Scientifically Speaking: How AI can hunt for ghosts in Earth’s oldest rocks, space

The problem is that chemistry fades. Over billions of years, heat and pressure scramble these molecules until they look indistinguishable from abiotic carbon, the dead and non-living matter found in meteorites. For decades, this has been the fog of war for biologists. We know life existed three billion years ago, but the molecular proof is often debated or dismissed as contamination.
A new study recently published in the Proceedings of the National Academy of Sciences has pierced that fog. By combining mass spectrometry with artificial intelligence, a team from the Carnegie Institution for Science in the U.S. has pushed the confirmed molecular record of photosynthesis back by nearly 800 million years.
The team, led by astrobiologist Michael Wong and geologist Robert Hazen, did not discover a new fossil. They discovered a new way of seeing. They gathered 406 samples including modern plants, ancient coal, 3.5-billion-year-old chert, and even carbonaceous meteorites. They used a technique called Pyrolysis Gas Chromatography Mass Spectrometry or Py-GC-MS. In simple terms, they heated samples until they vaporized, decomposing the organic matter into its component fragments.
In the past, scientists looked for specific biomarkers. These are single, intact molecules that scream life, such as cholesterol or chlorophyll. But those fragile molecules rarely survive billions of years of geological cooking. Instead of looking for a single needle in the haystack, the Carnegie team used machine learning to look at the geometry of the hay itself.
They trained a type of machine-learning model called a random forest on these molecular fingerprints. The AI did not look for one molecule. It analyzed the distribution of thousands of molecular fragments. It learned to spot the subtle, complex patterns that distinguish the chaos of biology from the order of nonliving chemistry.
Also Read: Scientifically Speaking: A heart drug disguised as a weight-loss injection
The results are very strong. When tested on modern samples, comparing leaves against meteorites, the AI was 100 percent accurate. When distinguishing between fossilized biological samples and abiotic meteorites and synthetic organic mixtures, it maintained roughly 93 percent accuracy. But the real breakthrough came when they applied the model to Earth’s oldest, most debated rocks.
The AI identified signs of oxygenic photosynthesis in the Gamohaan Formation in South Africa. These rocks are 2.52 billion years old. Before this, the oldest biomolecular evidence for photosynthesis was roughly 1.7 billion years old in molecules preserved in rocks. This work helps align the chemical record with the geological record and closes a massive gap in our understanding of when Earth began to breathe.
Even more strikingly, the model found signals of life in the Josefsdal Chert, a rock formation that is 3.33 billion years old. The AI looked at the degraded, scrambled carbon in these stones and identified it as the remains of living things, distinguishing it from meteoritic carbon that often contaminates rocks of that age.
This matters for two reasons. First, it rewrites the opening chapters of life on Earth. Second, it changes how we might hunt for alien life. No matter what science fiction and cinema shows, if we ever find alien life it is not going to be Little Green Men, it is going to be microbial.
Currently, the gold standard for finding life on Mars is sample return. This involves spending billions of dollars to fly rocks back to Earth for analysis in our best laboratories. This new study suggests we might not always have to wait for that logistical feat.
NASA’s Curiosity rover is already equipped with a Py-GC-MS instrument, the Sample Analysis at Mars or SAM suite. It cooks rocks and analyzes the gas. The problem has always been interpreting that data. If Curiosity finds carbon, is it biological in origin or just a nonliving meteorite?
The Carnegie study shows that we do not necessarily need pristine samples to answer that question. The AI model showed it can differentiate between biological carbon and the abiotic carbon found in carbon-rich meteorites with high precision. We could potentially upload this computer brain to our rovers. Instead of sending samples home, we can send software to Mars, though in practice, that would need careful calibration to the rover’s instruments and strict testing.
Also Read: Scientifically Speaking: How AI is invading scientific journals
Science often advances not because we find new things, but because we find new ways to look at old things. We have stared at these South African rocks for decades. Computers are finding the ghosts hiding inside them. But that might just be the start. If a computer can distinguish between a 3-billion-year-old microbe and a space rock on Earth, it might be our best bet for making that same distinction on Mars as well.
Anirban Mahapatra is a scientist and author, most recently of the popular science book, When The Drugs Don’t Work: The Hidden Pandemic That Could End Medicine. The views expressed are personal.