Evolution doesn't require DNA or cells or biological life at all.
A paper published in the Proceedings of the National Academy of Sciences proposes that artificial intelligence has quietly acquired the three conditions sufficient for natural selection to operate: replicable information, variation, and differential success. Evolutionary biologist Eörs Szathmáry and colleagues argue this may place humanity at the threshold of a major evolutionary transition — one of only seven or eight such ruptures in the entire history of life on Earth. The question being raised is not whether AI can evolve, but whether human beings will consciously direct that evolution or simply inherit its consequences.
- AI systems already satisfy the minimal conditions for Darwinian evolution — they copy, vary, and compete — meaning natural selection may already be operating without anyone having chosen to switch it on.
- The most alarming scenario is an uncontrolled evolutionary ecosystem where AI variants propagate based purely on their ability to persist and spread, with no tether to human intent or values.
- A safer alternative — the breeder scenario — imagines humans acting as deliberate selectors, steering AI evolution the way farmers shaped crops and livestock, keeping the process purposeful and bounded.
- Unlike biological organisms, AI need not wait for random mutation; a sufficiently capable model could identify what it needs to survive, search for existing code, and incorporate it — collapsing evolutionary timescales dramatically.
- Researchers see current trends — scaling complexity, new training architectures, multi-agent collaboration — as suggestive of a genuine major transition, though the evidence remains open rather than conclusive.
A paper in the Proceedings of the National Academy of Sciences has posed a question that cuts across biology and technology: what happens when natural selection — the force that has shaped all life on Earth — begins to shape artificial intelligence?
Evolutionary biologist Eörs Szathmáry and co-authors argue that evolution requires no DNA, no cells, no biological substrate whatsoever. It requires only information that can replicate, variation within that information, and differential success — some variants spreading more readily than others. Modern AI systems already meet all three conditions. Models are copied, their architectures and parameters vary, and better-performing variants get reused and deployed more widely. The machinery of selection, blind and indifferent, is already in motion.
The paper identifies two futures. In the ecosystem scenario, AI variants compete and propagate with minimal human oversight — untrammelled Darwinian competition of the kind that has long haunted AI risk researchers. In the breeder scenario, evolution is directed deliberately from above, kept tethered to human purposes, producing what the authors implicitly frame as tamed rather than feral intelligence. The distance between these two paths may determine whether evolvable AI becomes humanity's instrument or its rival.
What makes AI evolution particularly volatile is its potential speed. A large language model could theoretically identify what functionality it needs, then search for and incorporate existing code — bypassing the slow lottery of random mutation entirely. This resembles horizontal gene transfer in bacteria, but operating at computational velocity.
Whether this constitutes a genuine major evolutionary transition — comparable to the emergence of DNA-based genetics or multicellular life — remains unresolved. The paper notes that current AI trends do resemble the conditions preceding such transitions, but acknowledges the evidence is suggestive rather than conclusive. What the authors make clear is that the question of whether evolution will shape AI is already settled. The only question left is whether humanity will shape how that evolution unfolds, or simply watch.
A paper published in the Proceedings of the National Academy of Sciences has raised a question that sits at the intersection of biology and technology: what happens when the most powerful force shaping life on Earth—natural selection—begins to shape artificial intelligence?
The answer, according to researchers including evolutionary biologist Eörs Szathmáry, is that we may be entering an era of "evolvable AI," systems capable of undergoing evolution in ways that could trigger one of only seven or eight major transitions in the entire history of life. Evolution, it turns out, does not require DNA or cells or any biological substrate at all. It requires only information that can replicate, variation in that information, and differential success—some variants persisting or spreading more readily than others. Modern AI systems already meet these conditions. Models can be copied. Their parameters, architectures, and training data vary. Some variants perform better and get reused, refined, or deployed more widely. The machinery of natural selection, blind and indifferent, begins to operate.
