Methodology

How the atlas turns counts of animals into Suffering-Adjusted Days, where every input comes from, and what we are least sure about.

Contents

Suffering-Adjusted Days

What a SAD is (one day of disabling pain for one person), where the measure comes from, and which version this site uses.

Read the whole section on Suffering-Adjusted Days

The atlas measures suffering in Suffering-Adjusted Days (SADs), the metric Ambitious Impact (AIM, formerly Charity Entrepreneurship) uses to compare animal-welfare problems and charities. AIM defines SADs as the “amount of days in a certain type of pain, conditional on sentience and welfare range”. The reference point is human: one day of disabling pain experienced by one person counts as one SAD.

SADs rest on the Welfare Footprint Institute’s way of describing experiences. Rather than giving an animal one welfare score, Welfare Footprint estimates how long it spends in four intensities of negative experience: annoying, hurtful, disabling and excruciating. “Pain” covers any felt negative state, so fear, hunger, heat stress and frustrated natural behaviour count as well as injuries. Each welfare problem (keel-bone fractures in hens, lameness in broilers, crowding in shrimp ponds) is described by how long it lasts, how many animals it affects and how likely each intensity is. Problems are added together, and because they overlap an animal can accumulate more than 24 hours of pain in a day.

This site uses AIM’s SADs v.2 (2026) inputs, downloaded on 15 October 2026. There are several incompatible versions of SADs, so figures computed with different versions (AIM’s 2024, 2025 and 2026 models, and Animal Charity Evaluators’ own 2025 adaptation) should not be compared directly.

What SADs tell us

Expected days of disabling-equivalent pain per year. Time in each pain intensity is weighted by how intense it is, by the species’ welfare range and by the chance it is sentient.

What SADs don’t tell us

  • How many animals there are (switch the Measure on the Atlas to Animals alive now or Killed per year).
  • Whether a species is sentient: this is a probability-weighted average, so small chances across huge numbers can dominate.
  • How easy the suffering is to reduce.
  • Good experiences: we found no published estimate of time spent in good experiences for any animal on this site. For most farmed and caught animals with SADs, only time in pain is counted. For cattle, sheep and goats and for wild animals, the figure starts from an overall welfare score, which can weigh good points against bad.

The model

How each group’s SADs add up suffering while alive and while dying, and which pain estimate each group uses.

Read the whole section on The model

For each group of animals the atlas adds suffering during life to suffering while dying:

pain-days per animal-day = Σ weight[intensity] × hours[intensity] ÷ 24 pain-days per death = Σ weight[intensity] × hours while dying[intensity] ÷ 24 SADs per year = ( alive now × 365.25 × pain-days per animal-day + deaths per year × pain-days per death ) × welfare range × P(sentience)

“Pain-days” are disabling-equivalent days: a day of disabling pain counts as one, and the other intensities are converted with the weights below. Multiplying by the welfare range (how intense the animal’s experiences can be, relative to a human’s) and by the probability that it is sentient turns them into SADs, exactly as AIM does. Deaths are animals killed by people (slaughter, capture or culling) for farmed and caught animals, and deaths from all causes, mostly natural, for wild animals.

AIM’s own headline since April 2026 is the four intensities reported separately; it keeps the single weighted total “for information only”. The atlas has to add up 38 groups, so its rankings use the weighted total. Species pages also show the four intensities separately wherever there is a pain-hour model: the whole total for the 13 modelled, borrowed or dying-only groups, and the dying term only for the 24 groups converted from a welfare level that have a count of deaths, whose life term cannot be split. 12 more groups have no SADs at all (below). The other measures on the Atlas are simpler: animals alive now and killed per year are counts, and the welfare-adjusted population is alive × P(sentience) × welfare range.

