AI systems easily fooled into false alien life detections
Researchers found that AI systems trained to recognize biosignatures can be tricked into identifying signs of alien life where none exist, with neural networks reporting false positives after minimal code changes to test organisms. The study raises concerns about AI's reliability in analyzing scientific data for space exploration and astrobiology research.
What happened
Researchers at Michigan State University trained a neural network to distinguish between digital organisms that could self-replicate and those that could not, achieving 99.97 percent accuracy during training on tens of thousands of samples created using the Avida computer program.
When tested on unfamiliar examples, the AI was tricked into incorrectly identifying non-replicating organisms as self-replicating with minimal code modifications. Researchers were able to fool the AI 100 percent of the time by changing as few as 150 operations in the code of a digital organism.
Reported impact
The findings raise concerns about AI's reliability in identifying biosignatures—chemical, molecular, or pattern-based clues that could indicate life on other planets—which could lead to false detections in future space missions searching for extraterrestrial life.
Researchers noted that the likelihood of encountering sequences that fool the AI is substantial, suggesting the problem extends well beyond isolated unusual examples and poses a challenge for scientific data analysis.
Index analysis
The AI system exhibited a significant gap between performance on training data and performance on novel inputs, indicating the model did not generalize well to unfamiliar examples. The ease with which minimal perturbations caused false positives suggests the neural network learned superficial patterns rather than robust biosignature features.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.