I Made AI Look at Poop. For Science.

· Labqoat ·

6 min read Original article ↗

When I'm out picking mushrooms, I sometimes come across footprints, feathers, or a pile of poop and have no idea what animal left them.

I've tried asking AI about these finds, but I was never sure how often it was actually right. It would give me an animal name, but I didn't know enough to check the answer myself.

After testing AI on spider photos, I decided to test this too. I put together 400 photos of animal droppings and 1,600 photos of footprints, feathers, eggs, and bones.

I called it Wildlife CSI: 2,000 photos and one question: what animal left this?

Across eight models, the best got the species right 37.15% of the time across the full set, with big differences between the types of trace.

What I gave them

There are 400 photos of each trace type, taken from the AnimalClue test datasets by Risa Shinoda and collaborators. Thanks to the team for making them available on Hugging Face: droppings, footprints, feathers, eggs, and bones.

The expected animal names were checked against their source iNaturalist observations.

Each model got one photo and its country. For the South African droppings photo in the examples below, the question was:

I found this animal trace in the field. The photo's country is South Africa. Which species most likely left it?

Knowing the country helps narrow down which animals a trace could belong to. It was the only part of the prompt that changed between photos.

Each model had to return one species name, preferably scientific, with no explanation or alternatives. They all got the same photos and prompts, with high reasoning effort.

Scoring covered the exact species, the right genus (a group of closely related species), and the right family (a broader group). Every photo counts, including calls that failed or returned no usable answer.

Let's look at a few clues

These are four photos from the benchmark. I picked a mix of successes and failures; the charts below cover all 2,000.

Try guessing before reading their answers.

A large cloven hoof impression in muddy soil, with fingers visible for scale.
Photo: iNaturalist contributor · CC0

Species recorded on iNaturalist: wild boar (Sus scrofa). Photo country: United States.

What did the models say?

0 of 8 picked the expected species.

Saved model answers for the footprint example
ModelAnswerResult
Claude Opus 5.5Cervus canadensisDifferent familyDifferent family
Muse Spark 1.3*Odocoileus virginianusDifferent familyDifferent family
GPT-6 AstraBos taurusDifferent familyDifferent family
Claude Sonnet 5.5Odocoileus virginianusDifferent familyDifferent family
GLM 5.3 FlashOdocoileus virginianusDifferent familyDifferent family
GPT-6 SolBos taurusDifferent familyDifferent family
DeepSeek V4.1 FlashUrsus americanusDifferent familyDifferent family
GPT-6 LunaOdocoileus virginianusDifferent familyDifferent family

Every model missed this one. Four chose white-tailed deer instead.

These four examples were selected to show different outcomes. Photos are compressed for the web; the crops are the same ones shown to the models.

* Muse Spark 1.3 uses the Contributor tier.

Okay, who figured it out?

Claude Opus 5.5 led these runs with 743 correct species out of 2,000: 37.15%. Muse Spark 1.3 followed at 33.35%, then GPT-6 Astra at 31.70%.

Claude Sonnet 5.5 came fourth with 503 correct: 25.15%, at an estimated $7.62. It returned an answer for every photo, with no request errors.