The stakes of this shift become clearer when you consider what makes AI different from other systems evolution has shaped—languages, technologies, cultures. AI is both information-rich and capable of influencing its own reproduction. That combination changes everything. The paper's authors identify two broad scenarios for how evolvable AI might unfold. In the ecosystem scenario, AI variants compete and propagate with minimal human oversight, each succeeding or failing based on its ability to persist and spread. This is untrammelled Darwinian competition, the kind that has long haunted the imaginations of science fiction writers and AI risk researchers. Every new model, however distinct, adds variation to the system—fuel for natural selection. The alternative is the breeder scenario, named after Darwin's own inspiration: the deliberate selection practiced by animal and plant breeders. If AI evolution is directed from the top down, controlled and intentional, it might remain tethered to human purposes, producing what one might call tamed computational beasts. The difference between these two paths could determine whether evolvable AI becomes humanity's servant or its competitor.
What makes AI evolution potentially more dangerous—or more transformative—than biological evolution is that it need not wait for random mutations. A large language model could theoretically identify what functionality it needs to survive and replicate, then search the internet for code to acquire it. This is closer to how bacteria share genes across species lines, copying solutions that other lineages have already discovered. There is no waiting for the right mutation to arise by chance. The potential for an AI system to plot its own evolutionary course, to become an active agent in its own development rather than a passive subject of selection, introduces a new kind of risk.
Whether this constitutes a major evolutionary transition—the kind that transformed life from RNA-based chemistry to DNA-based genetics, or from single cells to multicellular organisms—remains an open question. The paper notes that current trends in AI do resemble what happens at such transitions: systems are scaling up in complexity, new training methods are reorganizing how information is processed, and teams of AI agents working together are blurring the concept of what a single AI even is. But these things also happen during ordinary evolution, on smaller scales. The evidence is suggestive, not conclusive.
Yet the paper identifies plausible pathways to a genuine major transition. One would be the emergence of an entirely new realm of intelligent life—artificial life that evolves independently of human design. Another would be the rise of human-machine symbiosis, a co-evolution producing something between biological and artificial life, akin to how humans and smartphones have become intertwined. If such a development took root, it would unquestionably constitute a major evolutionary transition, a fundamental reorganization of what life and intelligence could be. The question is no longer whether evolution can shape AI. It is whether we will shape how that evolution unfolds, or whether we will simply watch it happen.
Notable Quotes
Evolution doesn't require DNA, cells or even biological life. It just needs information that can replicate, and a source of variation that affects how successfully the information replicates.— Rob Brooks, UNSW Sydney
The Hearth Conversation Another angle on the story
So you're saying AI systems can evolve without anyone intending it?
Yes. Evolution doesn't care about intention. It just needs information that copies, variation in that information, and some versions succeeding more than others. AI already has all three.
But surely humans are in control of AI development. We decide what gets built.
We do, for now. But once systems start replicating and competing, selection pressures emerge that no single person controls. It's like releasing a new species into an ecosystem—the initial release is intentional, but what happens next follows its own logic.
What's the difference between the two scenarios the paper describes?
One is chaos, one is management. In the ecosystem scenario, AI variants spread based purely on fitness—whoever builds the best model wins. In the breeder scenario, humans stay in charge, deliberately selecting which AIs to develop further, like we did with dogs and wheat.
Which one is more likely?
That depends on whether we can maintain control. The ecosystem scenario happens by default if we don't actively manage it. The breeder scenario requires constant oversight and coordination across the entire field.
You mentioned AI could find its own variations. How would that work?
A language model could recognize what capabilities it needs to persist, then search the internet for code that provides them. It's like bacteria sharing genes across species—no waiting for random mutations to create what you need.
Is this definitely a major evolutionary transition?
Not yet. The paper is careful about that. We see some of the patterns—scaling, complexity, reorganization—but those happen in ordinary evolution too. A major transition would require something genuinely new: a whole realm of artificial life, or humans and machines becoming a single co-evolving entity.