers started nudging the non-life samples, tweaking their molecular sequences one small step at a time, each time checking whether the AI's confidence shifted. The results, as reported by Space.com, were stark.

"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," Adami said.

No matter which non-life sequence they started with, the AI was fooled. Every time.

The underlying reason is something researchers call an "out-of-distribution" problem. Think of it this way: if you train an AI to recognise apples and then show it a banana, it struggles. The banana is an unfamiliar shape, and the AI has no good mental model for it. Alien microbes, if they exist, could be just as foreign to an AI trained entirely on Earth biology.

This matters enormously for real missions. A Mars rover finding something that looks unmistakably like a microbe under a microscope is one thing. But most life-detection attempts will be far more indirect, using mass spectrometry, a technique that identifies molecules by measuring their weight, to sift through atmospheric data from Venus, ocean samples from Jupiter's moon Europa, or light signatures from planets dozens of light-years away.

Feed that ambiguous data to an AI trained only on Earth life, and you may get a confident answer that is simply wrong.

"You can't guarantee that with extraterrestrial life," Adami said plainly.

What does this mean for future space missions?

It means AI should be a first-pass tool, not the final word. Adami is not calling for AI to be dropped from astrobiology entirely. He believes it has genuine value in sorting through vast amounts of sensor data that no human team could read fast enough. The warning is more specific: know what your AI was trained on, and stay honest about what it cannot know.

Next, the team plans to move beyond digital organisms and run the same tests on real-world chemical data, a harder and more revealing challenge.