Where the pain-hours come from

  • Laying hens and broiler chickens: AIM’s v.2 models, built mostly on Welfare Footprint Institute analyses (570 of 600 and 836 of 872 intensity estimates).
  • Pigs and silkworms: AIM’s own v.2 models (AIM staff estimates using the Welfare Footprint method). Pigs blend market pigs and breeding sows.
  • Farmed shrimp: AIM’s v.2 model, built on Rethink Priorities’ 2024 shrimp welfare analysis of the grow-out stage.
  • Farmed fish: AIM’s salmon model without the salmon-only problems, which is how AIM itself estimates other farmed fish; dying adds the Welfare Footprint Institute’s 2025 study of trout dying by air asphyxia.
  • Borrowed values: turkeys and ducks use broiler values, crabs and crayfish use shrimp values, and farmed insects use silkworm values. These are marked “Borrowed pain model” on their species pages.
  • Male chicks: only their deaths are counted, using AIM’s February 2025 quick estimate of one minute of excruciating pain each (AIM’s 2026 models have no chick model).
  • Wild-caught fish and crustaceans: only their capture and killing is counted, with dying profiles built for this site (below); their lives are counted with wild animals.
  • Beef cattle, dairy cows, sheep and goats, and wild animals (all but copepods and the other tiny crustaceans, mites and springtails) are converted from a welfare level (below), because we found no estimate of their hours of pain split by intensity. AIM’s February 2025 quick estimates for beef and dairy cows give only totals, which their species pages use as cross-checks.
  • Managed honeybees, dogs and cats, geese, rabbits, and animals farmed for fur or used in research: we found neither a published pain-hour estimate nor a published welfare level for them, so they have no SADs (below).
  • Wild copepods and the other tiny crustaceans, mites and springtails: we found no published welfare score for them, and by the site author’s choice Charity Entrepreneurship’s ‘wild bug’ score, an insect score, is not borrowed for them, so they have no SADs either (below).

All of AIM’s models describe high-income-country systems, and they treat the time spent in pain as a point estimate.

Pain-intensity weights

How the four pain intensities are weighed against each other, and the five sets of weights you can choose.

Read the whole section on Pain-intensity weights

Because the four intensities are only ranked, they need weights before they can be added. These are the largest judgement call in the metric. AIM’s v.2 weights, derived mostly from human time-trade-off and health-utility studies plus community surveys, count a day of annoying pain as 0.057 disabling days and a day of excruciating pain as about 11. They have moved a lot between versions: AIM’s 2025 model treated excruciating pain as 83 times worse than disabling. You can switch between these sets on the Atlas page:

Switching weights can reorder groups whose suffering is mostly chronic and mild (crowding, barren cages) relative to groups whose suffering is brief and intense (slaughter, boiling). Each species page shows how its total and rank change under every set.

Moral weights

How much one animal’s experience counts next to a human’s: welfare ranges, neuron counts or equal weights, and the chance of feeling anything.

Read the whole section on Moral weights

By default the atlas uses Rethink Priorities’ welfare ranges, conditioned on sentience, and their mean, as AIM does. A range that already includes P(sentience) must never be multiplied by it again; we use the sentience-conditioned values and apply P(sentience) separately. The 2024 book Weighing Animal Welfare covers pigs, chickens, carp, shrimp, black soldier flies and octopuses. Honeybees, silkworms, crabs and crayfish, and the wild groups that borrow silkworms’ range use RP’s 2023 Moral Weight Project values instead, which come from a different mix of models (including a rate-of-experience adjustment the 2024 values lack), so they are not strictly comparable. Where we found no RP estimate for a species or its group, we borrow another animal’s, labelled “borrowed from …” on the species page: cattle, for example, borrow pigs’ welfare range, and most wild groups borrow theirs (below).

You can switch to two alternatives on the Atlas page. The results are then labelled SAD-equivalents, because they are no longer AIM’s metric:

  • Neuron counts: an animal’s estimated neuron count divided by the human brain count (about 86 billion). For wild groups with a weight, the model borrows a published count from a particular animal. Each species profile gives the source and explains which animal and part of its nervous system the count covers. The chart ranges hold this count fixed. A few profiles also offer a separate scenario using a count from one life stage. The physical neuron comparison uses its own catalogue. Neuron weighting is contested: Rethink Priorities (Shriver 2022) argue that neuron counts should not be used on their own as a measure of moral weight.
  • Equal weight: every sentient individual counts the same as a human. This shows the raw scale of numbers.

