When AI hallucinates aliens: Some think NASA should be nervous

Michigan State University researchers found that an AI system trained to distinguish between digital organisms that can replicate and those that cannot became almost 100 percent certain that nonliving code was alive after small changes, raising concerns that NASA's use of AI for life detection on Mars or icy moons could lead to false positives. The team, led by doctoral student Ankit Gupta and professor Christoph Adami, will present their findings in August at the 2026 Conference on Artificial Life in Waterloo, Canada.

When AI hallucinates aliens: Some think NASA should be nervous Elisha Sauers /author/elisha-sauers NASA's moon and Mars missions https://mashable.com/article/why-nasa-going-to-moon-artemis , chatting up astronauts https://mashable.com/article/nasa-artemis-2-moon-commander-reid-wiseman and history-making discoverers https://mashable.com/article/black-hole-milky-way-discovery-sagittarius-a , and jetting above the clouds https://mashable.com/article/solar-eclipse-flight-delta-2024 . Through 17 years of reporting, she's covered a variety of topics, including health, business, and government, with a penchant for public records requests. She previously worked for The Virginian-Pilot in Norfolk, Virginia, and The Capital in Annapolis, Maryland. Her work has earned numerous state awards, including the Virginia Press Association's top honor, , and https://www.youtube.com/watch?v=ruwl6a8d2Z4 Best in Show for narrative storytelling. For each year she has covered space, Sauers has won https://featuresjournalism.org/contests-2/sfj-27th-annual-winners-by-category/ national recognition , including first place for her https://www.headlinerawards.org/2024-online/ National Headliner Awards Sex in Space https://mashable.com/series/sex-in-space series. Send space tips and story ideas to or text 443-684-2489. Follow her on X at /cdn-cgi/l/email-protection 66030a0f150e0748150713031415261c0f00000b03020f074805090b email protected . https://twitter.com/elishasauers @elishasauers Read Full Bio /author/elisha-sauers If a chatbot can insist it's right while being totally wrong, imagine the stakes when NASA https://mashable.com/category/nasa starts leaning on to help with life detection. https://mashable.com/category/artificial-intelligence artificial intelligence Could the rovers https://mashable.com/science/nasa-future-moon-mars-rover-prototype-artificial-intelligence suddenly hallucinate ? https://mashable.com/article/nasa-ufo-uap-study aliens Michigan State University researchers wanted to know, so they trained an AI system to spot the difference between digital "organisms" that can copy themselves and others that can't, using a computer program called Avida https://avida.devosoft.org/ . In test runs, the algorithm nailed the distinction, getting it right 99.97 percent of the time. You May Also Like Then, the team tried to fool it. They took nonliving chunks of code and made small changes over and over, nudging the AI to feel more confident that it was looking at something alive. Eventually, the system became almost 100 percent certain that many of the fakers were alive, even though they never exhibited the one behavior defined as "alive" in that virtual world: replication. The work shows how AI can latch onto patterns in its training data without fundamentally understanding what makes something alive, said Ankit Gupta, a doctoral student in computer science and engineering. Gupta and his coauthor, Christoph Adami, a professor of microbiology and molecular genetics, physics, and astronomy, warn that this weakness could foil space https://mashable.com/category/space missions that send smart instruments to , https://mashable.com/science/nasa-mars-curiosity-rover-pure-sulfur-new-theory Mars , or https://mashable.com/article/nasa-dragonfly-why-flight-beats-rover icy moons to hunt for signs of life. https://mashable.com/article/venus-volcano-volcanic-activity-discovery other extraterrestrial environments The team will present their findings https://arxiv.org/abs/2604.11915v2 in August at the 2026 Conference on Artificial Life in Waterloo, Canada. "AI is highly accurate at classifying things that are 'usual,'" Gupta told Mashable. "The problem occurs when it is presented with examples that are unlike the common things, which AI will misclassify with confidence. This is much more likely with extraterrestrial life, for obvious reasons." At the same time, NASA and its partners are pouring money into AI‑driven life‑detection projects https://carnegiescience.edu/news/teaching-ai-detect-life-carnegie-scientist-co-leads-nasa-funded-effort , betting that machine learning can sift through huge, complex data sets that scientists could never fully analyze on their own. It's setting the stage for a future in which AI helps flag promising clues, with humans making the final