A study by Christoph Adami shows machine learning models proposed for extraterrestrial life detection classify artificial life outputs as biotic with near 100% confidence despite lacking capacity for life (arXiv:2604.11915).

The paper demonstrates that ML methods trained on terrestrial biotic and abiotic molecular mixtures are vulnerable to out-of-distribution samples generated via artificial life platforms, a limitation prior coverage of AI-for-astrobiology proposals such as those using mass spectrometry classification (Marshall et al., PNAS 2021, https://www.pnas.org/doi/10.1073/pnas.2103395118) did not test. Related astrobiology missions including Viking labeled-release experiments and ALH84001 meteorite analysis exhibited similar ambiguity when encountering non-terrestrial chemical patterns (NASA Astrobiology Strategy 2015).

Synthesis with assembly theory (Walker et al., Nature 2023, https://www.nature.com/articles/s41586-023-06600-9) indicates that complexity measures grounded in selection and evolution provide a more agnostic alternative to pure ML pattern matching; the Adami work reveals mainstream AI coverage missed how artificial life directly informs the false-positive risks for Mars Sample Return and Europa missions by exposing distribution shift failures inherent to supervised models.