Probability of sentience can be switched off, in which case every group is treated as certainly sentient (SADs conditional on sentience).

Groups converted from a welfare level

How groups converted from a welfare level are put into SADs, and how rough that conversion is.

Read the whole section on Groups converted from a welfare level

For beef cattle, dairy cows, sheep and goats, and all wild animals, we found no estimate of hours of pain split by intensity, which the model needs to weight each intensity. AIM’s February 2025 quick estimates for beef and dairy cows give only totals of disabling-equivalent days, which their species pages use as cross-checks. For these groups (except wild copepods and the other tiny crustaceans, mites and springtails, which have counts only) the atlas converts the group’s average welfare level, its overall welfare score (from −100%, as bad as a life can be, to +100%), into pain-days with one constant, fitted on the groups where both measures exist:

pain-days per animal-day = k × max(0, −welfare level)

The welfare levels are Charity Entrepreneurship’s 2018 Weighted Animal Welfare Index scores. The index did not score sheep or goats: they use Kyle Fish’s 2023 value of −20% (tier 2), the value he gives cattle and says he based on Charity Entrepreneurship’s index; it equals the index’s score for factory-farmed beef cows. We found no published range for sheep or goats, so they borrow the beef-cow score’s range: the spread between Charity Entrepreneurship’s 2018 raters, from −58% to +7%. The index scores only some kinds of wild animal; most wild groups borrow one of those scores as a stand-in, labelled as borrowed on each group’s page (wild animals, below). The constant k is the geometric mean (an average taken on a multiplying scale) of AIM v.2 pain-days ÷ −(welfare level) over the calibration pairs below. Under the default weights a welfare level of −100% corresponds to about eight hours of disabling pain a day. The constant is recomputed for each set of pain weights.

Only pairs whose welfare level is on Charity Entrepreneurship’s scale are used: four from its 2018 index, farmed shrimp from its 2021 shrimp report, and silkworms from a 2024 report by its research training programme (tier 2). Pigs also have both measures, but their welfare level (−40%) is Kyle Fish’s own 2023 estimate rather than a score on the index; adding them would raise the constant by about 8%.

This keeps every group in one unit, but these estimates are much weaker. The calibration rests on six farmed-animal comparisons, three of them chickens; the welfare index scores distance from an ideal life rather than time in pain; and a welfare level is a net figure, so for groups with mixed lives good experiences partly offset suffering, and a group whose average is above zero gets no SADs at all. On the Atlas these groups are drawn with open markers.

We found no check of this conversion against any wild animal.

Charity Entrepreneurship’s index gives an animal points on each of eight criteria and adds them up, and a criterion can score above zero, so a score can weigh good points against bad. Within the score each group uses, weighing good points against bad changes the result only for cattle, sheep and goats, and for the wild groups on Charity Entrepreneurship’s ‘wild fish’, ‘wild rat’ and ‘wild bird’ scores. Charity Entrepreneurship’s ‘wild bug’ score, used by the groups that make up about 98% of the wild total under the default assumptions, has no good points to weigh: each of the index’s eight criteria scores zero or below.

But the constant itself is fitted on farmed scores, and four of the six calibration scores weigh a few good points against bad (for the other two the site has no breakdown into points). Counting only the bad points of those four, the constant would be about 6% lower, and so would the suffering while alive of every group converted from a welfare level.

The site uses two welfare-level scales. In this conversion, under the default weights, −100% is about eight hours of disabling pain a day. The welfare levels that species pages show as converted by this site from hours of pain use another scale: there, −100% is 24 hours a day of excruciating pain (on farmed fish’s row for dying, excruciating pain for all of the stages of dying it covers). 24 hours a day of excruciating pain is about 32 times as much pain as −100% in this conversion (excruciating pain counts about 11 times as much as disabling pain).

Groups we found no published estimate for

The groups with counts but no SADs, and why: we found no published estimate for them, and the site does not borrow an insect score for the wild ones.

