Here’s a fairly safe prediction: much of the near-term potential harm from AI is potential harm to nonhuman animals.

It’s a safe^[1] prediction because there are so many more nonhuman animals than there are humans. Many of these animals presently live in conditions where their lives are influenced or controlled by human activity. Since AI will increasingly substitute for human activity, many animals’ lives might therefore someday be controlled or influenced by AI. Indeed, while there are just over 8 billion people, there are around a zillion^[2] animals farmed each year, the majority on factory farms, where their lives are nearly entirely determined by human choices. Factory farms, then, are an enormous and terrifying arena for AI-caused harm.

Accordingly, when we talk about AI governance and risks from AI, risks to nonhuman animals should be a major part of that conversation, alongside e.g. biorisks and discrimination. Unfortunately, this is basically never the case. For instance, the mostly commendable recent International Scientific Report on the Safety of Advanced AI: Interim Report, a long and comprehensive document, with a long and comprehensive name, written by committee and designed to achieve global consensus on AI risks, does not even mention animal welfare risks – does not even mention animals, or agriculture, or anything like that – at all.

This blog post is a small attempt to rectify the imbalance. I want to share my thoughts about a specific AI technology: precision livestock farming. I’ll argue that, while I am uncertain about whether AI precision farming technology will be good or bad for animal welfare on a per-animal basis, I think we can be much more confident that it will be bad on a total population basis, since it will likely increase the total number of farmed animals. Largely in view of this argument, but also for political feasibility reasons, I think we should consider advocating for restrictions on precision livestock farming in industrial settings.

‘AI’ can mean a lot of different things

When we talk about the effects of AI on animal welfare, it is essential to distinguish different types of AI systems. ‘How will AI affect animals?’ is uselessly broad. There are a number of different AI technologies we should distinguish between: precision livestock farming, AI genetics and breeding programs, language models used by the general public, AI adoption by activists, AI adoption by alternative protein producers, models used to decode animal thoughts, and others. All quite different, and it’s hard to know what the aggregate effect of these technologies will be.

A lot of these technologies are important, but my focus here is on precision livestock farming.

What is precision livestock farming?

Precision livestock farming (PLF), as I understand it, basically means putting in place a bunch of sensors and instruments that gather data about individual farmed animals, and then having some algorithm – this could be a neural network, but might be something much more basic – manage features of the animals’ environment, based on the gathered data, in order to ‘optimise’ farming. The classic example is an automated feeding system, which gathers data about individual animals’ feeding behaviour and health and stuff and automatically distributes food in response to the gathered data.

In theory, systems like this can be designed to make sure all animals are having a great time. In practice, systems like this will often be designed, I expect, to further intensify animal agriculture.

What do we know about the particular technologies that are currently being implemented?

According to this paper (Schillings et al, 2021), the following technologies are already commercially available:

In cattle, smart camera systems using computer vision combined with deep learning can monitor eating time and feed availability at group level, while neck collars equipped with 3D accelerometers continuously monitor rumination and eating time in individual animals. Gastrointestinal health can also be monitored using boluses sitting in cattle reticulum which measure pH and temperature. In pigs, RFID ear tags are used as part of electronic feeding systems, while in the aquaculture sector, hydroacoustic-based technologies and cameras combined with machine learning allow to monitor fish pellet consumption and appetite. Finally, water consumption can be monitored with commercially available boluses in cattle and with sensors in cattle, pigs, and poultry.

However, given the pace of AI and robotics advances, we should expect the landscape to change considerably. New PLF technologies will be available soon – potentially much more capable ones.

Schillings et al also offer us some information about technologies in development:

Other systems which are currently in the development stages (categories P1 to P2) can monitor ingestive behaviors in free-ranging cattle, goats and sheep using acoustic monitoring (Navon et al., 2013; Chelotti et al., 2016). In poultry, Aydin (2016) developed a sound-based monitoring system to detect short-term feeding behaviors of broiler chickens by recording pecking sounds. RFID systems have been used to monitor feeding patterns in pigs (Maselyne et al., 2016b; Adrion et al., 2018), turkeys (Tu et al., 2011) and laying hens (Li et al., 2017). Image analysis and binocular vision techniques have been developed to monitor feeding in pigs (Yang et al., 2020) and poultry (Xiao et al., 2019), while sensor-based systems can monitor feed intake in goats (Campos et al., 2019) and turkeys (Chagneau et al., 2006). Technologies at phase P2 also introduced the possibility to use 3D-vision to automatically assess reticulo-ruminal motility in cattle (Song et al., 2019). Finally, drinking behavior can be monitored using RFID in pigs (Maselyne et al., 2016a) and a combination of sensors and RFID have been used in cattle (Williams et al., 2020). Accelerometers have been used to monitor drinking in calves (Roland et al., 2018), while camera-based systems have been developed to monitor drinking behavior in pigs (Kashiha et al., 2013a) and chickens (Xiao et al., 2019).

