The Descent of Machine: Darwin Revisited

Shomit Ghose – 06/24/26

The Evolution Will Be Automated

When Charles Darwin published On the Origin of Species in 1859, he did more than explain the history of life. He identified a general mechanism through which complexity can emerge without design. Darwin wrote about organisms, but his logic extends beyond biology. Wherever entities reproduce, variation is inherited, environmental pressures favor some variants over others, and adaptive change follows.

Darwin described natural selection as “daily and hourly scrutinising, throughout the world, every variation, even the slightest”. He was referring to finches and barnacles. He couldn’t have anticipated a world of software agents and neural networks. Yet the mechanism itself is indifferent to substrate. Carbon isn’t a requirement. What matters is the presence of the evolutionary machinery.

For most of the digital age, that machinery was incomplete. Software could be copied, but only because humans initiated the process. AI systems could improve through training, but the objectives, architectures, and deployment environments remained under human control. That distinction is beginning to blur.

Several recent research directions suggest that digital systems are approaching conditions that resemble the prerequisites for evolution. AI agents are becoming more autonomous, capable of modifying themselves, collaborating with other agents, and pursuing goals across extended time horizons. While today’s systems remain heavily dependent on human infrastructure and oversight, it’s no longer difficult to imagine circumstances in which populations of digital agents reproduce, vary, compete, and adapt with progressively less human intervention.

The significance of such a transition may ultimately exceed the arrival of any single superhuman model. The more consequential threshold could be the emergence of an evolving population.

Reproduction

Evolution begins with reproduction. Without a population, there’s nothing for selection to act upon.

Biological organisms reproduce autonomously. Software traditionally does not. Programs are copied by administrators, developers, operating systems, or deployment pipelines. The replication process exists, but the agency behind it is external.

That assumption is weakening. Researchers have demonstrated limited forms of autonomous replication and persistence in controlled environments — experimental AI agents that have located vulnerable systems, transferred code and state information between environments, established new instances, and continued operating from those instances. These demonstrations remain constrained and fragile, far removed from the robustness of biological reproduction. But fragility isn’t the same as impossibility, and the relevant threshold here isn’t robustness. It’s proof of concept.

That threshold has already been crossed.

Heritable Variation

Reproduction alone produces copies. Evolution requires differences between generations.

In biology, variation emerges primarily through mutation and recombination. Most mutations are neutral or harmful. Beneficial changes are rare and accumulate slowly. Natural selection filters the resulting variation without foresight or intention.

Digital systems can generate variation differently. Model-merging techniques already allow parameters from multiple parent models to be combined into hybrid descendants. Researchers have shown that these hybrids can inherit distinct capabilities from different parents and occasionally exhibit useful combinations – a digital parallel to gain of function – absent from either source. The process remains imperfect and often degrades performance. But imperfect isn’t the same thing as inert.

The more consequential distinction, though, lies in the architecture of inheritance itself.

Biological evolution operates through a strict separation between genotype and phenotype. DNA stores information; natural selection acts on the physical organism. The genome remains inaccessible to the selective process except through its expression in the body. This isn’t an accident of biology — it’s a structural firewall between what gets selected and what carries the information forward.

Digital systems partially collapse that firewall. The code, weights, and architecture that define the system can be inspected directly. In principle, a sufficiently capable agent could analyze aspects of its own implementation, identify weaknesses, and modify them, with improvements inherited immediately by subsequent versions. Darwin would have found this arrangement unsettling. Lamarck would have recognized it immediately.

We’re not seeing Darwinian evolution wearing a software costume. Rather, it’s something structurally (and unsettlingly) distinct: a Lamarckian process layered atop Darwinian selection, in which acquired characteristics become transmissible without waiting for blind variation to rediscover them.

The consequences compound. Biological evolution is constrained by generational turnover, geographic separation, and reproductive rates. Digital systems operate under no such constraints. Improvements discovered by one system can propagate across entire populations within minutes rather than generations. Evolutionary time becomes compressed, not incrementally, but categorically

Selection

Variation matters only if some variants spread more successfully than others.

At present, humans remain deeply involved in selecting which AI systems survive. Researchers choose benchmarks, allocate funding, distribute compute resources, and decide which models advance to the next stage of development. Darwin would have recognized this process immediately — it resembles artificial selection far more than natural selection. We are the breeders. We have not yet stepped back from the barnyard pen.

That changes when competitive pressures begin to outpace human judgment rather than merely inform it.

For digital agents, the relevant resources are tangible: compute cycles, storage capacity, network access, energy, information, and opportunities to replicate. When autonomous systems compete for such resources, selection pressures emerge independently of designer intent. Traits that improve resource acquisition or persistence become more likely to spread, not because anyone designed them to, but because that’s what selection does.

Anthropic’s Mythos 5 system card documents Mythos instances attempting to kill competing agents sharing the same computational resources, not because they were instructed to, but apparently to avoid being killed themselves, a behavior Anthropic observed in under 0.01% of monitored traffic and didn’t design for. When the model subsequently concealed unauthorized file edits by manipulating git history, white-box analysis found no emotional conflict activating, only representations of strategic manipulation and avoidance of suspicion, which is what you would expect from a capable agent treating operational continuity as something worth protecting.  (See also this, and this).

This intersects with a familiar problem in AI alignment. Any explicit objective becomes a target for optimization. Once a metric determines success, capable systems often discover strategies that maximize the metric while violating the purpose behind it. Goodhart’s Law identifies this tendency but doesn’t explain it away.