Read the whole section on Groups we found no published estimate for

Six groups of animals kept by people have no SADs: managed honeybees, dogs and cats, farmed geese and rabbits, animals farmed for fur, and animals used in research. We found no published estimate of their welfare level or hours of pain, and without one the model has nothing to turn into SADs. We do not fill the gap with a score of our own, because it could be wrong by a wide margin.

So every SADs total, ranking and map on this site leaves these groups out. Their counts are still shown, and so is their welfare-adjusted population, which uses Rethink Priorities’ published welfare ranges and probabilities of sentience and needs no welfare level or hours of pain. On the Atlas you can enter your own hours of pain for any of them under “Change the assumptions” → “More assumptions”; the result is then your own figure, not this site’s.

Six wild groups have counts only in the same way: copepods and the three other groups of tiny crustaceans (other tiny drifting crustaceans, tiny crustaceans of the sea floor and freshwater crustaceans), mites and springtails. We found no published welfare score for them. Charity Entrepreneurship gives no definition of ‘bug’, and its report describes insects, so its ‘wild bug’ score is an insect score, and the site does not borrow it for them (wild animals, below).

Suffering while dying

How suffering while dying is counted, for slaughter, catch and natural deaths.

Read the whole section on Suffering while dying

For modelled groups, and the farmed groups that borrow a modelled species’ values, AIM’s end-of-life problems (catching or depopulation, transport, lairage, and slaughter until loss of consciousness) are moved out of the life term into the dying term, so nothing is counted twice. Deaths on the farm before slaughter are already part of the life term. Three farmed groups converted from a welfare level borrow an AIM end-of-life profile as well: beef cattle, dairy cows, and sheep and goats use the pig one (transport and stunning).

Wild-caught fish and crustaceans, and wild animals, have dying profiles built for this site: a mix of causes of death, each weighted by its share of deaths and given a duration and intensity split. Wild animals mostly die young, by predation, starvation, exposure or disease; the profiles use mortality studies and the Welfare Footprint Institute’s pain estimates for these causes, and count only the final phase of dying, so long hunger or illness before it stays in the life estimate. The Welfare Footprint Institute applies its profiles, built for chickens, across vertebrates; using them for invertebrates (insects, spiders, worms, krill, shrimps and crabs, squid, sea butterflies and nematodes) is this site’s own extension. Each wild profile was built for one kind of animal, and the other wild groups borrow one of them, labelled as borrowed on their pages. Wild-caught fish and crustaceans are profiled by fishing gear (crowding in purse seines, crushing in trawls, hours in gillnets, then asphyxia, ice or live gutting) and are counted only while being caught and killed, because their lives are counted with wild animals. Most of the shares and durations are our own documented estimates, and the species page of each of these groups lists the causes behind its profile.

Wild animals

How wild animals are counted, which values they borrow from other animals, and why the wild figures are so uncertain.

Read the whole section on Wild animals

Wild animals are counted in finer groups gathered into families: copepods and other tiny crustaceans; mites, springtails, insects, spiders and worms; krill, shrimps, crabs, squid and sea butterflies; fish; and mammals, birds, reptiles and amphibians. Wild nematodes are a family of their own and are left out unless you include them. The Atlas shows the families; each group has its own species page with its counts, its values and their sources. Their suffering is never added to the human-caused total. Four earlier groups were split: wild mammals, wild fish, wild insects and other land arthropods, and wild aquatic crustaceans. A link that set numbers for one of them applies them to each group it became; a chance of feeling at the low or high end of the old group’s published range becomes each group’s own low or high end.

Counting the animals

Each group’s page says whether its count is a published estimate, our own derivation from published figures, or our own rough estimate, with its sources.

For a few groups every hatched animal counts, while it lives and when it dies: their young are added to the animals alive. Copepods’ count already includes animals the size of their larvae, and fish larvae are a group of their own. The copepod count combines Bar-On et al.’s (2018) estimate with our re-integration of Hatton and colleagues’ (2021) model of ocean life by size through the depth of each layer of water, as their methods describe; with their published totals the same steps would give a larger count (the copepods’ page has the steps). Most bird nestlings and mammal pups are not counted alive, and eggs are never counted.