And they provide this table, of PLF system development broken down by species:

TABLE: ‘Development stages of PLF technologies related to the “nutrition” domain of the Five Domains Model for different species (expressed in phases—P1, proof-of concept stages; P2, validation stages; P3, commercialization phases).’

Given the difficulty of predicting which specific technologies will be feasible, and recent rapid advances in AI, I wouldn’t place too much stock in this though.

PLF is a bundle of many different technologies

One thing that should be clear from the above is that, not only are there many different types of AI systems that matter for animals, there are many different types of PLF systems that matter. It’s reasonable to question whether it’s worthwhile to think about PLF in general, rather than considering each individual PLF system separately.

Sometimes, the answer to this question is ‘no’, ‘PLF’ is too broad a label. As I’ll discuss, some PLF systems are plausibly good for animals and some are likely bad – it seems important to distinguish these.

Still, I think it’s sometimes worth thinking about PLF systems collectively, because:

Many PLF systems have the common attribute of decreasing per-animal production costs, as I’ll discuss in a second.
PLF systems might be regulated collectively, and it might not be possible to disambiguate good and bad systems in political contexts, making it important for animal advocates to understand the collective effect of the bundle of technologies.
I’d like AI governance work to begin considering risks to animals as a high-level category of AI risk. Considering the underdeveloped state of this discourse, I expect that basic awareness of PLF as a large AI-driven risk to animals probably needs to predate nuanced analysis of individual PLF systems.

Is precision livestock farming good for animals?

It’s important to distinguish between technologies that improve (or worsen) animal welfare on a per-animal basis, and technologies that improve (or worsen) total animal welfare. These can come apart when the number of animals changes. For instance, a technology that has a very small positive effect on every chicken in a farm will make each individual chicken better off. But if this technology also allows the farm to scale up production, many more chickens might be farmed in total, increasing the number of animals who live lives of misery. As a consequence, total welfare would diminish.

Ultimately, when I ask whether PLF is good for animals, I’m asking about total welfare. But it is helpful to decompose the problem into two parts: the effect of PLF on each animal, and the effect on the total population size.^[3]

Per-animal effects are uncertain

I think the expected average (i.e. per-animal) welfare change from implementing more PLF practices is highly uncertain.

Reasons for optimism

On the one hand, farms using PLF have an enhanced ability to monitor animal welfare problems, and fix them.

For instance, a PLF system can monitor animal health, catching disease before it spreads. This could reduce the number of disease-afflicted animals, and reduce the need for culls, which often employ more painful methods of death than normal slaughter methods.

Additionally, the ability to track new data on animal welfare opens up some cool possibilities. Perhaps regulators could require farms to report their welfare data, which could be used for welfare labelling (rather than using more simple proxies for welfare, like stocking densities). Activists could advocate for more public reporting of welfare metrics, and insist that companies achieve certain targets.

AI monitoring might also reduce the need for painful systems used for identifying and monitoring individual animals at present, like ear tags, earmarking, and branding.

Reason for pessimism

On the other hand, farms using PLF have an enhanced ability to more cheaply / intensively farm animals, sometimes at a cost to each animal’s welfare. For instance, maybe more capable monitoring systems enable higher stocking densities, reducing animal welfare. Or maybe AI algorithms that determine the most efficient time to inseminate sows mean that pigs undergo more pregnancies, and therefore suffer from increased risk of death from prolapses.

In a nice article for Aeon, Virginie Simoneau-Gilbert and Jonathan Birch also point out that, if PLF is efficiency enhancing, and more likely to be implemented at factory farms than at smaller, higher welfare farms, factory farms might increasingly outcompete smaller farms, meaning that a greater proportion of animals might be farmed in the inferior, factory conditions. This would make the life of the average individual animal worse, since it would be more likely to be a factory farmed life.

Still on the negative side of the ledger, it’s also worth noting that the history of technological innovation in farming is a devastatingly painful one, with technological development enabling new cruelties and worsening welfare standards. This macrohistorical ‘outside view’ provides some reason to think that this particular set of technological innovations might also be likely to be bad on a per-animal basis.

Reasons for confusion

One reason for additional uncertainty is the effect of relative price changes. PLF could change the relative costs of producing different animals, meaning, for instance, cheaper cow meat relative to chicken meat, or vice versa. This change in relative prices would then lead to a change in demand for beef and chicken. If cows generally experience, on a per-animal basis, better lives in factory farms than chickens do, more demand for cow meat would mean higher average welfare.

