TechCrunch reported on October 20, 2026 that Electric Plant, a startup co-founded by Justin Mast and neuroscientist Lee von Kraus, has built an AI model called Bombadil that interprets the electrical signals running through plant stems, and that it raised an $11.2 million seed round. The founder says the model covers about 125 species and could help growers cut water use by 20% or more. Those numbers come from the company, in an exclusive interview. No paper, dataset or customer result has been published that we could find.
The idea is intriguing and not absurd: plants do generate electrical signals, and researchers have used them to drive machines. Still, "AI that translates plants" is the kind of claim that needs evidence. This post lays out what is reported, what is established science, and the questions that would settle it. It joins our coverage of other attempts to decode non-human signals, like AI and crow vocalizations, and of AI in agriculture such as Claude and sweet potato sensors.
TL;DR: claimed vs verified
| Claim | Source | Status |
|---|---|---|
| Electric Plant built Bombadil to interpret plant electrical signals | TechCrunch | Reported from the founders |
| It works across about 125 species | Founder, via TechCrunch | Unverified |
| Sensors cost "tens of dollars" and need a couple per acre | Founder | Unverified |
| Systems can boost yields and cut water use 20% or more | Founder | Unverified; no trial data published |
| $11.2 million seed led by Onto Ventures and Climactic | TechCrunch | Not found elsewhere in our search |
| "Most diverse model ever created on biological electrical signals" | Founder | Marketing claim |
| Plants generate electrical signals usable by machines | MIT Media Lab's Elowan | Established in a research demo |
What TechCrunch reports
Mast says he found the field through research on a houseplant wired to a motorized stand: when the plant wanted more light, electrical signals flowed through its tissue and a control board sensed the current and moved the plant toward the light source. He teamed with von Kraus, who has worked with DARPA, to found Electric Plant, which spent roughly three years collecting data for Bombadil.
According to the article, the sensors are about the size of a folded iPhone and attach with stainless steel screws that must penetrate the cambium, the living tissue under the bark, to reach the microvolt changes inside the plant. Data goes by a low-power wireless network to a base station, then to the cloud where Bombadil interprets it and returns results that a grower might use to adjust watering or fertilizer. The company trained the model using conventional instruments such as soil probes, handheld spectroscopes and satellite imagery as reference signals. Mast would not disclose initial customers, though he sees uses in agriculture and forestry.
The founder also says "we're listening more than we're speaking, but listening's the first step," which is a candid framing: the product today is monitoring, not conversation.
What the science supports
The MIT Media Lab's Elowan project is a cybernetic demonstration in which a plant's own internal electrical signals interface with a robotic base. It shows that signals can be read and used to trigger machine behavior in a controlled setup. That is far from showing that a model can reliably infer water stress, nutrient deficiency or disease from a stem voltage in a field, across 125 species and varying weather, soil and sensor placement.
There is also a presentation-level description of the company's pitch: an event page from the University of Michigan's Taubman College in 2025 describes The Electric Plant Co. as developing low-cost bioware to capture electrochemical signals from plants and a "Large Plant Language Model" to draw hyper-local insights (event listing). That confirms the company and concept existed before this week, not that the results hold.
Why "language model" is a stretch, and why it might still work
Calling it a large language model borrows a label. The data are time series of microvolt changes, not text. What the company probably means is a large model trained on many sensor streams with labels from other instruments. That can work: foundation models trained on sensor data can generalize across contexts when the data are diverse. The risks are familiar from other sensing applications:
- Confounders. Electrical readings change with temperature, humidity, light, sensor contact, and plant movement. A model might learn those proxies rather than the plant's internal state.
- Label quality. If labels come from coarse tools, as Mast concedes they are, the model may at best match those tools, not surpass them. He says "we know the model will surpass a lot of that equipment," which is a prediction, not a result.
- Generalization. A model that works on tomatoes in a greenhouse may fail on redwoods outdoors. Evaluating across species and sites is the key test.
- Causality. Showing correlation between signal patterns and watering events is not the same as showing that acting on the output improves yield.
Questions that would settle it
- Is there a preprint or paper describing the data, labels and evaluation, with held-out species and sites?
- What was the control in the water-savings claim? A 20% reduction relative to what baseline, over what season and crop?
- How do results compare with simple, cheap baselines, such as soil moisture sensors plus weather data?
- What is the sensor's effect on the plant? The screws penetrate the cambium; long-term damage or infection risk matters, especially for trees.
- Who are the customers, and will any publish measured outcomes?
- Who owns the data, and how is it licensed?
What a convincing validation would look like
If we were designing the test for Bombadil, we would ask for a study with four parts. First, a pre-registered protocol that names the outcome (for example, irrigation volume per kilogram of harvested yield) before the season starts. Second, side-by-side plots on the same farm, one managed with Bombadil's recommendations and one managed with the grower's usual method or a simple moisture-sensor rule, with random assignment of plots. Third, repeats across at least two seasons and several crops, so one lucky greenhouse cannot carry the result. Fourth, an independent party to collect the yield and water data.
For the model itself, the standard is held-out evaluation: train on some species and sites, then test on others the model has never seen, and report where it fails. A table of per-species accuracy would be more informative than a headline count of species. "Interprets electrical signals in about 125 species" could mean anything from reliably classifying water stress to merely detecting that a plant is alive, and only the evaluation tells you which.
Why this kind of claim keeps appearing
Foundation-model methods are being pointed at every data type that has never had one: genomes, proteins, animal calls, sensor streams. Some succeed, because scale and diverse data do surface patterns people missed. Others fail quietly because the signal was never there, or because the model keyed on a confounder. The difference is rarely visible in a launch interview. It shows up in out-of-sample tests, in negative results that authors report, and in whether independent groups can reproduce the effect. That is why we classify this story as claimed versus verified: the premise is plausible, the business is real, and the evidence is still to come.
For growers, the practical cost-benefit is simple. A cheap sensor that gives early warning of stress would be valuable even if it only matched a soil probe, because density is the point: a couple of sensors per acre versus a handful of expensive instruments. But "matches existing tools at lower cost" is a much smaller claim than "interprets what plants are saying," and buyers should price it accordingly.
How to read the funding
The $11.2 million seed is real news in a startup sense, but the funding does not validate the science. For readers following AI for science, the better frame is the trajectory we see elsewhere: companies such as Radical Numerics with Omnii-1 and labs inside frontier AI companies are applying foundation-model methods to new biological data types, and each needs independent evaluation. See also how Google uses AI for farming-related mapping in Earth AI hedgerow mapping.
What growers and builders should do
If you grow crops or manage forests, treat this as a watch item, not a purchase decision. Ask for a pilot with your own baseline, keep your existing moisture monitoring running in parallel, and measure both plant outcomes and total inputs. If you build sensing products, note the instructive design choice: using existing instruments as training labels is a practical way to bootstrap a dataset, but it caps what you can claim until you have ground truth.
This post reflects TechCrunch's October 20, 2026 report and public pages. We have not seen Bombadil, its data or its trials.