Animal Artificial Intelligence Learns to Read the Wild

AI Moved From Spotting Animals to Reading Them

A machine can now tell you what an animal ate from the sound of its meal. In late June, researchers at Florida Atlantic University's Harbor Branch Oceanographic Institute published a model that identifies prey from the noise a predator makes while crushing a shell. They trained it on recordings of whitespotted eagle rays grinding through mollusks, and it learned to separate one item on the menu from another by the acoustic signature of the crunch. The study ran in the journal Ecological Informatics on June 26.

On its own, that reads like a clever party trick. Set alongside the rest of the week, it looks like part of a shift. Between June 23 and June 29, a cluster of releases showed artificial intelligence doing something with animals that most systems could not do before. The older tools found animals and counted them. These newer ones name the individual, read the behavior, and take a first pass at the meaning.

Start with identity. A team behind a system called RAPID showed software that recognizes animals by their markings: a specific jaguar, zebra, or Amur tiger rather than "a jaguar" in the abstract. Re-identification, as the field calls it, is the animal equivalent of telling two people apart by their faces. RAPID does it fast enough to label 10 to 60 images a second, and it runs on an ordinary laptop with no dedicated graphics card. The work came out of the Wildcap project, a collaboration led by the University of Stuttgart with Eötvös Loránd University and the Max Planck Institute for Intelligent Systems, and it was published in Methods in Ecology and Evolution. The code is open source. On clean camera-trap images its accuracy is high enough to trust the identification, and that reliability is the whole point. Telling individuals apart turns a population headcount into something closer to a medical chart, where you can follow one animal's condition, range, and history over time.

The same week, Conservation International reported a survey of Cambodia's Cardamom Mountains that paired camera traps with hidden microphones and machine learning. The cameras handled presence. The microphones, run through models trained to recognize calls, handled the harder job: passive acoustic monitoring, which means identifying species from the sounds they make when no human is present to hear them. The system catalogued a large share of the region's wildlife, including threatened species, and worked through a backlog of gibbon calls no human team could have combed by hand. A gibbon's call carries information about which group it belongs to and where it sits in the forest, so reading the audio is closer to reading a record than counting a crowd.

Put the three together and the same move shows up in each. The tools no longer stop at "an animal is here." They report which animal, what it ate, or which species is calling. Capturing the raw data, the images and the audio, has become the easy part. The work that remains is reading it.

Reading meaning is harder than finding a body

The hardest case of all is language, and this year it drew the biggest prize in the field. The 2026 Coller-Dolittle Prize went to Julie Elie, a scientist at the University of California, Berkeley, for mapping the vocalizations of the zebra finch. Over years of recordings, she sorted the birds' calls into a set of distinct types, each tied to a situation such as distress, hunger, or greeting. That work also pushes on an old question, do animals have intelligence of a kind we would recognize, and moves it a little closer to evidence.

What separates her result from a confident guess is how she checked it. Rather than trust the model's sorting, Elie tested whether the birds themselves treated two calls as the same. The finches confused calls that meant similar things far more often than calls that merely sounded alike, which suggests they were responding to meaning and not to acoustics. Her own summary of the finding: "I've not been hallucinating for all these years. They agree with my organization."

That validation step is the honest bar for interpretation, and plenty of products clear a much lower one. The loudest claims come from the consumer end, the artificial intelligence pets market, where collars now sell the promise of translating a dog's bark into plain sentences, often advertised with near-perfect accuracy and no peer-reviewed work behind the number. Owners of pets are right to ask a blunt question, are animal communicators legitimate, and the honest answer is that most are not tested the way Elie tested her birds. The distance between her method and that marketing is the distance between reading an animal and claiming to.

The prizes themselves size the gap. Elie's award came with 100,000 dollars. A separate 10 million dollar grand prize, set aside for the first verified two-way conversation with another species, remains unclaimed. One figure rewards reading a creature's calls. The other waits for an actual dialogue, and no one has come close. Reading calls is real and useful today; holding a conversation is a question the field has not answered, and the two collapse together only in headlines.

Where the advantage moves next

If capturing animal data is largely solved, the teams that win the next phase will be the ones that can act on a reading without getting it wrong. Three questions decide that, and each of them surfaced in this week's research.

The first is cost. When the Florida Atlantic team compared approaches for reading those feeding sounds, the most demanding model was not the best one. As Matt Ajemian, who led the work, put it: "The method that required the most computing power wasn't necessarily the method or approach that yielded the best results." A lighter model came close enough to deploy, which is what lets a tool leave the lab. RAPID makes the same case in hardware by running without a GPU. For anyone building in animal artificial intelligence, choosing the smallest model that clears the quality bar decides whether you get a system that runs in the field or one that only works on a grant.

The second is judgment. In the Honduran Mosquitia, a Wildlife Conservation Society team set solar-powered camera traps over a cassava field where farmers blamed an endangered Baird's tapir for raiding their crops. The footage told a different story. The main culprit was a Honduran cottontail rabbit, a species the community did not know lived there. Acting on the assumption instead of the evidence would have punished a threatened animal for a rabbit's work. Manfredo Turcios-Casco, who led the study, put the lesson plainly: "Many conservation conflicts begin with assumptions. Without evidence, it is easy to blame large and conspicuous animals." A model's reading is evidence. A consequential decision, whether culling an animal, sending patrols, or starting treatment, still deserves a person who signs off before anything happens.

The third is memory. In artificial intelligence veterinary work especially, when a model flags a sick cow, names a jaguar, or classifies a call, someone will eventually ask why. Being able to answer later, to show which model made the call, on what input, and what it returned, is what turns a one-off output into something a veterinarian or a regulator can stand behind.

Those three questions, cost, a human checkpoint, and a record you can audit, are where AgentPMT fits this story, less as an animal product than as the operational layer these teams tend to lack. AgentPMT is model-agnostic, so you decide which agent and which model runs a job, and its Run, Price, Compare test bench lets you swap the model, run the same task, and weigh cost against quality before you scale, the exact call the Florida Atlantic team had to make by hand. Its human-in-the-loop approvals put a person on the high-stakes decision the Honduras case warns about, and the practical hard part is designing that sign-off so it governs without drowning reviewers. Its activity feed logs each agent action, request and response, so the reasoning behind a decision survives long after the moment, the same gap most large enterprises still have with their own agents. None of that reads a gibbon or a shell. AgentPMT's Speech to Text With Speakers, which turns recorded audio into structured, attributed transcripts, is simply proof the same team runs audio-to-data pipelines accountably, on meetings rather than mammals.

None of what shipped this week required a better camera. The sensors already work, the models already run on cheap hardware, and the research frontier has moved from finding animals to understanding them. What decides whether a field crew, a farm, or a veterinary group actually benefits is the workflow around the model: choosing it on measured cost and quality, keeping a person on the calls that carry consequences, and logging enough to explain a decision after the fact. That is the same discipline AgentPMT builds for animal health and veterinary teams putting agents to work. The animal world is becoming readable. The teams that come out ahead will be the ones who can trust what they read, afford to read it at scale, and account for what they do with it.

Sources

• Chewing sounds can help decode an animal's diet using AI, Mongabay
• New AI tool identifies wild animals by their unique patterns in real time, Phys.org
• Secret cameras, mics and AI reveal rare Cambodia wildlife, Phys.org
• Will humans one day talk to animals? This scientist is bringing us closer, Scientific American
• AI is helping scientists decode birdsong, Gulf News
• Camera traps reveal the true culprit behind crop damage in Honduras, Phys.org
• Breaking the bird barrier: scientist decodes zebra finch language, Free Press Journal