It’s not super clear to me which species of animals could see the largest reduction in production cost as a result of PLF, so it’s hard to know how relative prices will change. While, as I mentioned earlier, present adoption has been greatest in the dairy industry, it’s not at all clear that this will continue in the long run.^[4] Especially since advances in AI capabilities (like cheaper and more powerful computer vision systems) could change things.

Another reason for uncertainty is that the exact same PLF system might be used to increase or to decrease per-animal welfare, depending on underlying conditions. Consider an AI model like this one that predicts the rate of animal heat-induced stress in a factory farmed animal population, given the temperature, which can be adjusted using a cooling system (which costs money to run). Imagine that, before the model is implemented, farmers use a suboptimal amount of cooling C which causes a heat-stress level of S per animal.

What will farmers do once they can more reliably predict how changes to C will affect S?

Maybe farmers realise they can actually reduce cooling a lot more, with only moderate consequences for heat-stress levels, and they decide that a moderate increase to heat-stress level is worth incurring. Maybe they realise they can actually reduce cooling without any increase to heat-stress, just by changing the time of day they do cooling, so C goes down and S stays constant.

Maybe farmers realise that they can actually improve heat-stress levels a lot, with only moderate cooling cost increases, and they decide that a moderate increase to cooling costs is worth it. Maybe they realise that they can actually reduce heat-stress without any increase in cooling costs, just by changing the time of day they do cooling, so S goes down and C stays constant.

Or maybe they reduce their cooling costs a bit, and reduce heat-stress a bit, too.

Cooling costs and heat-stress levels could go up, go down, stay the same. We don’t know; it depends on what the underlying trade-off relationship between C and S turns out to be, and what farmers’ goals are. And we don’t know what the underlying trade-off relationship is like, or just how much farmers want lower heat-stress vs lower cooling costs. So this consideration raises our uncertainty about the goodness of the technology.

(However, I do think that, in practice, economic incentives make the welfare-reducing use more likely.)

The per-animal upshot

If I had to bet, I’d say the collective upshot of all these considerations is maybe more likely bad than good? But I think the fairest thing to say is that we don’t know: the per-animal welfare effects of PLF are highly uncertain.

PLF will likely increase the number of farmed animals

In contrast to the per-animal effect, I think the expected aggregate welfare change from implementing more PLF is negative, and more certain.

The reason is that PLF is likely to be implemented only if it is profitable for a meat producer. The main way that PLF would be profitable is if it decreases the average cost of producing an animal to be slaughtered. Almost certainly, this would correspond with a lowering of the marginal cost of production, meaning that it would become more profitable to scale up the number of animals farmed.

Consider an automated feeding system, our paradigm case. This may have benefits for individual animals: by automating feeding times, perhaps the PLF system could be designed to make sure animals are healthier and happier than they would be if their feeding schedule was set by a human. But the system would also lower the marginal cost of production. That’s like the whole point! This paper, citing a Patience et al 2015, says “feed is the most important cost component in many livestock operations and in commercial growing–finishing pig-production systems and represents between 60% and 70% of overall production costs.” By implementing the PLF system, a farm can reduce feed costs, making each animal less expensive to produce. Meaning that more animals can be farmed in total. Bad.

Predictably, a quick scan of the websites of some manufacturers of PLF systems reveals that they advertise their products primarily on the basis of efficiency gains. And this review of animal welfare arguments against PLF claims “the PLF-systems that are being brought to market and adopted by farmers are those that focus on production efficiency and farmer quality of life.”

I’m unsure exactly how much farms would be likely to scale their operations if they had a low-cost automated feeding system. (One statistic I found had it that ‘Rotary dairies with automated technologies have been shown to have 43% higher labor efficiency and 14% higher milking efficiency’.) Good economic / business research here would be really helpful. (It might exist, I haven’t had time to look thoroughly.) But given the cost of feed mentioned in that paper, it sounds like scaling could be quite significant.

As with per-animal welfare, changes to the relative prices of meat by species also affect the total number of animals farmed. Imagine what happens if PLF makes cow meat more expensive relative to chicken meat. This change in relative prices might lead to some of the demand for cow meat being replaced with demand for chicken meat. When demand for meat from large animals (like cows) is replaced with demand for meat from smaller animals (like chickens), this means more total animals farmed, since more small animals are needed to produce the same quantity of meat. This is sometimes called the ‘small animal replacement problem.’

Since we don’t know which species will see relative price decreases from PLF, this increases our uncertainty about the population size effects of PLF adoption. It is hard to say whether the effect of relative price changes is an upward or downward pressure on the total number of factory farmed animals.

Still, even if the effects from relative price changes happen to be welfare-promoting, they seem unlikely to compensate for the effect of making all animal agriculture cheaper. So I feel relatively confident that the ultimate effect of PLF will be to increase – maybe dramatically – the number of animals in factory farms.

