Overview
- Researchers at Michigan State University presented the experiments on Tuesday showing a neural network trained to spot digital ‘life’ could be fooled every time in a controlled test.
- The team used Avida, a simulation that creates tens of thousands of digital organisms, to train a neural network that reached about 99.97 percent accuracy on its training set.
- When researchers made small, incremental edits to non-replicating digital code—on the order of roughly 150 changes—the model began reporting high-confidence false positives for life.
- The authors say this is a deliberate proof-of-concept that exposes a gap between training performance and real-world generalization and they urge human-in-the-loop verification for mission-critical detection tasks.
- Related work from the Southwest Research Institute found AI-generated lunar crater catalogs underperform human catalogs, which suggests this weakness in automated science tools could affect multiple fields and the MSU team plans to repeat tests with real biological data.