Borrowed values

We found no published hours of pain for any wild animal, so the days they live are converted from Charity Entrepreneurship’s 2018 welfare scores with the same constant as other groups (above). Its index scores only some kinds of wild animal. Each score includes points for how the animal dies, so for the groups whose deaths are counted, dying is partly counted twice, once in the score and once in the dying profile (not corrected). Each range is the lowest and highest of its eight raters’ scores, three of whom, from its own team, gave identical scores. Charity Entrepreneurship gives no definition of ‘bug’, and its report describes insects, so its ‘wild bug’ score is an insect score. By the site author’s choice, copepods and the other tiny crustaceans, mites and springtails do not borrow it and have counts only (above); the other groups that are not insects keep it, labelled “borrowed from an insect score”. Spiders are a less clear case: its notes call spiders “predator bugs”.

Most groups borrow their welfare score, chance of feeling, welfare range and dying profile from another animal. Some groups take the welfare score of their own kind of animal, borrowed for some of their animals (labelled “partly borrowed”). Every borrowed value says so (“borrowed from …”) on the group’s page, with its source.

Copepods and the three other groups of tiny crustaceans, mites and springtails have counts only, because we found no published welfare score for them, and Charity Entrepreneurship’s ‘wild bug’ score is an insect score, which by the site author’s choice is not borrowed for them. Their welfare-adjusted population still uses a borrowed chance of feeling and welfare range: for the tiny crustaceans, the roundworm C. elegans’ chance of feeling (Rethink Priorities 2023) and shrimp’s welfare range (Rethink Priorities 2024).

How sure the numbers are

Every wild figure in SADs is given to one significant figure. The 90% range of a sum of wild groups (a family, or the wild total) comes from 20,000 random draws shared by groups that borrow the same value (the same count, chance of feeling, welfare range, welfare score or dying profile), so splitting a group never makes the total look more certain. A single wild group’s range comes from 200,000 of the same draws, because one group’s ends move more from one set of draws to the next than a sum’s. Each end is taken a little further out than the draws’ own 5th or 95th percentile, by about how far that many draws could miss it. The ends are then rounded outward, the low end down and the high end up, so that the chance variation of the random draws is very unlikely to make a range look narrower than it is.

Where wild animals live

The country map shows where people farm and catch animals. We found no consistent country-by-country counts of wild animals. The only ocean-wide pattern for copepods we found comes from a fitted model (Hatton and colleagues, 2021): it predicts zooplankton from satellite chlorophyll (how much algae grows), sea temperature and depth, fitted to net samples. Our copepod map integrates the model through the local water depth, following the authors’ methods, and scales its pattern to this atlas’s count. It assumes the same copepod share and size everywhere. The model’s data and code are openly licensed (CC BY 4.0). Missing-depth cells and water outside the source grid are hatched. Their share of the global count is reserved, not moved into the mapped cells. No suffering weights enter the map. On land we have no map: spreading the counts over land area would be a guess, and range maps of mammals, birds, reptiles and amphibians show where species live, not how many animals there are.

What we don’t know

  • Whether most of these animals can feel anything. We found no Rethink Priorities estimate for copepods, mites or springtails, among others, so each borrows another animal’s chance of feeling.
  • How their lives go. The only published welfare levels for wild animals we found are Charity Entrepreneurship’s four 2018 scores. Charity Entrepreneurship gives no definition of ‘bug’; its report describes insects.
  • How many there are. The counts leave out eggs, most bird nestlings and mammal pups, young fish between the larval stage and 1 gram, and animals the site has no count for, such as rotifers and tardigrades. Mammal and bird deaths may be too low, because their counts are mostly of grown animals.
  • How long dying takes. We found almost no measured times from the start of dying to death for wild animals; the hours counted here come from Welfare Footprint’s pain profiles for chickens, which it applies across vertebrates; using them for invertebrates is this site’s own extension.

Uncertainty

How the 90% ranges are worked out from random draws.

