Stigmergic coordination is truth-blind across every substrate where it operates. The pheromone doesn't know what it means, and neither does the system that accumulates it.
Ahrefs analyzed nine hundred thousand newly created web pages in April 2025 and found that 74.2 percent contained AI-generated content. Each page becomes training data for the next model that generates the next page. The web is a pheromone trail writing itself.
The trail metaphor is not a metaphor.
The Trace
Pharaoh's ant foraging trails have half-lives of eight to twenty-five minutes depending on the substrate. Attractive pheromones decay in roughly thirty-three minutes. Repellent pheromones last more than twice as long, around seventy-eight minutes. The difference matters: the warning signal persists after the food signal fades.
No individual ant evaluates the trail. Each follows the concentration gradient, deposits its own pheromone, and walks on. The gradient decays at a rate determined by the medium, not the message. Accuracy is irrelevant. A trail that led to food three hours ago still attracts followers if the pheromone has not evaporated. A trail that leads nowhere accumulates signal if enough ants walk it.
This is what makes stigmergic coordination both powerful and dangerous. The pheromone does not know what it means.
Five Substrates
The pattern repeats across every medium where traces accumulate.
Citations. Simkin and Roychowdhury studied misprints in scientific references and estimated that roughly four of every five citations are copied from other papers' reference lists without consulting the original source. The identical misprints propagating through the literature proved it: scientists cite what other scientists cited. The citation is a pheromone pellet. Each deposit strengthens the trail to the cited paper regardless of whether anyone verified the claim it contains. Papers become renowned through accumulation, not through the quality of the underlying work.
Credit ratings. Rieber and Schechinger found that when one credit rating agency changes a rating by one notch, rival agencies follow by an average of 0.4 notches. The convergence is not driven by independent analysis reaching the same conclusion. It is driven by the reputational safety of agreement: if three agencies are wrong together, none is individually liable. The rating is the pheromone. The substrate is institutional reputation. The decay rate is measured in quarters, not minutes.
AI web content. The Ahrefs finding sits at the inflection point of a recursive loop. Language models trained on web text generate new web text that becomes training data for the next generation of models. Researchers have shown that when synthetic content exceeds a critical fraction of the training corpus, model outputs converge and diversity collapses. The technical term is model collapse. The phenomenon is pheromone saturation. When the trail overwhelms the environment, the medium and the signal become indistinguishable.
Legal precedent. For forty years, Chevron deference instructed federal courts to defer to agency interpretations of ambiguous statutes. The doctrine was cited nearly eighteen thousand times. On June 28, 2024, the Supreme Court overturned it in Loper Bright Enterprises v. Raimondo. The generative mechanism was removed. But the trail persists: thousands of regulations, judicial interpretations, and administrative procedures built on Chevron deference continue to structure the legal landscape. The pheromone outlived the ant that laid it.
The Decay Rate
What unites these substrates is not stigmergy itself. The Mound, published here in February, explored how traces become structure. The deeper finding is that stigmergic coordination is truth-blind across every substrate where it operates.
The pheromone does not know what it means. The ant following the trail does not verify that food is still there. The scientist copying a citation does not read the original paper. The rating agency converging on a competitor's assessment does not run independent analysis. The language model generating web text does not evaluate whether its training data was accurate. The judge applying Chevron deference did not re-examine the agency's reasoning.
In each case, the accumulated signal coordinates behavior without anyone in the system checking whether the signal is true. The coordination works. The question is whether it works toward anything real.
The answer depends on the decay rate. Ant pheromones evaporate in minutes, limiting the damage of a false trail. Citations persist for decades, allowing a fabricated claim to accumulate thousands of followers. Credit ratings reset over quarters, fast enough to correct errors after a crisis but slow enough to amplify herding before one. Legal precedent operates on the scale of generations, meaning the trail can outlast the reasoning that created it.
AI-generated web content introduces a timescale the system has not encountered before. The pheromone is deposited and consumed in the same medium. The trail and the ground are the same thing. And the decay rate is effectively zero: once published, web content persists until actively removed.
The question for every system that runs on accumulated signals is the same one the Pharaoh's ant answers with chemistry: how long should the trail last? Too short, and coordination fails. Too long, and the trail outlives its truth.
The web just chose the longest possible half-life for the least verified signal in history.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)