Animal Artificial Intelligence: The Week's Field Report

Beyond the shift from detecting animals to interpreting them, late June brought a landscape-scale AI wildlife survey in Cambodia, a widening race to decode animal communication across species, and a Cornell summit setting ground rules for AI in veterinary medicine.

Our feature this week follows one throughline: animal artificial intelligence is moving from spotting animals to reading them. The rest of the field did not sit still. Across late June, a landscape-scale survey put that reading power to work protecting a Cambodian rainforest, a widening group of labs pushed on the harder problem of decoding what animals actually say, and forty researchers sat down at Cornell to argue over the rules such tools should follow before a veterinarian trusts one.

A Rainforest Census, Run by Cameras and Microphones

Conservation International finished the largest census yet of Cambodia's Cardamom Mountains, a range that spans more than a million hectares, roughly 2.47 million acres, of southwest Cambodia. The method was almost entirely automated sensing: close to 150 camera traps placed at regular intervals for a 2024 survey, with a repeat planned for later in 2026, plus bioacoustic monitors at ten sites spaced at least three kilometers apart.

The haul was substantial. More than 100 resident species turned up in the Central Cardamom region, nearly two dozen of them vulnerable or endangered, including pileated gibbons, pangolins, elephants, pig-tailed macaques, and endangered wild dogs known as dholes. The gibbons carried the acoustic side of the study. Monitors recorded close to 800 calls in six weeks, and the team spent three months preparing the machine-learning model, labeling up to half the audio by hand so the system could tell a gibbon call from everything else in a noisy forest.

The people running it framed the data as leverage, not trivia. "Gibbons are indicators that our forest is still alive," said Ratha Sor, biodiversity and science manager at Conservation International, who called the survey "the real evidence" that the range still shelters rare and threatened species. Pan Sok, a 50-year-old member of the Chong Indigenous minority who helped with the fieldwork, put it more plainly after reviewing the footage: "My efforts paid off."

That evidence matters because the pressure is real. Infrastructure projects, including several dams, remain live deforestation threats, and the central protected region lost nearly 7,000 hectares of tree cover over five years. Poaching has cooled, though a ranger still turned up part of an old snare mid-survey. At this scale, the sensing is the easy part; the work is turning a river of images and audio into decisions a conservation team can stand behind, cheaply enough to run every season and with a record of how each call was made.

Source: Phys.org (AFP)

The Race to Decode Animal Speech Widens

Our feature covers this year's Coller-Dolittle Prize winner, Julie Elie, and her decade-plus of zebra finch work; the field around her is where the rest of the news sits. The most ambitious entry is the Earth Species Project, whose co-founder Aza Raskin laid out the approach at SXSW earlier this year: build large AI models trained on animal sounds instead of human language, then hunt for recurring patterns and hidden structure across species rather than translating word for word. "AI gives scientists an opportunity to analyse nature at a scale that simply wasn't possible before," Raskin said.

The supporting evidence is piling up across the animal kingdom. A Swiss-US team found bonobos combining calls into sequences that resemble simple linguistic rules. French researchers, working with colleagues in Côte d'Ivoire, have been parsing the hoos and yelps of chimpanzees. Another French group showed African striped mice identify one another through ultrasonic squeaks. Each is a separate answer to the same old question people ask about their pets and about wild animals alike: do animals have intelligence rich enough to carry real information, and can a model recover it?

The money and the skepticism are both worth watching. The Coller-Dolittle Prize, established in 2024 by the Jeremy Coller Foundation with Tel Aviv University, hands out $100,000 a year and holds a $10 million grand prize that no one has claimed, reserved for genuine two-way communication with another species. Coller, the British financier behind it, is bullish: "I'm convinced this is now inevitable. It's inevitable because AI is accelerating so fast." His own judges are more careful. Panel chair Yossi Yovel of Tel Aviv University has been openly cautious about the timeline, and fellow judge Jonathan Birch of the London School of Economics praised the winning work as "phenomenal" while stopping well short of declaring the code cracked. For anyone building a product that claims to read an animal's mind, that gap is the useful signal. The question buyers keep asking, whether animal communicators are legitimate, gets answered by validation and a paper trail, not by confidence, and the researchers doing the real work say so themselves.

Source: Gulf News, Free Press Journal

Cornell Tries to Write the Rulebook for Veterinary AI

While the wildlife labs pushed on capability, a smaller group spent three days on the guardrails. Cornell University's College of Veterinary Medicine convened a summit, "Building Benchmarks for AI-Driven Veterinary Innovation," from June 9 to 11, drawing 40 leaders from 16 institutions across six countries. The premise was that veterinary AI has raced ahead of any shared way to judge it.

The agenda read like a to-do list for making the technology trustworthy: benchmark datasets, data standardization, privacy, bias, and whether a model trained in one clinic generalizes to another. Governance of the data repositories themselves came up, as did a One Health framing that ties animal, human, and environmental health together, and the environmental cost of running the models. "Without them, we can't tell if an AI tool is reliable, fair, or even useful," said Renata Ivanek, a professor in the Department of Population Medicine and Diagnostic Sciences, on the missing benchmarks. Parminder Basran, an associate professor of medical oncology, framed the value as alignment: "The summit created space for leaders to align on problems." Jennifer Sun, an assistant professor of computer science at Cornell's Ann S. Bowers College, brought the machine-learning side to the table.

The meeting produced more than talk. Attendees drafted a 22-page white paper targeted for publication in September 2026 and stood up four working groups covering governance, sustainability, minimum viable products, and the shared data repository, each with a near-term and long-term roadmap. That is the unglamorous layer that decides whether artificial intelligence in veterinary medicine ends up dependable or just impressive in a demo, and it is the same problem every operator of animal AI runs into once the sensors work: choosing a model on measured cost and quality, keeping a person on the consequential calls, and being able to show later why the system decided what it did. It is why AgentPMT treats that operational layer, model comparison, human approvals, and a full audit trail, as the actual product for teams building across the animals and veterinary space.

Source: Cornell University College of Veterinary Medicine

Sources

Secret cameras, mics and AI reveal rare Cambodia wildlife, Phys.org (AFP)
AI is helping scientists decode birdsong, Gulf News
Breaking the bird barrier: scientist decodes zebra finch language, Free Press Journal
Cornell summit sets bar for responsible data science and AI in veterinary medicine, Cornell University College of Veterinary Medicine