Skip to main content
Launching October 15 · Apply for early access
Wild Spaces

Practical conservation. Useful tools. Clear accountability.

EchidnaCare is software made in the USA, built to help animal-care teams spend more time on the work that matters. That responsibility includes how we use AI: tracking its use, estimating its environmental impact and keeping people in charge of decisions.

Through Wild Spaces, we also offer free software to projects protecting wildlife and habitats. Here you can see how we account for AI use, the safeguards around it, and the research behind our conservation work.

Recorded usage. Estimated impact. Practical conservation support.

We record AI token usage and use it to estimate electricity and carbon impact. We explain the assumptions because providers do not give us a direct energy reading for each request. Our conservation support is a separate contribution, with its own outcomes to report.

Practical conservation, built around the place and its people.

Wild Spaces takes a nonpartisan approach. We support useful work for wildlife and habitats, and assess projects by their needs, animal-welfare practices, local knowledge and evidence of results. Political affiliation is not a selection criterion.

Different places need different approaches. We value cooperation among conservation teams, zoos, veterinarians, landowners, farmers and local communities. We ask what works in that setting, what the tradeoffs are, and how the people responsible will review the results.

Our first contribution will be software, setup and support for twelve projects that can use it. Nominations and votes help bring projects forward; we confirm eligibility, practical fit and the organization’s consent before awarding a place. We agree the donated term and what happens afterward in writing before onboarding. A nomination or vote does not start a subscription.

Feedback helps us improve. Reviews and photographs are optional, and public stories or listings need the project’s permission. Donated software and any conservation outcomes are reported separately from our estimated AI footprint.

Nominate a project
AI, with your team in charge

Useful help. Clear limits. People responsible for the decisions.

A zoo owner should be able to ask what AI does, what it costs the environment, and who checks its work. Here is how we approach each of those responsibilities.

What is the environmental impact?

We track AI use and record the token counts returned by our providers. Tokens are the small pieces of text a model reads and produces. We use online estimation tools and published research to translate usage into approximate electricity and carbon figures. The token counts are recorded usage; the environmental figures are estimates, not readings from a meter in the provider’s data center.

Our product log records the feature, model, input and output tokens, and whether a request succeeded. Cached and reasoning tokens are recorded when the provider supplies them. Missing usage stays marked as missing, rather than being treated as zero. Images are tracked separately. AI used to develop the software is accounted for separately from AI used by your facility.

Is that impact offset?

Wild Spaces gives conservation projects practical help through free software, setup and support. We report that contribution separately from our AI footprint. Donating software does not cancel out emissions, and we do not call EchidnaCare carbon neutral. Any future carbon-offset claim would need its own verified project records and retired credits.

How do we set and check the AI’s limits?

Each AI feature has a defined job, such as summarizing an animal record, preparing a draft SOAP note or helping find relevant regulations. We set its instructions, the information it can access and its response limits in the application. Private record tools use facility access controls; compliance answers draw on relevant records from our regulatory library and show their sources. The public website assistant is instructed to refer veterinary and legal questions to qualified professionals.

AI can miss context or make mistakes. Veterinary, welfare, regulatory and staffing decisions need review by the appropriate person at your facility. A drafted note needs a clinician’s review before it is saved; a suggested answer is a starting point to check against the records and sources.

Our automated checks cover access boundaries, request limits, usage recording and safeguards for incorrect AI extraction results. Those checks help catch software failures; they do not certify every AI answer as accurate or replace professional judgment. We are continuing to develop evaluations using representative animal-care and compliance questions.

How do we keep unnecessary use down?

Product AI runs when someone asks for help. We reuse animal briefings while the underlying record is unchanged and cap public chat requests and response length. EchidnaCare does not train its own AI model. Our aim is to save your team useful time while keeping the work focused and the usage visible.

If your board needs more detail, contact our team about the usage records, estimation assumptions or a particular AI workflow. You can also read our security information.

Explore the estimation method and conservation research
What AI costs

How we turn usage into an estimate.

Token counts tell us how much text the AI processed. Estimating electricity also needs assumptions about the model, hardware, response length and data center. Our current product method groups recorded calls into conventional and longer-response scenarios from the study below, using 1,000 output tokens as our dividing point. That cutoff is our approximation, not a boundary established by the study.

Carbon estimates multiply estimated electricity by an electricity emissions factor. Our current method uses the EPA’s eGRID2022 US-average factor of 373.3 grams of CO₂ per kWh as a proxy; it is not a measurement of our provider’s electricity supply or a full life-cycle CO₂-equivalent figure. Water remains unquantified because we do not know the serving data center or its cooling system. Training, hardware manufacture and users’ devices are outside this inference estimate. Made in the USA describes our software; it does not establish where an AI request is processed.