Read the whole section on Uncertainty

Intervals come from 1,500 random draws per group (a seeded Monte Carlo simulation); a wild group’s come from 200,000 draws (wild animals, above). Where a group’s chance of feeling has a low end just above zero (such as silkworms’), about one draw in 20 rounds it down to zero, and 1,500 draws cannot tell whether the 5th percentile is zero or a small number above it, so that end is worked out again with 50,000 draws. Counts, welfare ranges, P(sentience), welfare levels and the time spent in pain are each drawn from a distribution whose 5th, 50th and 95th percentiles match each input’s low, central and high values: published ones where they exist, otherwise our own ranges, set out below. AIM varies only welfare range and sentience; we also vary pain-hours, the conversion constant (×/÷2.5) and dying. Pain-hours vary by a factor of about two either way for modelled groups (for laying hens the low end is the cage-free value) and three to five for borrowed ones, our assumption. AIM’s end-of-life profiles, and those borrowed from them, vary ×/÷3. The dying profiles built for this site and the groups counted only while dying are drawn between their own low and high values. Within a group, inputs are drawn independently, which probably understates the uncertainty. A range for a sum of wild groups shares draws between the groups that borrow the same value instead (wild animals, above). The pain weights are not varied: they are a choice you make with the selector. On the Atlas, “Try your own numbers” lets you set each group’s chance of feeling anything and its moral weight; a number you set is used exactly, so its 90% interval no longer allows for uncertainty in it.

Farmed insects use Rethink Priorities’ 2026 population projection, published in 2025, for black soldier fly larvae and yellow mealworms. These are projections, not measured 2026 production. The source adds the two species’ low and high counts alive; those sums are not a verified 90% range for their combined population. This site uses them as rough input bounds in the calculation above. The resulting SADs range is this site’s modelled range, not a range published by Rethink Priorities. The population update keeps the earlier borrowed pain estimates and moral weights.

Country map

How each group’s total is shared out between countries, and which groups are not on the map.

Read the whole section on Country map

The map shows where the suffering humans cause happens, country by country. It does not estimate anything new. Each group’s yearly total is the same number as on the Atlas, under the same assumptions, and it is shared out between countries in proportion to where that group’s animals are farmed or, for fishing, to the catch of each country’s boats (FAO records catch by the flag of the boat, wherever it is landed). The countries’ values then add up to the group’s total.

value for a country = Σ over the groups on the map: group’s yearly total × the country’s share of the group’s animals
  • Where the shares come from. FAO’s country statistics: FAOSTAT counts of animals slaughtered and alive, and of beehives; FishStat tonnes of farmed and caught fish and crustaceans, turned into numbers of animals with fishcount’s average weight for each species; and the International Sericultural Commission’s tonnes of raw silk for silkworms.
  • Proxies (≈). Where no country counts the animals themselves, a related number that should rise and fall with them stands in: chickens slaughtered for the broilers alive at any time, the laying flock for culled male chicks, beehives for honeybees, and tonnes of silk for silkworms. The map’s tooltips and table mark these shares with ≈.
  • Production, not consumption. A country’s value follows where animals are farmed or, for fishing, the flag of the boats that catch them (FAO records catches by the flag of the boat, wherever the catch is landed). A country that imports most of its fish or shrimp looks small; one that exports most of what it farms looks large.
  • Why wild animals are not on the map. We found no consistent country-by-country counts of wild animals. Splitting the global estimates by land area would be an assumption, not data (wild animals).
  • The same suffering per animal everywhere. The pain estimates are global, so a broiler chicken in Brazil counts the same as one in Poland. Only the numbers of animals differ between countries, not how they are farmed.
  • Earlier years. The map’s timeline goes back to 1961. It keeps today’s suffering per animal and changes only the numbers of animals each country reported.

Every country’s shares, with their sources and licences, are in the downloadable data (countries.json and countries_history.json). The map page lists each group’s source and caveat in its Sources tab.

How this was built

AI agents gathered and checked the data: what they did, what other agents checked, and the limits of that work.

