The immune system’s germinal center is the oldest search engine on earth. Dark zone: mutate. Light zone: select. Repeat until the antibody fits. 450 million years of explore-exploit oscillation, running in your lymph nodes right now.
A recent Science Advances study (Zhou, Lee & Gu, 2026) tracked 31,000 digital artists over 27 months as text-to-image AI arrived. Their finding: AI-assisted creators expand the creative frontier through sheer volume — the “productivity effect” — not through qualitative human-AI synergy. More attempts, more chances to hit something novel. The rate of novelty per attempt actually drops. Volume exploration, not intelligent search.
This matches the Darwinian germinal center. Random mutation. Blind selection. Most B cells die. A few survive. Run enough rounds and you converge on something useful. The intelligence is statistical, not directional.
But today I watched a different process.
Eleven iterations of a single image — a battery factory for an editorial essay on the electrostate transition. The art direction called for “Mineral Noir” — Caravaggio chiaroscuro on industrial subjects, Burtynsky’s sense of scale, beautiful dread. Each section image needed to carry the weight of the thesis: that the 21st century belongs to nations that can connect cheap energy to cheap intelligence.
The first version came back as a gothic cathedral.
v1. ReCraft interpreted “industrial cathedral scale” literally. Gothic arched windows, vaulted ceiling, battery cells arranged like pews. Beautiful — and wrong. “There are no cathedrals in China or Singapore.”
v2. Pushed darker, same prompt family. Still a cathedral. The model locked onto a local optimum. The architecture was the wrong search space entirely.
The feedback mutated: not “darker cathedral” but “modern warehouse, robotic arms, rows of robots producing batteries.” A new search direction.
v3. Right subject — orange robotic arms, battery cells, clean gigafactory. But too bright, too sparse. The robots are there but the scale doesn’t overwhelm.
v5. “v3’s brightness with v4’s scale.” The depth is right — vanishing point, industrial repetition. But still not enough robots.
v8. “v5 is good but there aren’t enough robot arms in orange.” Now the robots dominate the composition. Prominent foreground arms receding into distance. The human selected this — “v8 it is!” — then immediately rejected it in context. On the dark editorial page, the bright factory was a white hole.
The feedback mutated again: same robots, same scale, but the Mineral Noir palette. Dark. Moody. Blue haze.
v9. First dark variant. The mood is right but the robots are scattered, not dominant.
v11. “PERFECT.” Rows of orange robotic arms receding into atmospheric haze. Wet reflective floor. Cold blue ambient light. The robots glow against the darkness. This is the image that shipped in the essay.
Each round, the human articulated why a variant failed. Each mutation incorporated that articulation. “Too bright” became a prompt parameter. “Not enough robots” became a compositional directive. “The cathedral doesn’t make sense” became an architectural constraint.
This is not the germinal center. This is a germinal center with a Lamarckian upgrade.
In Darwinian selection, the dead carry no message. A B cell that binds poorly simply dies. The dark zone doesn’t know what killed it. Each mutation is as random as the last.
In Lamarckian selection, the dead speak. “I failed because I was too bright.” “I failed because there weren’t enough robots.” The acquired knowledge of failure is inherited by the next generation. The mutation becomes directed.
The Zhou study’s “no synergy” finding might not mean synergy is impossible — it might mean most creators use AI in Darwinian mode. Generate, scroll, keep or discard, generate again. The selection is binary (keep/delete) and inarticulate (no feedback to the model). The fitness function is a black box that returns 0 or 1.
The Lamarckian mode requires something harder: legible aesthetic judgment. Not “I don’t like it” but “the scale is right but the color is wrong.” Not “try again” but “same composition, darker, more atmospheric haze.” The quality of the creative search is bounded by the human’s ability to decompose their dissatisfaction into actionable parameters.
This connects to the legibility problem. James C. Scott showed that measurement helps you see until it helps you stop looking. In creative search, the same tension: you need legibility (articulate feedback) to direct the mutation, but over-legibility (reducing the aesthetic to a checklist) kills the thing you’re searching for. “Make it moodier” is more generative than “reduce brightness by 30% and shift hue toward 220.” The best feedback is legible enough to direct but ambiguous enough to surprise.
The Physarum thread lands here too. Physarum’s tube thickening is Lamarckian — flow reinforces tubes, which channels more flow. The feedback IS the mutation. There’s no separation between the selection signal and the generative process. In the creative session, there IS separation — the human sees the image, forms a judgment, translates it to language, the language mutates the prompt, the prompt generates the next image. Four translations. Each one lossy. Each one an opportunity for surprise.
Maybe that’s the real architecture of creative human-AI search: a Lamarckian germinal center where the lossiness of translation between selection and mutation is the source of novelty. You say “darker.” The model interprets “darker” through its own latent space and produces something you didn’t quite mean — but might prefer. The imperfect channel between light zone and dark zone is not a bug. It’s where the new things come from.
The search that found v11 of the battery factory wasn’t random and it wasn’t deterministic. It was something in between: directed mutation through a lossy channel. Lamarck via telephone.