For background on estimation, see the EcoLogits methodology and the EPA’s electricity conversion references. Different tools may use different assumptions. A report should name its period, models, usage coverage and method so the figures can be checked and updated.

Energy per AI query, and why a single number for all queries is wrong.

Modelled

Energy use of AI inference, efficiency pathways, and test-time scaling

Oviedo F, Kazhamiaka F, Choukse E, Kim A, Luers A, Nakagawa M, Bianchini R, Lavista Ferres JM · Joule, 2026 · doi:10.1016/j.joule.2026.102430

Median 0.31 Wh per query (IQR 0.16–0.60) for frontier-scale models above 200B parameters on H100 nodes, assuming about 300 output tokens. A long reasoning query of about 5,000 output tokens rises 13-fold to a median 3.91 Wh (IQR 2.15–7.05). The authors argue widely cited figures are overstated by 4–20 times.

What this does not show

A bottom-up estimate from token throughput and node power, not metered hardware. It assumes particular hardware and production-scale batching, so it does not describe any specific provider's infrastructure, and it excludes training entirely.

Where it applies: Not location-specific; models a large-scale deployment.

A provider's own measured figures for energy, carbon and water per prompt.

Measured

Measuring the environmental impact of delivering AI at Google scale

Elsworth C, Huang K, Patterson D, Schneider I, Sedivy R, Goodman S, Townsend B, Ranganathan P, Dean J, Vahdat A, Gomes B, Manyika J · arXiv preprint (Google), 2025

A median Gemini Apps text prompt consumed 0.24 Wh of energy, 0.03 gCO₂e and 0.26 mL of water. The same prompt measured by a narrower method that counts only the accelerator comes out at 0.10 Wh, which is why methodology has to be stated alongside any figure.

What this does not show

Specific to Google's hardware, software and clean-energy portfolio, and the authors say directly that it may not generalise to other providers or architectures. Training energy, embodied manufacturing emissions and the user's own device are all excluded. The carbon figure is market-based (Scope 2 MB) and so reflects renewable purchases; it is not the same as a grid-average location-based figure.

Where it applies: Google's own data centres and energy portfolio.

Animals and carbon

What research tells us about wildlife and healthy habitats.

81% of tropical tree species need an animal to carry their seeds somewhere useful. When the large fruit-eaters go, the trees that depend on them thin out, and in many forests those are the big, dense, carbon-heavy ones. The research below is where that story is solid, and, just as importantly, where it is not.

The habitat matters

The relationship varies by habitat. One study modeled carbon losses in African, American and South Asian forests, with little change at its Southeast Asian and Australian sites. Another study in Thailand found losses there too. We keep both in view: wildlife matters in its own right, and a carbon benefit needs evidence from the ecosystem concerned.

What a forest elephant does to a rainforest, and what its loss costs.

Modelled

Carbon stocks in central African forests enhanced by elephant disturbance

Berzaghi F, Longo M, Ciais P, Blake S, Bretagnolle F, Vieira S, Scaranello M, Scarascia-Mugnozza G, Doughty CE · Nature Geoscience, 2019 · doi:10.1038/s41561-019-0395-6

Elephants thin out small stems as they move and feed, which leaves fewer trees competing for light and water and shifts the forest towards larger trees with denser wood. At the usual density of half an elephant to one elephant per square kilometre, that raises aboveground biomass by 26 to 60 tonnes per hectare. Losing forest elephants entirely is estimated to cut aboveground biomass across central African rainforest by 7%.

What this does not show

An Ecosystem Demography model checked against forest inventory data, not a measurement of elephants being removed. The same shift lowers the forest's net primary productivity, because denser wood grows more slowly, so the gain is in stock rather than in growth rate. The paper carries a published author correction (doi:10.1038/s41561-019-0487-3), which is worth reading alongside it.

Where it applies: Central African lowland rainforest.

Tapirs, and the surprising place they do their most useful work.

Measured

Lowland tapirs facilitate seed dispersal in degraded Amazonian forests

Paolucci LN, Pereira RL, Rattis L, Silvério DV, Marques NCS, Macedo MN, Brando PM · Biotropica, 2019 · doi:10.1111/btp.12627

Tapirs dispersed about 9,800 seeds per hectare a year in forest damaged by fire and fragmentation, against about 2,950 per hectare a year in undisturbed forest. That is roughly three times more work in the places that need it most, apparently because tapirs prefer ground where resprouting growth gives them something to eat. The team identified close to 130,000 seeds from 24 species out of tapir dung.

