A recent breakthrough in artificial intelligence (AI) research has enabled scientists to uncover the oldest molecular evidence of photosynthesis on Earth, pushing back the timeline of life’s early biochemical activity by hundreds of millions of years. This advance not only deepens our understanding of Earth’s primordial biosphere but also promises to revolutionize the search for extraterrestrial life.
The research, published on November 17, 2025, in the Proceedings of the National Academy of Sciences, details how a team of scientists employed machine learning techniques to analyze molecular signatures preserved in ancient rocks, some as old as 3.3 billion years. This is a significant leap forward, more than doubling the previous record for identifying molecular signs of life, which stood at about 1.6 billion years.
One of the most remarkable findings of the study is the detection of biomolecular evidence indicative of photosynthesis in rocks dating back 2.5 billion years—approximately 800 million years earlier than previously confirmed through molecular data. This discovery aligns with other geochemical evidence but marks the first time such biological activity has been identified through molecular signatures alone. The implications extend beyond Earth’s history, as the methodology could be adapted to search for signs of life on other planetary bodies, such as Mars or the icy moons of the outer solar system.
A key advantage of this approach is its potential for real-time analysis aboard space missions. According to Michael Wong, the study’s lead author and an astrobiologist at the Carnegie Institution for Science, their machine-learning algorithm could operate directly on rover missions, eliminating the need to transport samples back to Earth for detailed laboratory examination. This capability could greatly accelerate the pace of astrobiological discovery by enabling in situ detection of biosignatures.
The technique is especially promising because it approaches the search for life in an “agnostic” manner, independent of Earth-centric assumptions about biology. Karen Lloyd, a biogeochemist at the University of Southern California not involved with the study, praises the method for its ability to analyze a wide variety of organic molecules, whether derived from living organisms or not. This flexibility is invaluable when examining ancient Earth rocks, where traditional fossil evidence is scarce or ambiguous, and even more so when analyzing extraterrestrial materials whose biological origins are uncertain.
Understanding Earth’s early life has long been a challenge because the planet’s oldest rocks have been heavily altered by geological processes over billions of years. Beyond roughly two billion years ago, no pristine, unaltered Earth rocks exist. Such extensive alteration often obliterates or obscures any direct signs of ancient life, making it extremely difficult to identify definitive biosignatures. Traditional fossil evidence—like dinosaur bones, plant remains, or recognizable microfossils—represents only the most recent 10 percent of Earth’s 4.5-billion-year history. The preceding billions of years, often called the Precambrian, are dominated by microscopic life whose presence is mostly inferred through chemical traces such as lipids and amino acids.
According to Robert Hazen, a geologist at Carnegie and the study’s lead author, the challenge has been that biomolecules degrade and vanish over time. Their new method overcomes this limitation by focusing not on individual molecules but on complex patterns within molecular fragments, much like how AI facial recognition systems reconstruct identities from partial data. Hazen likens this to how AI was used to decipher the burnt scrolls of Herculaneum, where human eyes see only smudges but AI can reconstruct coherent text.
To train their AI model, the researchers collected over 400 samples spanning a diverse range of origins: modern biological material, ancient rocks both with and without known fossils, abiotic organic compounds from meteorites, and samples containing organic molecules yet lacking obvious biological indicators. These samples were analyzed using pyrolysis gas chromatograph mass spectrometry (Py-GC-MS), a technique that vaporizes the material and sorts molecular fragments by mass and chemical properties. The resulting data sets, rich with tens of thousands to hundreds of thousands of distinct chemical peaks, provided a detailed “chemical landscape” for each sample.
The AI was trained on approximately 75 percent of these samples and then tested on the remaining 25 percent. Impressively, it correctly distinguished biotic from abiotic samples with over 90 percent accuracy in the test set. However, accuracy decreased for older and more degraded samples; for rocks older than 2.5 billion years, the AI assigned biotic origins to less than half of the samples and with reduced confidence.