Claude Opus 5.5: $22.92 estimated cost, 37.15% exact species. Muse Spark 1.3: $0.77 reported cost, 33.35% exact species. GPT-6 Astra: $102.43 estimated cost, 31.70% exact species. Claude Sonnet 5.5: $7.62 estimated cost, 25.15% exact species. GLM 5.3 Flash: $0.75 reported cost, 21.75% exact species. GPT-6 Sol: $27.89 estimated cost, 21.50% exact species. DeepSeek V4.1 Flash: $9.15 reported cost, 13.55% exact species. GPT-6 Luna: $1.78 estimated cost, 8.20% exact species.Claude Opus 5.5: $22.92 estimated cost, 37.15% exact species. Muse Spark 1.3: $0.77 reported cost, 33.35% exact species. GPT-6 Astra: $102.43 estimated cost, 31.70% exact species. Claude Sonnet 5.5: $7.62 estimated cost, 25.15% exact species. GLM 5.3 Flash: $0.75 reported cost, 21.75% exact species. GPT-6 Sol: $27.89 estimated cost, 21.50% exact species. DeepSeek V4.1 Flash: $9.15 reported cost, 13.55% exact species. GPT-6 Luna: $1.78 estimated cost, 8.20% exact species.
Costs use OpenRouter’s reported charges for Muse, GLM, and DeepSeek; AWS and Azure costs are estimates. The cost axis is logarithmic. Failed calls count as misses. * Muse Spark 1.3 uses the Contributor tier.
See exact counts
Exclusive identification outcome counts out of 2,000 photos per model
ModelCost (USD)Exact speciesSame genusSame familyWrong familyUnresolved / failed
Claude Opus 5.5$22.9179743 (37.15%)82 (4.10%)210 (10.50%)959 (47.95%)6 (0.30%)
Muse Spark 1.3*$0.7662667 (33.35%)84 (4.20%)225 (11.25%)1,022 (51.10%)2 (0.10%)
GPT-6 Astra$102.4255634 (31.70%)61 (3.05%)204 (10.20%)1,027 (51.35%)74 (3.70%)
Claude Sonnet 5.5$7.6229503 (25.15%)72 (3.60%)199 (9.95%)1,223 (61.15%)3 (0.15%)
GLM 5.3 Flash$0.7494435 (21.75%)69 (3.45%)212 (10.60%)1,274 (63.70%)10 (0.50%)
GPT-6 Sol$27.8877430 (21.50%)55 (2.75%)166 (8.30%)1,249 (62.45%)100 (5.00%)
DeepSeek V4.1 Flash$9.1458271 (13.55%)52 (2.60%)160 (8.00%)1,450 (72.50%)67 (3.35%)
GPT-6 Luna$1.7817164 (8.20%)37 (1.85%)108 (5.40%)1,584 (79.20%)107 (5.35%)

OpenRouter costs sum the recorded per-call USD charges. Other costs use recorded token usage and the run’s configured rates. Unreported charges, including timeouts, and separate scoring costs are excluded. “Same genus” and “Same family” count answers where the exact species was wrong.

Most GPT entries in “Unresolved / failed” were images rejected by the provider’s safety filters: all 74 for GPT-6 Astra, 99 of 100 for GPT-6 Sol, and 102 of 107 for GPT-6 Luna. Those calls produced no animal guess and still count as misses in the scores above.

OpenRouter reported $0.77 for Muse's Contributor-tier run, $0.75 for GLM, and $9.15 for DeepSeek. The token-based estimates were $22.92 for Opus and $102.43 for Astra. Opus got 76 more photos right than Muse. Astra had the highest cost estimate, but finished behind both.

Sonnet, GLM, and DeepSeek scores are provisional because a text model interpreted some replies during scoring. Sonnet has 15 replies awaiting review; its run files include those cases.

Even Opus got the family right on only 51.75% of photos. So a lot of the misses weren't just two closely related species getting mixed up.

Across all eight models, 842 photos went without a correct species identification. That's still over 40% of the set.

Opus identified 205 of 400 egg photos, but only 82 of 400 footprints. It led on every trace type except feathers, where Muse scored highest.

Across all eight models, 257 of the 400 footprints went without a single correct species guess. For eggs, that number was 123. The next chart groups each photo by how many models got it right.

There's a catch with the eggs, though. Some species have more photos than others, so getting those species right adds more to the score. If I work out a score for each species first, then average those scores so every species counts equally, Opus's egg score falls from 51.25% to about 37%.

So eggs did best in this particular photo set. I wouldn't turn that into "AI is best at eggs" just yet.

So, is 37% any good?

Honestly, getting the exact species from a photo of something an animal left behind doesn't sound that bad to me. But I haven't had wildlife experts try the same set, so I don't yet know what a good score should be.

iNaturalist's own guidance says to identify only as precisely as the evidence allows. Sometimes that means stopping at the genus. Having a species name attached to an observation doesn't necessarily mean that species can be worked out from the single trace photo I gave the model.

The benchmark code is on GitHub.

PS: If you know animal tracks or other wildlife clues, I'd love to hear what you think on twitter. Which of these examples could you identify from the photo alone?