call. Terms of Use https://www.ziffdavis.com/terms-of-use and Privacy Policy https://www.ziffdavis.com/ztg-privacy-policy . In one NASA‑funded effort, a large team led by Carnegie Science's Michael L. Wong and NASA's Caleb Scharf is building machine‑learning tools to pick out patterns of life in complex chemical data. Their project, supported by a $5 million grant https://nspires.nasaprs.com/external/solicitations/summary.do?solId=%7B3D196E9B-0D83-A882-C488-CA112557D945%7D&path=&method=init , aims to train AI on at least 1,000 samples — from and rocks to fossils and living organisms. https://mashable.com/article/first-witnessed-meteorite-fall-ensisheim meteorites Some of the data comes from real instruments, including mass spectrometers https://youtu.be/ L4U6ImYSj0 like those already flying on NASA missions. The algorithm learns how a mix of molecules shifts when life has shaped that material versus when simple physics and chemistry have done the job. While Wong welcomes the fresh scrutiny, he isn't sure the experiments exposed a serious vulnerability in AI approaches. In his view, the way the study actually created its fake life matters. The authors used a kind of selection process that kept refining the code until it looked increasingly lifelike. "I don't think it's that bad, to be honest," he told Mashable. "To generate their life mimics — their spoofs — they actually had to perform natural selection, which is a living process. I don't know of any actual physical environment that performs that kind of selection, where it says, 'I will prefer entities that just look like life, but don't actually do lifelike things.'" But the research did push Wong to think more about what kinds of examples he feeds into life‑detection models. His group wants to continue training AI not just on things that are alive, but also on the remains and byproducts of life — the fossils, goo, and waste. After all, if the models are only taught to identify living things, they won't be able to recognize the dead ones, he said. That could create a massive blind spot if Mars or another world hosts only the chemical leftovers https://mashable.com/article/mars-rock-microbial-life-biosignature of a long‑extinct biosphere. "A fossil is evidence of life. So is the chair that I'm sitting on because those don't just magically pop into existence thanks to geology," he said. "A fossil and my chair are not alive, but I think they're part of life. They are a product of life." Wong's team also spends a lot of time trying to deceive its models. In some cases, the algorithm has surprised the scientists by picking up on patterns the humans initially missed. In one test, the system kept insisting that a sea squirt https://a-z-animals.com/animals/sea-squirt/ sample looked photosynthetic, even though sea squirts don't make their own food from sunlight. What initially looked like an error was anything but, Wong said: Those creatures often host algae that do use sunlight, and the AI had spotted the chemical fingerprints. Despite the advances in technology, the Michigan State team is not convinced the tools are ready for spacecraft. The team's lab https://msutoday.msu.edu/news/2026/07/ai-alien-hoaxes is also working on its own machine-learning methods to defend against this kind of trickery. "We can wait," Gupta said. "Clearly, we are not there yet." Topics Artificial Intelligence https://mashable.com/category/artificial-intelligence Elisha Sauers writes about space for Mashable, taking deep dives into NASA's moon and Mars missions https://mashable.com/article/why-nasa-going-to-moon-artemis , chatting up astronauts https://mashable.com/article/nasa-artemis-2-moon-commander-reid-wiseman and history-making discoverers https://mashable.com/article/black-hole-milky-way-discovery-sagittarius-a , and jetting above the clouds https://mashable.com/article/solar-eclipse-flight-delta-2024 . Through 17 years of reporting, she's covered a variety of topics, including health, business, and government, with a penchant for public records requests. She previously worked for The Virginian-Pilot in Norfolk, Virginia, and The Capital in Annapolis, Maryland. Her work has earned numerous state awards, including the Virginia Press Association's top honor, Best in Show https://www.youtube.com/watch?v=ruwl6a8d2Z4 , and for narrative storytelling. For each year she has covered space, Sauers has won https://featuresjournalism.org/contests-2/sfj-27th-annual-winners-by-category/ national recognition , including first place for her https://www.headlinerawards.org/2024-online/ National Headliner Awards Sex in Space https://mashable.com/series/sex-in-space series. Send space tips and story ideas to or text 443-684-2489. Follow her on X at /cdn-cgi/l/email-protection 9cf9f0f5eff4fdb2effde9f9eeefdce6f5fafaf1f9f8f5fdb2fff3f1 email protected . https://twitter.com/elishasauers @elishasauers