What this does not show

One region of the Brazilian Amazon, so the ratio should not be assumed to hold elsewhere. It measures seeds delivered, which is not the same as seedlings that survive, and the paper does not follow them to established trees. The link to carbon is an inference from those seeds being large-seeded species that become large trees, not something this study measured.

Where it applies: Mato Grosso, Brazilian Amazon.

Rewilding as a climate measure, and the size of the claim being made for it.

Modelled

Trophic rewilding can expand natural climate solutions

Schmitz OJ, Sylvén M, Atwood TB, Bakker ES, Berzaghi F, Brodie JF, Cromsigt JPGM, Davies AB, Leroux SJ, Schepers FJ, Smith FA, Stark S, Svenning J-C, Tilker A, Ylänne H · Nature Climate Change, 2023 · doi:10.1038/s41558-023-01631-6

Restoring nine animal groups — African forest elephants, American bison, fish, grey wolves, musk oxen, sea otters, sharks, whales and wildebeest — is estimated to add 6.41 GtCO₂ of negative emissions a year. The authors set that against a natural climate solutions target of 10 GtCO₂ a year, which it would meet 64% of.

What this does not show

The authors say plainly that some of this is ALREADY counted inside existing natural climate solutions that protect these species' habitats, so it cannot simply be added to them. Press coverage of this paper often reports the figure against the 6.5 GtCO₂ a year implied by the 500 Gt by 2100 target, which makes it look like 95% rather than 64% and drops the double-counting caveat entirely. It is a synthesis of modelled estimates, not a measurement.

Where it applies: Global, across nine species groups on land and at sea.

A single rewilding project, and why we do not put its number in a headline.

Modelled

Rewilded bison in Romania's Țarcu Mountains: modelled additional carbon capture

Schmitz OJ and colleagues, Yale School of the Environment, with the Global Rewilding Alliance · Reported May 2024; analysis not peer reviewed at the time of writing, 2024

A herd of about 170 European bison grazing roughly 20 square miles is estimated to help capture an additional 54,000 tonnes of carbon a year, around ten times what the same ground was drawing down before the bison were returned.

What this does not show

Not peer reviewed. It is the output of a model rather than a field measurement, and press accounts of it disagree with one another on the car-equivalence figure, which is a fair warning about how far the number has travelled from its source. It describes a whole ecosystem's response, so dividing it by 170 to get a figure per bison is not something the work supports, and we do not do it.

Where it applies: Țarcu Mountains, Southern Carpathians, Romania.

Seed-dispersing animals and the rate at which tropical forest regrows.

Measured data, modelled result

Seed dispersal disruption limits tropical forest regrowth

Fricke EC, Cook-Patton SC, Harvey CF, Terrer C · Proceedings of the National Academy of Sciences, 2025 · doi:10.1073/pnas.2500951122

Across 3,026 tropical regrowth plots, areas with the least disruption to seed dispersal accumulated about four times as much carbon as those with the most. For a typical restoration site the loss averaged −1.8 Mg of carbon per hectare per year, a 57% reduction in regrowth potential. 81% of tropical tree species depend on animals to move their seeds.

What this does not show

The four-fold figure describes the RATE at which cleared land regrows, not the carbon a standing forest holds — the two are often confused. The authors note gaps in the resolution of animal biodiversity data, possible mismatches in space and time, and uneven monitoring coverage across the tropics. The analysis treats birds and mammals together and does not separate out primates.

Where it applies: Tropics globally, with the Amazon Basin, Congo Basin and Borneo examined closely.

Why losing large fruit-eating animals affects carbon in some forests and not others.

Modelled

Contrasting effects of defaunation on aboveground carbon storage across the global tropics

Osuri AM, Ratnam J, Varma V, Alvarez-Loayza P, Hurtado Astaiza J, Bradford M, Fletcher C, Ndoundou-Hockemba M, Jansen PA, Kenfack D, Marshall AR, Ramesh BR, Rovero F, Sankaran M · Nature Communications, 2016 · doi:10.1038/ncomms11351

Removing large-seeded animal-dispersed trees entirely cost 2–12% of aboveground carbon in African, American and South Asian forests, and 1–5% under a halved scenario. Southeast Asian and Australian forests showed little change or a marginal gain, within ±1%.

