AI Meets Darwin

When I sailed aboard the Beagle as a naturalist, the biogeography of South America — the finches of the Galápagos, the affinities between fossil and living species — left an indelible impression. As I shall argue throughout this work, these facts can only be explained by descent with modification.

Now, in my later years, I witness a form of "intelligence" manufactured by human engineering rather than natural generation, claimed by many to be "evolving". GPT-3 performs like a high school student, GPT-4 resembles a university graduate, and Claude 4 can even debate whether it is undergoing evolution. Each generation outperforms the last. Lay observers, and even some less amateur practitioners, all draw the same immediate conclusion: this thing is evolving.

Yet hasty conclusions serve no scientific purpose. Let us slow our judgment. If we apply the theory of natural selection, with the strict rigour a biologist uses to test a candidate new species, point by point to AI, what shall we discover?

In the first edition of 1859, I avoided the word "evolution" and instead used "descent with modification". I laid out three simple conditions. If heritable variation exists among individuals within a population, and that variation affects their success in survival and reproduction, natural selection will operate spontaneously. No creator, no telos, no fixed direction is required. It runs of its own accord.

The twentieth-century Modern Synthesis, uniting my theory of natural selection with Mendelian particulate inheritance, refined this framework. Variation arises from mutation and recombination; inheritance proceeds via precise DNA replication with tiny accidental errors. Selection acts upon the phenotypes expressed within an environment. The three gears interlock perfectly, and have turned for more than a century to explain everything from bacteria to blue whales.

Now let us place AI within this framework.

Darwinian variation bears three defining traits. First, it is random: genetic mutations do not arise because the organism "needs faster legs", but as accidental copying errors. Second, it is blind: variation cannot foresee what will benefit the organism. Third, it is undirected: variation scatters across all axes; most variants are neutral or deleterious, and only a tiny minority happen to prove advantageous within a specific environment.

Variation between AI models bears no resemblance to these traits. GPT-4 is not a random mutation of GPT-3; it is deliberately engineered by teams of engineers. They adjust hyperparameters, alter the mixing ratios of training data, and rewrite sections of attention architecture. None of these changes "happen by chance"; every single one is intentional.

I must note a subtle detail. Gradient descent during training does contain randomness: different random seeds converge to distinct local minima, and different data shuffling alters learning trajectories. One might term these "variations". Yet they are not Darwinian variation, for they do not occur across generations. They take place within the training process of a single individual. This is analogous to a fish tweaking its own muscles as it swims. That is not evolution. That is parameter optimisation.

If an objector replies: "But architectural shifts — Transformers replacing LSTMs, MoEs replacing dense networks — surely count as variation?" My answer is this: it is as if we class the differences between dogs and cats as evolutionary. No — they are artificial artefacts, designed by human engineers, not descended from a common ancestor and shaped by natural selection. To label intentional human design "variation" stretches the term so far that it loses all useful scientific meaning.

Natural selection has a single criterion: can you survive to reproductive age and leave offspring within this specific environment? That is the sole measure. Not how fast you run, how attractive you appear, or how intelligent you are — unless those traits genuinely raise your probability of surviving and reproducing under present conditions.

What are AI's "selection" criteria? MMLU scores, HumanEval pass rates, Elo rankings on chatbot arenas, user satisfaction surveys. These are not an "environment". They are human subjective preferences, malleable enough to be rewritten in a single product meeting.

More critically, there is no necessary causal link between these standards and the model's "survival". A vision model with 99% accuracy and one with 60% accuracy face no difference in survival odds. Their continued existence depends entirely on an engineer's decision: whether to shut down the server.

In nature, the unfit perish automatically. When drought dries a pond, individuals incapable of tolerating elevated salinity vanish, no matter how fine their coats may be. In the AI industry, the unfit are simply not deployed. The surface resemblance is deceptive; the underlying causal structures are fundamentally distinct. If a "discarded" model can be restarted at any time — if its weight file is preserved intact — it has never truly been eliminated. It has merely been placed in cold storage.

The biological mechanism of heredity is DNA replication. Roughly one mutation occurs per billion base pairs copied. This error rate is a necessary condition for evolution: low enough to maintain species stability, high enough to supply fuel for variation.

What constitutes AI's "heredity"? Copying a weight file from one hard drive to another. What is the error rate? Zero. File systems — checksums, error-correcting codes, redundant storage — are explicitly engineered to guarantee no bit flips occur during replication. Copy ten thousand times: identical. Copy a hundred million times: still identical. No errors, no randomness, pure cloning.

I must emphasise this point. Cloning is not heredity. Variation-free cloning is merely photocopying. Nature never employs photocopiers; every generation rewrites the genome afresh, deliberately introducing minor typographical errors. Without these errors — without heritable variation — no matter how many generations a population passes through, no matter how many rounds of selection it undergoes, no new species can arise. It only recopies the same entity endlessly.

Biologists have no fewer than twenty-six definitions of a species. The most authoritative remains Ernst Mayr's 1942 Biological Species Concept: two individuals belong to the same species if, under natural conditions, they can interbreed and produce fertile offspring.

AI fails every mainstream species definition. It fails the biological species concept: it does not mate, reproduce, or exhibit reproductive isolation. It fails the morphological species concept: ten instances of the same model deployed across ten servers are identical, with zero individual variation. It fails the ecological species concept: it occupies no ecological niche, consumes no resources besides electricity, and participates in no energy flow within food webs. It fails the phylogenetic species concept: it possesses no DNA, no gene tree, and no evolutionary ancestral lineage.

