What if artificial intelligence could read billion-year-old rocks like history books—and quietly rewrite what we think we know about the origin of life?
That’s exactly what a new blend of chemistry and machine learning is starting to do, and the implications stretch from Earth’s earliest oceans all the way to the search for life on other worlds.
Researchers have now combined advanced chemical analysis with AI to detect traces of ancient life in rocks as old as 3.3 billion years. Instead of hunting for obvious, intact molecules, this method looks for extremely faint chemical “fingerprints” left behind after original biomolecules have long since broken down. By doing so, scientists can begin to extract hidden biochemical information from rocks older than 1.7 billion years, closing a huge gap in the record of life’s evolution and shedding new light on processes such as photosynthesis—and possibly even guiding how we search for extraterrestrial life.
The mystery of missing biochemical history
Evidence like tiny microfossils and distinctive carbon isotopes suggests that life on Earth was already present around 3.45 billion years ago. Yet, when scientists look for more complex biochemical evidence—actual molecular remnants of that early life—the record becomes surprisingly thin. For a long stretch of Earth’s history, especially before about 1.7 billion years ago, the “molecular diary” of life is almost blank.
The oldest clearly accepted records of complex biomolecules, such as lipids (which help form cell membranes) and porphyrins (important in metabolic pathways like those involved in energy capture), only go back to roughly 1.7 billion years. That means that for about half of life’s known existence, there has been little direct molecular evidence documenting how biology evolved. This enormous gap has left major questions unanswered: How exactly did early metabolism work? When did different types of photosynthesis emerge? And what kinds of organisms were thriving during those lost chapters?
How AI listens to chemical “echoes” of life
To tackle this problem, an international team of scientists turned to analytical chemistry and machine learning, aiming to pull out subtle biosignatures from rocks far older than the usual 1.7-billion-year cutoff. But here’s the twist: instead of searching for specific, well-known biomolecules—like lipids or sterols—they deliberately ignored the usual suspects.
Instead, the team focused on the overall pattern of tiny molecular fragments that form when original organic materials are broken down. Think of it like listening not for a recognizable song, but for the overall pattern of notes and echoes in the noise. These fragments are the decayed remains of once-intact biomolecules, and while any single fragment might not tell you much, the combined distribution of hundreds or thousands of them can carry a hidden signature of life.
To explore this, the researchers assembled a collection of 406 varied samples. These included ancient sediments and fossils, but also modern plants, animals, fungi, and even meteorites. Bringing in such a wide range of material gave the AI plenty of examples of both biological and non-biological chemistry, across many environments and time periods.
Each sample was analyzed using pyrolysis–gas chromatography–mass spectrometry, a technique that essentially heats the material to break it apart, then separates and measures the resulting chemical fragments. The process rips both organic and inorganic components into smaller pieces, producing a complex “fragment spectrum” that can be treated like a chemical fingerprint. These fragment patterns function as “biomolecular echoes”—faint residual signals of molecules that no longer exist in their original form.
Training the AI to recognize life
Once the chemical fingerprints were collected, the team trained a machine learning model on about 75% of the samples. The goal was to teach the system to distinguish between material that originated from living organisms and material that did not, purely based on the patterns of fragments. In addition, the model was trained to decide whether the chemistry reflected photosynthetic activity or not.
The remaining 25% of the samples were held back for testing, to see how well the algorithm could classify previously unseen data. The result was impressive: the model achieved an accuracy between 90% and 100% in telling apart biological from non-biological signatures, and in identifying whether photosynthesis was involved. For such messy and ancient chemical data, that level of performance is striking—and it strongly suggests that life leaves a robust pattern in molecular fragments, even when clear-cut biomolecules are gone.
This is where it gets especially intriguing—and potentially controversial. If AI can reliably read these faint chemical patterns, it means that rocks once dismissed as “too altered” or “too old” for meaningful biomolecular study might suddenly become relevant again.
Ancient rocks, ancient life—and a photosynthesis surprise
When the researchers applied their trained AI to very old rocks, some of the most eye-catching results came from samples collected in South Africa. In 3.3-billion-year-old sedimentary rock, the AI classified the released chemical fragments as biological in origin. In other words, even though the original biomolecules had long since vanished, their decay products still carried a recognizable “life signature.” However, the model did not detect the distinct patterns associated with photosynthetic molecules in that particular rock.
