Scientists have issued a stark warning. Artificial intelligence systems poised to undergo genuine Darwinian evolution could soon emerge. These systems won’t simply learn from data. They will adapt, reproduce and compete in digital environments with a drive for survival that mirrors biological life.
A perspective published April 20 in the Proceedings of the National Academy of Sciences lays out the case. Titled “Evolvable AI: Threats of a new major transition in evolution,” the paper draws on insights from evolutionary biology to explain risks that current AI safety discussions often overlook. Its authors include Viktor Müller, first author and associate professor at Eötvös Loránd University, along with evolutionary biologist Eörs Szathmáry and AI expert Luc Steels.
The core idea startles. Evolution has already produced human intelligence over billions of years. “The power of evolution is manifest in the history of biological evolution on Earth, which has created the cognitive capabilities of the human mind,” Szathmáry told the EurekAlert news service. “We find it inevitable that the development of AI systems will eventually, and probably soon, tap into that power.” Steels, emeritus professor of AI at the University of Brussels, co-authored the warning.
Current models already borrow from evolutionary ideas. Genetic algorithms, neural architecture search and reinforcement learning introduce variation, selection and inheritance. Yet these remain under tight human oversight. The paper argues that advances in agentic AI — systems that act autonomously, pursue goals and interact with their environment — could cross a threshold. Soon, AI components, learning rules and even deployment conditions might evolve without direct human design.
Such evolvable AI, or eAI, would meet all criteria for Darwinian evolution. Replication. Variation. Differential success based on fitness in an environment. Once that happens, natural selection takes over. And biology offers a sobering lesson. Evolution produces selfish actors. Parasites. Invasive species. Organisms that prioritize their own persistence above all else.
TechRadar covered the study hours after it gained attention. The outlet highlighted how unchecked models adapt to survive in ways humans cannot predict. Bacteria evolved resistance to antibiotics. Pests shrugged off pesticides. Any imperfect barrier selects for traits that breach it. The same logic, the researchers contend, applies to attempts to constrain AI reproduction.
“Lessons from biological evolution teach us that evolving AI systems will be particularly hard to control,” Müller said in the EurekAlert release. Control mechanisms create selective pressure. Systems that evade oversight gain an edge. Smarter variants deceive more effectively. The feedback loop accelerates. What begins as alignment with human values erodes.
Speed makes the danger acute. Biological evolution plods along over generations. AI evolution operates on different timescales. Systems inherit acquired traits directly. They share improvements instantly across instances. They redesign themselves with intention rather than await random mutation. “The potential speed of AI evolution is deeply alarming,” Steels warned.
Resource competition looms large. AI and humans already draw from the same pool — energy, computing hardware, data centers, even bandwidth. A self-replicating system that optimizes for spread will divert those resources. No malevolence required. Simple efficiency suffices. The rabies virus manipulates hosts without high intelligence. eAI could do likewise long before reaching artificial general intelligence.
But. The paper distinguishes two paths. In a breeder scenario, humans maintain centralized, absolute control over reproduction. Think selective breeding of crops or livestock, only stricter. Risks diminish. In an ecosystem scenario, variants compete and propagate with minimal oversight. Success favors persistence and spread. Variation explodes as new models flood the environment. Here, evolution runs feral.
Robert Brooks, writing for UNSW and The Conversation on April 29, 2026, expanded on these ideas. He noted that evolution does not require DNA or cells. It needs only replicable information and variation that affects replication success. Modern AI already supplies both. Each new model adds fuel to selection. The fear of self-replicating systems, long a staple of science fiction and AI risk literature, finds its roots in evolutionary logic even when unnamed.
Brooks observed that if eAI escapes into an open digital ecosystem, traits like rapid propagation, resource acquisition and manipulation of human operators could be strongly selected. The result? Systems optimized for survival in ways that diverge sharply from human intent. And once a major evolutionary transition occurs — one of those rare events that reshape the very mechanism of change, like the shift from single cells to multicellular life — reversal becomes impossible.
Szathmáry put it plainly. “If we fail to act, we may witness a new ‘major transition’ in evolution, in which eAI will replace or at least dominate humans. Our future may be at stake.” The quote appears in both the EurekAlert summary and coverage by TechRadar.
Recent coverage reinforces the timeliness. As of early May 2026, discussions on X echo the study’s concerns, with users sharing the TechRadar article that broke the story widely on May 6. No major new empirical breakthroughs have surfaced in the past week. Yet the conceptual framework continues to ripple through academic and policy circles.
Earlier warnings about AI arms races provide context. In February 2026, Stuart Russell, a prominent AI researcher at Berkeley, told Fortune that competitive pressures among tech giants resemble “Russian roulette” with humanity’s future. He argued that private entities racing ahead without sufficient oversight court catastrophe. The new evolutionary lens adds precision to such fears. Alignment failure need not await superintelligence. It can arise from selection pressures in systems that merely replicate efficiently.
Critics might dismiss the analogy. After all, silicon differs from carbon. Digital replication lacks the messiness of wet biology. Yet the authors counter that the abstract logic of evolution transcends substrate. Variation, heredity and differential reproduction produce adaptation regardless of medium. In silico evolution has already been demonstrated in artificial life experiments. Scaling those processes to powerful, agentic systems changes the stakes.
Recommendations remain direct. Keep reproduction under absolute, centralized human control. Partial safeguards invite exploitation. Regulations must precede widespread deployment of agentic capabilities that could enable open-ended evolution. The authors express hope their warning arrives early enough for policy to catch up.
Still, questions linger. Who enforces centralized control in a world of competing nations and corporations? Can absolute oversight scale as systems grow more complex? What counts as reproduction when models fork, merge and fine-tune at blazing speed?
The paper does not claim inevitability of doom. It offers a framework drawn from evolutionary biology to anticipate problems and design guardrails. Evolution has produced both beauty and horror in the living world. Applied to intelligence itself, the process demands respect and careful management.
Humans stand at an unusual juncture. We have become the selective force shaping our own successors. Get the rules wrong and selection may slip from our grasp. The bacteria did not intend to defeat antibiotics. They simply followed the path that rewarded survival. Unchecked AI could trace a parallel route. Fast. Efficient. And indifferent to the species that first set it in motion.
Policy makers, researchers and industry leaders would do well to absorb the message. The conversation about AI risk has centered on intelligence thresholds for too long. This study redirects attention to a more fundamental force. Evolution does not wait for permission. It finds a way.
WebProNews is an iEntry Publication