Imagine the risk if a chatbot could claim to be right when it’s completely wrong. NASA start leaning back artificial intelligence Useful for life detection.
Can it be done? rover Suddenly a hallucination occurs alien?
Researchers at Michigan State University wanted to find out, so they trained an AI system using a computer program called . Avida. In test runs, the algorithm made the distinction clearly and got it right 99.97% of the time.
So the team tried to fool it. They took a chunk of non-living code and made small changes over and over again to convince the AI that it was seeing something alive. Eventually, the system became almost 100% sure that many impostors were alive, even though they never exhibited the one behavior defined as “alive” in the virtual world: duplication.
The study shows how AI can figure out patterns in training data without fundamentally understanding why something is alive, said Ankit Gupta, a doctoral student in computer science and engineering. Gupta and co-author Christophe Adami, a professor of microbiology, molecular genetics, physics, and astronomy, warned that this weakness could lead to failure. space Mission to send smart equipment Mars, ice moonor Other extraterrestrial environments To look for signs of life.
Astronomers believe that some asteroids were comets all along. Here’s why:
the team will present their discoveries At the 2026 Artificial Life Conference in Waterloo, Canada, in August.
“AI classifies ‘usual’ things very accurately,” Gupta told Mashable. “The problem arises when presented with an example that differs from the common one, and the AI will confidently misclassify it. For obvious reasons, this possibility is much higher with extraterrestrial life.”
Mars’ Perseverance rover has collected rock samples containing fossilized material that may have been created by ancient alien microbes.
Credit: NASA / JPL-Caltech / ASU / MSSS
At the same time, NASA and its partners Life detection project using AIis betting that machine learning can sift through huge and complex data sets that scientists can’t fully analyze on their own. This is setting the stage for a future where AI helps flag promising leads and humans make the final decision.
mashable light speed
In one NASA-funded effort, a large team led by Michael L. Wong of Carnegie Science and Caleb Scharf of NASA is building machine learning tools to find patterns of life in complex chemical data. their project is $5 million grantaims to train the AI on at least 1,000 samples. meteorite From rocks to fossils to living things.
Some of the data is taken from real equipment. mass spectrometer Like those already flying on NASA missions. The algorithm learns how the mix of molecules changes when life forms the material and when simple physics and chemistry play a role.
Although Wong welcomes the new research, it is unclear whether the experiment has exposed any serious vulnerabilities in the AI approach. In his view, what matters is how the research actually created that false life. The authors used a kind of selection process where they kept refining the code until it looked more authentic.
“I don’t think it’s that bad, to be honest,” he told Mashable. “To generate their life imitators, their impersonators, they actually had to carry out natural selection. It’s a living process. I don’t know of any real physical environment that would carry out that kind of selection, where it says, ‘We’re going to favor entities that just resemble life but don’t actually do life-like things.'”
But this research led Wong to think more about what kinds of examples to feed into life detection models. His group hopes to continue training AI not only on living things, but also on life’s debris and byproducts, such as fossils, slime, and waste.
After all, he said, if a model was only taught to identify living things, it wouldn’t be able to recognize dead things. If Mars or another world is the host, it could create a big blind spot chemical leftovers of a long extinct biosphere.
Some ascidians host photosynthetic algae in their outer tissues.
Credit: Hal Beral/Getty Images
“Fossils are evidence of life, and so is the chair I’m sitting in, because they don’t magically suddenly exist thanks to geology,” he said. “Neither the fossil nor my chair are living things, but I think they are part of life. They are products of life.”
Wong’s team also spends a lot of time fooling the models. In some cases, algorithms have surprised scientists by finding patterns that humans initially missed. In one test, the system kept claiming: Hoya Even though sea squirts do not make their own food from sunlight, the samples appeared to photosynthesize. At first it looked like an error, but it wasn’t. Wong said these organisms often host sunlight-harnessing algae, and AI discovered their chemical fingerprints.
Despite advances in technology, the Michigan State team is not confident that its tools are ready for use on the spacecraft. The team’s lab is also working on its own machine learning techniques to protect against this type of trickery.
“We can wait,” Mr. Gupta said. “Obviously, we’re not there yet.”
Topics
artificial intelligence
