There is a faint electrical spike traveling through the soil beneath your feet right now. It might be nothing—random noise, the hum of a root brushing against a mineral deposit. Or it might be a message. A fungus extending a filamentous thread toward a nutrient source could be signaling to its neighbors, coordinating growth, or warning of a threat. The problem is that we have spent centuries walking over this conversation without ever learning the language.
That may be about to change. A small but growing group of researchers is applying machine learning to the electrical signals produced by fungal mycelium—the vast, underground networks of thread-like hyphae that connect trees, shuttle nutrients, and form what ecologist Suzanne Simard famously called the wood wide web. The goal is ambitious and slightly strange: to translate the biological chatter of fungi into something humans can read.
The Underground Internet, Minus the Metaphor
The comparison between mycelial networks and the internet gets thrown around loosely, but the structural parallels are real. Mycelium forms a distributed, branching topology. Information and resources move across nodes. Connections strengthen with use and atrophy without it. Trees use these networks to share carbon, dispatch chemical warnings about insect attacks, and even funnel nutrients to offspring.
But the electrical dimension of this network is where things get genuinely strange. Fungi generate measurable spikes of electrical potential along their hyphae—small but distinct voltage changes that propagate through the mycelium. Andrew Adamatzky, a professor at the University of the West of England, has spent years recording these signals using pairs of electrodes inserted into substrate colonized by species like Pleurotus djamor (the pink oyster mushroom) and Cordyceps militaris. His findings suggest the spikes aren’t random. They appear in patterns, clusters, and rhythmic sequences that resemble, at least superficially, the action potentials of neurons.
“The fungal electrical activity is not noise. It carries structure, and structure implies information. The question is whether that information is meaningful to the organism or merely a byproduct of its biology. We may never know for certain, but the patterns demand investigation.” — Andrew Adamatzky, Unconventional Computing Laboratory
This is where the research moves from curious to compelling. If the signals carry information, then the logical next step is to ask: what kind? And that is precisely the question machine learning is being built to answer.
Teaching Algorithms to Listen to Spores
Adamatzky’s approach borrows directly from natural language processing. The idea is straightforward in concept and fiendishly difficult in execution: record the electrical spikes, segment them into discrete “words,” cluster those words into potential “sentences,” and then use a classifier to determine whether the sequences follow rules that distinguish language from random noise.
The pipeline looks something like this:
- Signal capture — Electrodes implanted in fungal substrate record voltage fluctuations over hours or days.
- Spike detection — Algorithms identify individual electrical events above a noise threshold, isolating them from background activity.
- Feature extraction — Each spike is characterized by amplitude, duration, frequency, and the interval between it and the previous spike.
- Pattern clustering — Unsupervised learning models group similar spikes together, treating each cluster as a candidate “word.”
- Sequence analysis — The order and spacing of words is analyzed for grammatical structure—rules that govern how the signals combine.
The results have been provocative. Adamatzky’s team has reported that the number of distinct “words” in fungal electrical communication ranges from roughly 15 to 50 depending on the species, and that the sequences show statistical properties consistent with natural languages—specifically, patterns that follow Zipf’s law, the empirical rule that the frequency of a word in a language is inversely proportional to its rank. When random noise is analyzed this way, it does not produce Zipf-like distributions. Language does.
Whether that means fungi are saying something in the way we understand speech is a separate and harder question. But the statistical signature is real, and it gives machine learning something to grip.
The Problem of Meaning
Here is where a responsible account of this research has to slow down. Finding statistical structure in a signal is not the same as decoding meaning. The history of attempts to find language in non-human systems is littered with overreach.
Consider the cautionary precedents:
- Dolphin communication — Researchers spent decades recording dolphin whistles, found rich structure, and still cannot say with confidence what most of it means.
- Plant electrical signaling — Documented since the 1970s, but the leap from “signal” to “intentional communication” remains contested.
- Bacterial quorum sensing — Demonstrably a communication system, but chemical, not electrical, and far simpler in scope.
Fungi occupy a genuinely murky middle ground. They are not animals with centralized nervous systems, but they are also not passive. They make decisions—about where to grow, when to fruit, how to allocate resources across networks that can span hectares. The electrical signals correlate with these decisions. What we lack is a Rosetta Stone that connects a specific spike pattern to a specific biological outcome with enough consistency to call it a vocabulary.
Machine learning excels at pattern recognition but cannot, on its own, establish semantic meaning. A model can tell you that Spike Pattern A resembles Spike Pattern B. It cannot tell you that Pattern A means “there is a nitrogen source to the east” unless you already have ground truth data linking that pattern to that outcome. And right now, we largely do not.
Why It Matters Anyway
So why should anyone care about an incomplete translation project aimed at organisms most people only think about when they’re on a pizza?
Because the implications of even partial success are enormous. If fungal electrical signals can be mapped to biological states—nutrient availability, stress responses, growth coordination—it would transform how we monitor ecosystems. Mycelium is everywhere. A network of cheap electrodes in soil could theoretically read real-time biological conditions across forests, agricultural fields, and restoration sites. You would be listening to the landscape report on itself.
There are also stranger frontiers. Adamatzky’s Unconventional Computing Laboratory has explored using fungi as living computing substrates—exploiting their electrical behavior to perform logical operations, routing signals through mycelium the way you might route them through a circuit board. It sounds absurd until you remember that slime molds have already been used to optimize urban transit network design, and that biological computing is an active, if niche, field.
The deeper pull, though, is philosophical. We have built our civilizations around the assumption that intelligence and communication belong to creatures that look and act more or less like us. Fungi challenge that at the root. They are decentralized, distributed, ancient—they were solving the problems of networked resource allocation hundreds of millions of years before anything resembling a human brain existed. If they are communicating, even in some limited, alien grammar we can barely conceptualize, then the category of “things that talk” just got considerably wider.
The soil beneath us may turn out to be louder than we ever imagined. We are only just learning how to listen.