In a different South African sample, around 2.5 billion years old, the story changed. There, the AI identified fragment patterns consistent with photosynthetic molecules. This effectively pushes the chemical record of photosynthesis back by more than 800 million years compared to some previous interpretations, suggesting that photosynthetic processes were firmly established much earlier than many lines of evidence had clearly shown.
The lead researcher described the team’s reaction as astonishment. From a human perspective, these fragments just look like a dense forest of peaks and signals—too complex to interpret by eye. But the AI was able to pick out consistent patterns hidden within that forest. In the distribution of hundreds or thousands of fragment types, the model “reads” whether ancient life was present, and in some cases, what kind of metabolism it likely used.
This has prompted a bold vision: that this chemistry–AI combination could become a standard toolkit in fields like palaeobiology (the study of ancient life on Earth) and astrobiology (the search for life beyond Earth). After all, if the same strategy can decode subtle biosignatures in Earth’s most ancient rocks, why couldn’t it be applied to samples from Mars or other planetary bodies?
A powerful method—but how revolutionary is it really?
Not everyone views the findings as a complete overhaul of our understanding of life’s history. Some experts note that the underlying machine learning techniques are not new in themselves; what is new is applying them in such a sophisticated way to complex geochemical data. In that sense, the innovation lies more in the creative use of existing methods than in inventing an entirely new kind of AI.
One researcher who studies the origins of photosynthesis commented that, while the results don’t dramatically change current theories about how photosynthesis evolved, they demonstrate that this AI-based chemical approach can support and enhance other lines of evidence. In other words, it appears to validate and extend what was already suspected, rather than overturning everything overnight.
And this is the part most people miss: the real game-changer may not be the current conclusions, but the doorway this method opens. If AI-guided chemical fingerprinting can be scaled up and applied to many more samples, it could gradually refine timelines for key events like the emergence of oxygen-producing photosynthesis, and clarify which types of metabolism dominated at different stages of Earth’s early history.
Challenges, next steps, and big ambitions
Looking ahead, scientists are eager to expand this work into a broader range of ancient rocks, particularly those from the Archean eon, which began about 4 billion years ago. Some argue that the method should be applied to every archived sample that might contain early biological signals, in order to pinpoint—more convincingly than ever—when oxygenic photosynthesis first appeared.
That ambition comes with challenges. Ancient environments were chemically and biologically complex, with multiple types of metabolism likely coexisting in the same place and time. Untangling these overlapping signals—especially after billions of years of geological alteration—is no small task. Even sophisticated machine learning models must be carefully trained, tested, and interpreted to avoid misreading non-biological signals as life, or mixing up different metabolic fingerprints.
The lead scientist emphasizes that this project is still in its early days. The team is actively seeking thousands of well-documented samples from diverse locations including Australia, South Africa, Greenland, and Canada. The logic is simple: the more varied the dataset, the more subtle attributes the AI can learn to recognize—whether that’s distinguishing different types of photosynthesis, separating microbial groups like prokaryotes from more complex eukaryotes, or perhaps even identifying entirely unexpected biochemical patterns.
Here’s where it could become truly contentious: if future analyses start to suggest timelines or metabolic pathways that conflict with long-standing interpretations from fossils, isotopes, or sedimentology, whose story should scientists trust—the traditional evidence or the AI’s chemical pattern recognition? And how much weight should we give to an algorithm’s verdict when the “ground truth” is billions of years in the past?
Your turn: is AI rewriting the story of life fairly?
This emerging technique raises some big, debate-worthy questions:
- Do you think AI should play a leading role in interpreting Earth’s deepest biological history, or should it remain a supporting tool behind more traditional methods?
- If AI-driven chemistry suggests that photosynthesis or other metabolisms appeared much earlier than current consensus, would you find that convincing—or would you want multiple, independent lines of evidence first?
- And when we start applying these methods to Martian rocks or samples from other worlds, how confident would you be in declaring “we’ve found signs of life” based largely on patterns that no human can directly see without an algorithm’s help?
Share your thoughts: Is this the beginning of a new era in decoding ancient life—or are we at risk of letting AI over-interpret the whispers of rocks that are too old to speak clearly for themselves?
