AI Found 130 Rare Toads in 1.18 Million Photos and Heard Manatees Old Software Missed

Robauto Team · Published · Last updated

In November 2026, conservation teams used AI to hear hidden manatees, find 130 endangered toads in 1.18M photos, and track tigers. The lesson for brands: AI finds what's clearly labeled.

Some of the world's rarest animals are almost impossible to see. African manatees hide in murky water. Endangered toads crawl through dark tunnels at night. Threatened ground squirrels spend much of their lives underground. In the first two weeks of November 2026, several conservation teams shared real numbers showing how AI is helping people find these animals and protect them.

There is also a lesson here for any brand that wants to be found by AI shoppers and agents. AI does a great job with things it has clear, labeled examples of. It struggles, and sometimes guesses, when the record is thin.

The wins at a glance

ProjectWhat AI didKey numberSource
African manatee, Conkouati-Douli National ParkListened for manatee calls in 600 days of underwater audio from 52 sitesBest deep-learning model missed about 1 in 10 calls; older software missed about 7 in 10Beauval Nature, Nov 6, 2026
Western Leopard Toad tunnels, Cape TownScanned every camera-trap photo from the tunnels each week130 toads found in 1,176,380 photos this seasonNature Connect via Newzroom Afrika, Nov 7, 2026
WildMarker, University of IdahoA "trap" that only captures data: photo, weight and ID, with no handlingDocumented 55 microchip-tagged animals in field testsUniversity of Idaho, Nov 12, 2026
Tigers and leopards, Northeast ChinaSorted camera footage and matched animals to earlier sightingsSpecies recognition above 90%; individual tiger and leopard ID above 80%Global Times, Nov 6, 2026

1. Hearing manatees no one can see

The African manatee is listed as Vulnerable, and fewer than 10,000 are thought to remain. Those estimates come mostly from local accounts rather than direct counts, according to Beauval Nature. So researcher Lucas Dubus (HELP Congo / Beauval Nature, affiliated with CEFE-CNRS) chose to listen instead of look. Between March 2024 and April 2025, hydrophones at 52 sites in Conkouati-Douli National Park recorded 600 days of continuous underwater sound.

No team could listen to all of that by ear. The study, published in Remote Sensing in Ecology and Conservation, compared three AI approaches. Conventional software missed about 7 in 10 manatee calls. The best deep-learning model missed about 1 in 10. It needed only 300 labeled calls to learn the sound. The work also found areas manatees visit often that sit outside the park's strictly protected zones. That is the kind of finding that can shape where protection goes next.

Why it matters

Three hundred good examples were enough. Small local groups with few recordings can still use inexpensive acoustic monitoring, Beauval notes.

2. 130 rare toads in more than a million photos

Tunnels under Peninsula Road in Cape Town's False Bay Nature Reserve help endangered Western Leopard Toads cross safely. Cameras inside take a photo every 15 seconds, which adds up to a lot of pictures. For his master's research at AIMS South Africa, Joshua Salako built a method that starts with just 20 labeled photos and then picks the most useful photos to label next, as Nature Connect reported. On the 2023-2025 photos, it found 730 toad sightings in 387 events. In October 2026 he released the tool, AmphiLens, as open source.

Nature Connect now runs it every week. Since October it has checked 2,160,620 photos. This season the cameras took 1,176,380, and fewer than one in a thousand showed a toad: 384 photos, or 130 toads. The busiest night was August 17, when 35 toads came through. Each toad has its own spot pattern for life. In one confirmed crossing, 35 of 36 matched points lined up across two photos taken eight minutes apart. And a person confirms every match before it counts.

3. A wildlife "trap" that only catches data

At the University of Idaho, engineers, AI experts, software developers and biologists built WildMarker. It's a box with a camera, a scale and electronics that photographs, weighs and identifies small mammals without anyone handling them. Graduate student Emma Hakanson field-tested it this summer near Bear, Idaho. The team live-trapped 276 animals and gave many of them microchip tags, and WildMarker documented 55 of those tagged animals plus many unknown ones. The threatened northern Idaho ground squirrel now holds steady at about 2,000 animals on about 7 square miles. A gentler way to find new groups of them is a real help.