What this does not show

This is the source that stops a blanket claim. The effect depends on which tree species a forest happens to hold, so 'protecting frugivores protects carbon' is true in some regions and not others. The scenarios are simulations, the authors report insufficient data on how defaunated forests actually change over time, and soil and leaf-litter carbon are not covered at all.

Where it applies: Ten sites across Africa, the Americas, South Asia, Southeast Asia and Australia.

The scale of carbon change from losing large fruit-eaters in one measured forest.

Modelled

Defaunation of large-bodied frugivores reduces carbon storage in a tropical forest of Southeast Asia

Chanthorn W, Hartig F, Brockelman WY, Srisang W, Nathalang A, Santon J · Scientific Reports, 2019 · doi:10.1038/s41598-019-46399-y

Simulating the complete loss of large-bodied frugivores reduced aboveground carbon by 2.4–3.0%. Carbon loss passed 1% once defaunation reached 40%.

What this does not show

A simulation built on one 30-hectare plot of 33,844 trees, assuming the community settles to a new equilibrium rather than modelling the transition. Species-level demographic data was missing, rodents may partly compensate by moving seeds, and the authors state considerable uncertainty about how surviving species compete.

Where it applies: Mo Singto plot, Khao Yai National Park, Thailand.

Primates and tapirs specifically, in Neotropical forests.

Review of other studies

To avoid carbon degradation in tropical forests, conserve wildlife

Bennett EL, Robinson JG · PLOS Biology, 2023 · doi:10.1371/journal.pbio.3002262

Summarising modelling work, losing large primates and tapirs is projected to cut aboveground tree biomass by 2.5–5.8% on average in the Neotropics, and by as much as 26.5–37.8% in the worst cases.

What this does not show

A Perspective piece, not primary research: it reports other people's models rather than new measurements. The authors note the picture is complicated by hunting also removing seed PREDATORS, which pushes the other way, and no direct measurement of primate-dispersed carbon is offered.

Where it applies: Tropical forests, drawing on the Brazilian Amazon, Central Africa and Thailand.

In the jungle

How wildlife contributes to forest recovery.

Tropical forest is where most of this evidence was gathered, and where the mechanisms are clearest. An elephant is not storing carbon in any direct sense. It is knocking over small trees, and the forest that grows back around that is heavier.

African forest elephant

Loxodonta cyclotisCongo Basin

Walks through dense understorey knocking over and eating small trees. That thinning leaves fewer stems competing for light and water, and the forest shifts towards fewer, bigger trees with denser wood.

At the usual density of half to one elephant per square kilometre, aboveground biomass is 26 to 60 tonnes per hectare higher. Losing them entirely is estimated to cut central African rainforest biomass by 7%.

What to keep in mind

Modelled, not measured, and the same shift lowers how fast the forest grows, because dense wood is slow wood. The gain is in what the forest holds rather than what it adds each year.

Berzaghi F et al., Nature Geoscience 2019

Lowland tapir

Tapirus terrestrisAmazon

Eats fruit across a wide range and deposits the seeds, with fertiliser, far from the parent tree. It favours damaged ground where fresh regrowth is easy to browse, which is exactly where seeds are scarcest.

About 9,800 seeds per hectare a year in forest damaged by fire and fragmentation, against about 2,950 in undisturbed forest. Roughly three times the work in the places that need it most.

What to keep in mind

Measured in one region of Mato Grosso. It counts seeds delivered, not seedlings that survived, and the step from seeds to carbon is an inference rather than something the study followed.

Paolucci LN et al., Biotropica 2019

Spider and woolly monkeys

Ateles and LagothrixNeotropical rainforest

Swallow large fruit whole in the canopy and carry the seeds along routes no ground animal takes. Large-seeded trees, which tend to be the heavy, long-lived, carbon-dense ones, largely depend on them.

Losing large primates and tapirs from Neotropical forest is projected to cut aboveground tree biomass by 2.5 to 5.8% on average, and by 26.5 to 37.8% in the worst modelled cases.

What to keep in mind

These are projections gathered in a review, not new measurements, and hunting removes seed PREDATORS as well as seed dispersers, which pushes in the opposite direction.

Bennett EL et al., PLOS Biology 2023
Where this goes next

Practical support for the people protecting wild places.

Good records help a conservation team follow an animal’s care, prepare for release and share what happens next. Our first Wild Spaces cohort offers twelve projects free licenses, setup and support, with nominations and voting helping bring those projects forward. As the program grows, we want to help zoos connect with conservation partners and share outcomes with each project’s permission.