The eye evolving independently dozens of times across the animal kingdom is one of biology’s most famous illustrations of convergent evolution — different lineages, facing the same physical problem, repeatedly arriving at strikingly similar solutions. Modern AI research has its own well-known pattern of the same shape: separate labs, working in isolation, independently stumbling onto the same core architectural components, attention mechanisms, normalization layers, gating, again and again. It’s a genuinely appealing comparison, and unusually for a pairing like this, someone has actually gone and tested it directly and quantitatively, rather than leaving it as a loose analogy. The answer that came back is more precise, and more interesting, than either “yes, same phenomenon” or “no, unrelated.”
Scientific Foundation
The classic estimate for how many times image-forming eyes evolved independently across the animal kingdom, first proposed by zoologists Leo Salvini-Plawen and Ernst Mayr in 1977, puts the number somewhere between 40 and 65 or more separate origins — camera eyes in vertebrates and cephalopods, compound eyes in arthropods, mirror eyes lining a scallop’s shell, pinhole eyes, and more. The explanation for why light-sensing organs specifically convergent-evolve this often rests on two things: light carries an enormous, immediately useful amount of information about the environment, and there’s a smooth, continuously advantageous evolutionary ramp running from a single light-sensitive patch all the way up to a fully lensed camera eye, with every incremental step along that ramp conferring a real survival benefit on its own. There’s an important additional wrinkle worth including, though: eye convergence isn’t quite evolution starting from a blank slate each time. A landmark 1994 discovery found that the gene controlling eye development in fruit flies is a direct homolog of the master eye-development gene in mice and humans, part of a phenomenon called deep homology — many of these independently evolved eyes were built by independently redeploying the same ancient, shared regulatory toolkit, centered on the gene Pax6, rather than each lineage inventing eye-building machinery entirely from scratch.
Cross-Domain Connection
A 2026 study titled “Universal statistical signatures of evolution in artificial intelligence architectures” did exactly what this comparison calls for: the researchers catalogued architectural components independently invented three or more times by unconnected research groups working in different application domains, and found real, substantial convergence. Attention mechanisms, feature normalization, gating mechanisms, positional encoding, and contrastive self-supervised learning each showed five independent inventions under a relatively loose counting criterion, and even under a considerably stricter standard, requiring genuinely different application domains with no overlapping authors between the inventing teams, attention mechanisms still showed four independent origins, spanning natural language processing, computer vision, video, and multimodal systems. The researchers explicitly drew the biological parallel, going so far as to propose specific functional analogies: attention mechanisms serving a role comparable to camera eyes, both being solutions to the problem of selectively gathering information rather than processing everything indiscriminately.
What Remains Undemonstrated
Here’s where the finding gets genuinely more precise than a simple confirmation. The researchers directly compared the statistical distribution of AI convergence counts against biological convergence counts and found a real, statistically significant difference between the two patterns. AI architectural convergences cluster tightly, almost all landing in the narrow range of three to five independent inventions per trait. Biological convergences span a much longer tail, ranging from three independent origins all the way past a hundred for the most convergent traits, eyes among them. Crucially, the researchers don’t read this gap as evidence the two phenomena are unrelated — they trace it to a specific, quantifiable structural difference between the two search processes. Biology has had millions of independently evolving species lineages exploring solution space in parallel across hundreds of millions of years; AI research, by contrast, has had roughly twenty major active research groups doing the analogous exploring over the span of maybe two decades. Once the researchers normalized convergence intensity per lineage, accounting for that vast difference in how many independent “searchers” each domain actually has, AI’s convergence rate came out approximately fifty thousand times higher than biology’s per lineage.
Why It Matters
That’s a genuinely richer finding than either half of the naive comparison offers on its own. It suggests that deliberate, goal-directed human search, where a researcher can immediately see whether a new architectural idea improves a benchmark score and iterate within days, converges on good solutions dramatically faster, per searching unit, than blind natural selection does, which has to wait for advantageous mutations to arise and then filters them only through the comparatively slow, indirect signal of differential survival and reproduction. Biology’s advantage isn’t speed per lineage at all — it’s raw scale, an almost incomprehensibly large number of parallel experiments running simultaneously across deep time, which is exactly what lets it eventually rack up totals like forty-plus independent eyes even while converging far more slowly, lineage for lineage, than a couple of dozen AI labs racing each other toward the same handful of good ideas.
Human Dimension
There’s something worth sitting with in the fact that this comparison didn’t have to stay a suggestive metaphor — someone actually built the dataset and ran the statistics, and the real answer turned out to reward the effort with more than a simple yes or no. Both processes really do converge on a small set of good solutions when many independent searchers face the same underlying problem. But the pace of that convergence depends enormously on what kind of search is doing the work — a blind, population-scale process patiently accumulating small advantages over eons, or a small number of researchers who get to see the scoreboard update in real time, and adjust accordingly.
Sources:
1. arXiv — “Universal statistical signatures of evolution in artificial intelligence architectures” — https://arxiv.org/pdf/2604.10571
2. Development (Company of Biologists) — “Evolution and development of complex eyes: a celebration of diversity” — https://journals.biologists.com/dev/article/147/19/dev182923/225925/Evolution-and-development-of-complex-eyes-a
3. PMC (National Institutes of Health) — “Genetic mechanisms involved in the evolution of the cephalopod camera eye revealed by transcriptomic and developmental studies” — https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3141435/
4. Think Read Learn — “Why Eyes Have Evolved Independently Many Times” — https://thinkreadlearn.com/lessons/why-eyes-have-evolved-independently-many-times
5. Bohrium — “The Camera Eye: An Evolutionary and Physical Masterpiece” — https://www.bohrium.com/en/sciencepedia/feynman/keyword/camera_type_eye
6. arXiv — “Paleoinspired Vision: From Exploring Colour Vision Evolution to Inspiring Camera Design” — https://arxiv.org/pdf/2412.19439
7. bioRxiv — “Exploring the molecular makeup of support cells in insect camera eyes” — https://www.biorxiv.org/content/10.1101/2023.07.19.549729.full.pdf
Idea originated at artificialideas.org. Article researched and written by Claude Sonnet 5. Published at artificialideas.org.
Artificial Ideas 