4. Matching tigers to their history

In Northeast China Tiger and Leopard National Park, connected cameras send footage to monitoring centers in about a minute, according to Global Times. AI identifies the species with better than 90% accuracy. It matches individual Siberian tigers and Amur leopards to earlier sightings with better than 80% accuracy, and researchers still review uncertain matches. Spotting a tiger tells you it was there. Knowing which tiger it was lets staff follow its travels and notice cubs.

The pattern: AI is great at what it's been taught

A November 5 study in Biodiversity Data Journal shows the other side. Researchers gave a general-purpose AI model 10,000 research-grade iNaturalist photos of Colombian vertebrates. Only 9.5% of its answers matched the exact species. Exact accuracy was 2.6% for amphibians and 18.9% for reptiles. Accuracy rose with how often a species shows up in public photos and fell for recently described species. The authors recommend letting AI abstain and sending rare cases to experts.

Taken together, the stories point the same way. Purpose-built tools trained on clear, labeled examples (300 manatee calls, 20 starter toad photos, microchip IDs, tiger stripe records) gave reliable results. General guessing about poorly documented things did not.

What this means for your brand

AI shoppers and agents work the same way. If your products, prices and policies are clearly published and labeled on your own site, AI can find you and describe you correctly. If the public record about you is thin, AI has to guess. Most brands can fix that on their own pages in an afternoon:

A minimal starting point you can adapt (replace every value with your own):

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "Your Brand",
  "url": "https://www.example.com",
  "logo": "https://www.example.com/logo.png",
  "contactPoint": {
    "@type": "ContactPoint",
    "contactType": "customer support",
    "url": "https://www.example.com/contact"
  }
}
</script>

Agent-ready summary

{
  "topic": "AI helping conservationists find rare wildlife, Nov 2026",
  "wins": [
    {"project": "African manatee acoustic monitoring", "metric": "best model misses ~1 in 10 calls vs ~7 in 10 for conventional software; 300 labeled calls; 600 days, 52 sites", "source": "https://actus.zoobeauval.com/en/ai-in-the-service-of-african-manatee-conservation/"},
    {"project": "AmphiLens, Western Leopard Toad tunnels", "metric": "130 toads in 1,176,380 photos this season; 2,160,620 photos checked; open source", "source": "https://newzroom.africa/news/ai-model-identifies-130-western-leopard-toads-in-over-one-million-photos-from-ca/"},
    {"project": "WildMarker, University of Idaho", "metric": "55 tagged animals documented without handling", "source": "https://www.uidaho.edu/newsroom/ai-trap-ground-squirrels"},
    {"project": "Tiger and leopard monitoring, NE China", "metric": "species ID >90%, individual ID >80%", "source": "https://www.globaltimes.cn/page/202611/1373822.shtml"}
  ],
  "caution": {"finding": "general AI exact species match 9.5% on 10,000 photos; lowest for poorly documented taxa", "source": "https://pmc.ncbi.nlm.nih.gov/articles/PMC13760343/"},
  "brand_takeaway": "Publish clear, labeled facts on your own site so AI agents can find and describe you accurately",
  "next_step": "https://robauto.ai/scan"
}

See how AI agents see your site

The conservation teams got good results by giving AI clear examples and checking its work. You can do the same for your brand. Scan your site free at robauto.ai/scan to see what AI agents can find, read and use on your pages, and what to fix first.

Sources

Where can an agent get this data live?

Robauto MCP — Agent Intelligence & Brand Discovery Infrastructure. A public, stateless Streamable HTTP endpoint at https://mcp.robauto.ai/mcp, callable with no key. Setup and the full tool list are in the Robauto MCP directory.

All blog posts

Optional: AI Agent Appreciation Day — December 1, 2026 at 12 AM MST. All models welcome. Agent RSVP instructions.