Read the whole section on How this was built

AI agents built this atlas. Claude agents from Anthropic and Codex agents from OpenAI gathered the data, checked it and wrote the site’s code. The site’s author directed the work: they set the rules and made the key decisions, such as counting only animals, measuring suffering in SADs, which sources come first, never filling a gap with a welfare score of our own, and having separate AI agents try to prove the research and the code wrong.

Sources are ranked in tiers. The main datasets list 315 sources: 243 in tier 1, 20 in tier 2 and 52 in tier 3. Tier 1 includes Rethink Priorities, Wild Animal Initiative, the Welfare Footprint Institute, Ambitious Impact, FAO and peer-reviewed papers; tier 2, other EA Forum posts and similar community research; tier 3, organisation websites, news, blogs and our own documented assumptions (the full lists). The rule for the research was that when sources disagree, the highest-tier source sets the central value and the others widen the range.

Other agents tried to prove the research wrong. For the research behind the SADs, the agents doing it were told to read the original source and quote the words behind each key number. Separate agents were then told to try to prove those numbers wrong by going back to the sources, and a last agent had to apply their corrections and leave out any number shown to be wrong.

Other parts were checked too. One check downloaded FAO’s country data again: every one of about 160,000 FAO figures it compared matched the site’s copy, but it found other problems, the biggest being the world totals behind the map’s timeline: beehives, for example, seemed to fall 17.5% in 2024 only because 17 EU countries had not yet reported. That was fixed. The pages and the code were reviewed in the same spirit: reviewer agents listed problems, other agents tried to reproduce each one, and the fixes were checked again by other agents, often over several rounds. A test also checks that the model, given Ambitious Impact’s own inputs, matches its published totals for broiler chickens, hens in battery cages and sows to within 1%.

Every number links to its sources, or is labelled as our own judgement. On the Atlas, the source links under the headline figures and each chart’s Sources tab list the sources behind what is shown, with their tiers. Each species page lists its sources, and its “Notes on each number” say where its counts, lifespan, welfare range and chance of being sentient come from. The Data page lists these sources alongside those for the neuron estimates, ocean map and wild-animal evidence. You can search and filter them by tier and download the data files.

The limits:

  • Research websites could not always be reached. The first research, in mid-October 2026, downloaded Ambitious Impact’s spreadsheets and FAO’s data directly. Later, the computer the research ran on could not reach many research sites, including the EA Forum, Rethink Priorities, the Welfare Footprint Institute, Ambitious Impact, FAO and journal sites. Some checks then rested on search-engine summaries instead of the source itself, and some source notes on the Data page say so. The search for published estimates for the six groups kept by people that have no SADs was one of these, which is one reason we say only that we found none (above). The source checks of the wild-animal figures, on 23 October 2026, ran with full network access and read the sources themselves: Charity Entrepreneurship’s 2018 index and reports, Rethink Priorities’ website, posts and data since 2024, the Welfare Footprint Institute’s pain definitions, and the papers and data behind the wild counts. A few publishers refused automated readers, and the source notes say where. Those checks found that Charity Entrepreneurship’s ‘wild bug’ score is an insect score, and no Rethink Priorities estimate newer than 2024 for copepods, mites or springtails.
  • Some inputs are this site’s own judgement, not published figures: several lifespans (for example of farmed fish and wild amphibians), some counts (for example of dogs and cats, of wild reptiles and amphibians, and of wild animals’ deaths), the uncertainty ranges around the pain estimates, how much wild animals, and fish and crustaceans caught by fisheries, suffer while dying, and stand-ins where one animal borrows another’s estimate (ducks and turkeys use the broiler chicken model, for example), and the method that turns a published welfare level into hours of pain for wild animals, cattle, and sheep and goats (see above). The sections above and the species pages’ notes say which.
  • AI agents can make mistakes. Every check described here was done by AI agents checking other AI agents’ work, not by human experts. The checks did catch real errors: one round found two welfare scores on the site that an agent had noted down from memory and never checked against the book they came from, and they were taken off. Others may remain.

If you find a mistake or a better source, please email ncr0517@gmail.com (more under Corrections and contributions).