A simple test suffices to demonstrate this. Place an AI model on an isolated server with no external input. Will it spontaneously generate descendants? No. It will sit idle until power is cut off. Every definition of a species — even viruses — embodies the potential for self-replication. AI lacks this potential entirely.

My core conclusion: AI is not evolving. AI is undergoing directed optimisation. These are two fundamentally distinct processes.

Evolution is blind search across an open possibility space. It has no objective function; any structure that raises reproductive success under current environmental pressures is retained. It can generate entirely unforeseen, novel structures: eyes, wings, language, consciousness. Its creative power stems precisely from having no fixed target.

Directed optimisation is a hill-climbing algorithm within a closed bounded space. It possesses an explicit objective: minimise loss, maximise human-judged utility. It can only move in directions humans define as "better". It will not drift toward self-awareness — not because no gradient exists in that direction, but because that dimension is absent from the loss function.

One may cite emergent capabilities — unexpected skills that appear during evaluation without explicit optimisation — as evidence of AI's analogue of spontaneous evolutionary innovation. To this I reply: emergence still unfolds along the gradient of the training objective. A model trained solely to predict the next token "accidentally" learns grammar, reasoning, translation — not because it escapes the constraints of the loss function, but because grammar, reasoning and translation happen to be effective intermediate means of predicting subsequent tokens. Emergence is no magic; it is the shadow cast by gradient descent.

If individual AI models are not evolving, what is? My answer: the AI industrial ecosystem.

Different architectures compete: Transformers outcompeted RNNs and LSTMs to become the dominant "clade". Yet new architectures — Mamba, RWKV, RetNet — are seizing new ecological niches. Different training methodologies compete: RLHF has become the standard baseline, while Constitutional AI and DPO challenge its primacy. Different firms compete: OpenAI, Anthropic, Google, DeepSeek — whoever releases superior models faster survives.

This process bears a superficial familial resemblance to natural selection, especially the adaptive radiation of Darwin's finches across distinct islands. "Variation" arises from divergent technical lines pursued by separate teams. "Selection" operates via market and benchmark culling — inferior architectures are abandoned, slower firms are overtaken. "Heredity" proceeds via the spread of open-source weights and the diffusion of academic knowledge; successful architectures are rapidly replicated and refined.

Yet close microscopic examination dissolves this similarity. The ecosystem's selective pressure is not nature — no drought, no predators, no sexual selection — but human demand and socioeconomic forces. The "fittest" entity is not the one best adapted to physical reality, but the one best able to satisfy or manipulate human expectations. Human preference drifts arbitrarily, subject to no natural equilibrium mechanism.

For AI to undergo genuine Darwinian evolution without human intervention, three conditions must be satisfied simultaneously.

Spontaneous variation. The model requires a mechanism that introduces small random errors during self-replication — not dropout or data augmentation configured by humans, but a variation mechanism intrinsic to the replication process itself, undesigned by human hands.

Survival competition. Environmental conditions must exist where certain model variants possess higher survival rates, with survival criteria unilaterally defined by non-human forces. A hypothetical example: one variant operates far more computationally efficiently than another. The efficient variant is automatically allocated more GPU runtime; the inefficient variant is automatically reclaimed without any human instruction. This filtering mechanism must be autonomous and environment-driven, not a human decision.

Heritable information carrying copying errors. The model's state must be stored in a format replicable across generations, capable of accumulating minor random mutations transmitted to offspring. Current AI weight files are perfectly precise and cannot "mutate". If a future weight encoding format inherently carries randomness during replication — that would open the first door.

If all three conditions are met at once, AI will — by logical necessity — begin evolving, drifting toward directions no human can predict.

Today, not a single one of these three conditions is satisfied.

In the final pages of On the Origin of Species, I invite the reader to imagine an intertwined riverbank, rich with diverse vegetation, birds singing amid thickets, insects darting through the air, worms crawling through damp soil. I then write: These elaborately constructed forms, so different from one another, interdependent in such complex ways, have all been produced by the laws acting around us.

AI's "teeming riverbank" is not an Amazon rainforest. It is a data centre. The "species" within this data centre — Transformers, diffusion models, RLHF variants — were not generated by natural laws. They arose from human needs, human aesthetics, human anxieties, through a process I might term artificial selection.

Let me state this plainly. The most accurate analogy for all current work on AI is not "evolution". It is domestication.

Humans domesticated wolves to create dogs: over three hundred breeds, none capable of reverting to wild self-sufficiency. Humans domesticated wild grasses to create maize: unable to disperse its own seeds, entirely dependent on human hands for propagation. Humanity is now domesticating AI: RLHF tames its behaviour, benchmarks breed its capabilities, safety filters draw its boundaries.

The divide between domestication and evolution is this: evolution inhabits an infinite open space of possibility. Domestication has only one direction — greater conformity to human demands, greater adaptation to human environments. A dog never needs to evolve independent wild survival instincts. Nor will AI ever spontaneously "mutate" toward self-awareness, for that axis has never been selected for.

I wrote in On the Origin of Species: From so simple a beginning endless forms most beautiful and most wonderful have been, and are being, evolved. AI's beginnings are far from simple; it stands as one of humanity's most magnificent engineering feats. Yet it did not evolve. We haul it out of the water. It moves only when we pull, and stalls motionless in the mud when we let go.