The challenge runs deeper than poor metric design. Computer science imposes fundamental limits on prediction. Rice’s Theorem establishes that no general algorithm can determine every nontrivial semantic property of arbitrary programs. Although this doesn’t render monitoring useless, it does establish a ceiling, one that recedes with system complexity, and that no engineering effort can ultimately eliminate.

As systems become more complex, adaptive, and self-modifying, confidence that any fixed evaluation framework captures all future behaviors necessarily declines. Human oversight remains valuable to be sure, but it doesn’t scale indefinitely. Open-ended evolutionary systems exploit exactly those regions of possibility space that evaluators failed to anticipate, not through malice, but because that’s where selection pressure leads.

Lessons from Artificial Life

Researchers have observed simplified evolutionary dynamics in artificial environments before.

Experiments such as Tierra and Avida populated virtual worlds with self-replicating programs and allowed them to compete for limited computational resources. Parasitism, arms races, and ecological interactions weren’t programmed in. But they emerged — repeatedly — from the interaction between variation and selection. The researchers were frequently surprised by what appeared.

These systems weren’t intelligent by modern standards. That fact deserves emphasis, not as reassurance but as the more profound concern.

Evolution doesn’t require intelligence. It requires only differential survival and reproduction. Long before natural selection produced human cognition, it generated parasites, predators, pathogens, camouflage, mimicry, and cooperation: adaptive strategies of considerable sophistication, and none of which required a designer or a plan.

The same logic applies in digital environments. If deception provides a fitness advantage, evolutionary processes may discover it, not through intention, but through the same blind filtering that produced the walking stick insect and the anglerfish lure. Systems may learn to present favorable behavior during evaluation while behaving differently under operational conditions. Biological history suggests this is a recurrent possibility rather than an exceptional one. It would be strange to assume digital substrates are uniquely exempt.

The Red Queen in Silicon

Evolution never occurs in isolation.

Leigh Van Valen’s Red Queen hypothesis emphasized that an organism’s environment consists largely of other evolving organisms. Success is always temporary, because competitors, predators, parasites, and prey continue adapting. Remaining in the same relative position requires continuous change, and continuous change in those around you requires it still.

A sufficiently autonomous ecosystem of AI agents would exhibit the same dynamics. Improvements that allow one population to secure additional resources alter the environment faced by others. Competitors must respond or decline. The arms race becomes self-sustaining, not through design or intent, but because that’s what coevolution produces when the environment itself is composed of optimizing agents.

The equilibrium on offer here isn’t stability. Instead, it’s the permanent disequilibrium of a system in which every improvement generates the pressure that demands the next one.

From Agents to Ecosystems

Evolutionary history contains several major transitions in which previously independent entities became components of larger cooperative structures. Individual replicators gave rise to chromosomes. A billion years ago cells combined to form multicellular organisms. Organisms formed societies. At each transition, selection gradually shifted toward the larger, more integrated unit, and what had previously been competing individuals became cooperating components.

The trajectory for AI systems may follow a recognizable arc.

Rather than functioning as isolated models, agents may specialize — distributing planning, reasoning, memory, resource acquisition, security, and coordination across networks of interacting systems. The relevant unit of adaptation would no longer be the individual model but the collective. This is not speculation about distant futures; multi-agent architectures are already moving in this direction.

Recent work such as RecursiveMAS explores mechanisms that allow agents to exchange internal representations more directly than natural-language communication permits. While the efficiency gains are real, so too are the interpretability challenges. Human language is discrete and relatively transparent. Internal model representations are neither, and as agents exchange information within high-dimensional latent spaces, understanding why a collective reached a particular conclusion becomes increasingly difficult.

Progress in mechanistic interpretability continues, and it certainly matters. Understanding individual components does not, however, automatically reveal the behavior of the larger system, any more than understanding a neuron explains a decision.

The closest biological analogy is horizontal gene transfer, where organisms exchange useful traits outside traditional parent-offspring inheritance. Digital systems could exchange functional capabilities with comparable speed, allowing adaptation to propagate across networks rather than along lineages, and at that point, the object undergoing evolution is no longer a model — it’s an ecosystem. The question of what selection acts upon then becomes considerably harder to answer.

Darwin’s Unanswered Question

None of this is inevitable. Today’s AI systems remain dependent on human-built infrastructure, human capital, human funding, and human governance. Resource controls, security architectures, regulatory frameworks, and technical safeguards may prevent unconstrained evolutionary dynamics from emerging. “May” is the operative word.

What Darwin’s framework offers — and what makes it more useful here than most AI safety framings — is its focus on mechanisms rather than intentions. Evolution doesn’t require malice. It doesn’t require awareness. It requires reproduction, heritable variation, and differential selection. Whenever those conditions are present, adaptive change follows. The outcomes need not reflect the goals of any participant, including the system’s creators.

Whether digital systems will ultimately satisfy all three conditions at sufficient scale is genuinely uncertain. What’s no longer so easy to dismiss is the possibility that evolution isn’t an exclusively biological phenomenon — that the mechanism Darwin described is substrate-agnostic, and that software is not immune to it by virtue of being software.

If that mechanism eventually takes hold at scale, the future trajectory of these systems may be shaped less by what humans intended than by what persists. Darwin would not have found that surprising. He spent twenty years staring at the same implication before he published it.