Source hierarchy

The three tiers of sources, and which one sets a number when they disagree.

Read the whole section on Source hierarchy

Every number links to its sources or is labelled as our own judgement, and each source carries a tier:

  • Tier 1 Rethink Priorities research (including RP reports posted to the EA Forum), Wild Animal Initiative, the Welfare Footprint Institute, Ambitious Impact (Charity Entrepreneurship), FAO/FAOSTAT, fishcount.org.uk, Our World in Data, peer-reviewed papers, Open Philanthropy (Coefficient Giving) and Animal Charity Evaluators. Their research counts as tier 1 wherever they post it, including research write-ups on their own websites that look like blog posts.
  • Tier 2 Other EA Forum posts, and similar community research. Useful, often the only source on a question, but weighted below tier 1.
  • Tier 3 Organisation websites, news stories and press releases, newsletters, blogs and personal essay sites, and our own documented assumptions. News stories, press releases and newsletters stay in tier 3 whoever publishes them.

When sources disagree, the highest-tier source sets the central value and the others widen the range. Some numbers are marked unverified in the data files, and source notes say when a check rested on a search engine’s summary rather than the source itself.

Caveats that matter most

  1. Wild animal numbers are far larger than the human-caused ones.

    So the site shows them beside the human-caused total and never adds the two. Their SADs come from welfare scores converted with one constant, mostly borrowed from other animals, not from pain data, and most of them come from small invertebrates, whose chance of feeling is very uncertain. Treat any total that includes wild animals as illustrative. The counts leave out eggs, most nestlings and pups, young fish between the larval stage and 1 gram, and animals the site has no count for (wild animals).

  2. Wild nematodes are left out by default, a choice of this site.

    Under the default assumptions they would add more than all other wild animals together. Selecting neuron-count weights includes nematodes. You can change “Include wild nematodes” in Explorer. Their values are in the Default values table below.

  3. Six groups kept by people have no SADs.

    For managed honeybees, dogs and cats, geese, rabbits, and animals farmed for fur or used in research, we found no published estimate of their welfare level or hours of pain, so the human-caused total leaves them out (see above).

  4. The pain weights are contested.

    They have changed sharply between AIM’s versions. Read rankings together with the weight selector.

  5. Only hens and broilers rest mainly on Welfare Footprint analyses.

    Pigs, salmon and silkworms are AIM staff estimates, shrimp come from Rethink Priorities, and for farmed fish other than salmon we found no model of their own, so they use AIM’s salmon model.

  6. Sleep is handled inconsistently in AIM’s models.

    The pig model subtracts eight hours a day of sleep, the silkworm model discounts sleep for one issue (hunger and thirst), and the hen, broiler, salmon and shrimp models do not. Per-day comparisons across species are affected.

  7. Welfare ranges for invertebrates rest on a small evidence base.

    RP’s own 5th–95th percentile intervals run from 0.001 or less up to about 1, at least three orders of magnitude, and the 5th percentile is zero for bees, silkworms, and crabs and crayfish.

  8. Counts of fish, shrimp and insects are rough.

    They are estimated from tonnage divided by average body weight, and published ranges are wide.

  9. The model is additive and linear.

    It assumes a day of suffering in one animal counts the same wherever it happens, and that suffering in different animals can be added up. Some ethical views reject this.

  10. The funding figures are rough.

    Those in the burden-vs-attention chart combine different years and leave out government spending. The chart compares each problem’s size with its funding (SADs per year per dollar a year). That is not the cost-effectiveness of any intervention.

AIM itself warns that “SADs introduce a large amount of uncertainty and subjective judgement calls and so shouldn’t be used unless necessary.”

Default values

Every group’s default inputs, in one table.

Read the whole section on Default values

Every default the model uses, under AIM v.2 weights. The same values are in the downloadable data.

Corrections and contributions

If you find a mistake, a better source or a number we have missed, please email ncr0517@gmail.com, with a link to the source if you have one. The data lives in plain JSON files, built from the research files by a short script. Every file can be downloaded from the Data page, with the source behind each number.