This idea, that PLF could cause more total animals to be farmed, is not mentioned anywhere that I can find. Neither the Simoneau-Gilbert and Birch essay nor this paper on the topic discuss the idea. Nor is it one of the twelve listed threats in Twelve Threats of Precision Livestock Farming (PLF) for Animal Welfare. Accordingly, I think existing welfare assessments of PLF are dramatically underweighting just how bad PLF could be.

More research, though, especially research that quantifies the production cost effects of PLF by species, would be very welcome.

Welfare-effects, all things considered

To summarise: We can’t say with much certainty whether PLF will have positive or negative per-animal welfare effects. We can say, with some greater confidence, that PLF probably will increase the total number of animals living miserable factory farmed lives.

Taking these considerations together, on net, PLF seems likely to decrease total animal welfare.

An aside about the environmental effects of PLF

Precision livestock farming has also been proposed as an environmentally beneficial technology that could reduce the carbon emissions of industrial farms. For instance, with an automated feeding system, farms can reduce the amount of feed they require per animal. And feed is carbon intensive.

However there is an important analogy to welfare: while the per-animal feed requirements could decrease with an automated feeding system, the simultaneous reduction in the marginal cost of animal agriculture would likely cause farms to increase the total number of animals. This could lead to more feed consumed in total, meaning more emissions, on net.

When I quickly looked through articles about the environmental sustainability of PLF, I found a number of papers that described PLF as a positive for environmental sustainability, without considering this effect at all. Concerning.

What should the animal advocacy movement do?

Promote positive uses of PLF?

One idea is to try to promote positive uses of PLF. At the recent AI, Animals, and Digital Minds conference, Walter Veit spoke about some promising uses of the technology to monitor animal welfare, and argued that positive uses of PLF should be made obligatory for farms.

However, since PLF is, I’ve argued, likely to be bad on net, we need to be careful. We should make sure that anytime we support welfare-focused PLF, we don’t, as a side-effect, make it cheaper or more politically acceptable to implement efficiency-increasing PLF too. For instance, animal advocates should not fund dual-use biosensing equipment for farms with the optimistic hope that the equipment will only be used for welfare-focused stuff.

We should also be worried that positive rhetoric about a few niche use-cases might act as rhetorical cover for the industry to increase the political acceptability of very harmful (but profitable) use-cases.

In general, I’m not super optimistic about this idea.

Advocate for restricting the use of PLF on factory farms!

Another idea is to advocate for restricting the use of PLF on factory farms. This could look like a ban on AI models above a certain number of parameters being used as part of Concentrated Animal Feeding Operations. Or it could be a restriction on using certain types of PLF or using it for certain specific industries.

Along these lines, Simoneau-Gilbert and Birch’s article proposes some intriguing principles for adoption and regulation of PLF, including the principles (among others) that

PLF should not be used to increase stocking densities, and

farms should be required to be both transparent about and responsive to welfare problems that PLF systems diagnose.

Because the benefits of the good use-cases of PLF seem likely to be outweighed by the harms that PLF can enable if it is used to increase the efficiency of animal agriculture operations, I think we should be supportive of even broad, blanket bans on PLF systems.

If this strategy fails, advocates can still press for positive uses of PLF-systems later. However the opposite is less feasible. Once PLF systems are in place, it becomes more difficult politically, logistically, and financially to remove them or restrict further use.

Why restrictions on PLF systems in factory farms might be feasible

I think regulations on PLF systems in factory farms might be feasible, actually? I offer four reasons…

The particular composition of who has influence in AI policy
Possible coalition with small farmers
Background support for the idea from the public
The financial costs of the policy occur in the future

…And one caveat: these reasons are a deliberate attempt to assemble grounds for optimism about the feasibility of restrictions on PLF – not a neutral list of reasons for and against feasibility. There are obviously also a bunch of reasons for pessimism, like: the industry can point to a few positive-welfare uses of PLF as a rhetorical distraction from the bad uses, it’s just really hard to put any restrictions on the animal agriculture industry at all, etc., etc..

Disheartening caveat out of the way, let’s take these in turn.

1. The particular composition of who has influence on AI policy

As far as I’m aware, the animal agriculture industry hasn’t done any lobbying specifically aimed at key figures or platforms in AI policy. Perhaps there is an opportunity to preemptively shape the conversation in AI policy, before the animal agriculture industry’s considerable political heft gets directed that way?

Also, because the PLF industry is still nascent, the PLF industry presumably has not yet built up a sophisticated lobbying operation.

Furthermore, a number of key people in AI policy have a history of engagement with the animal advocacy movement, attend shared conferences with the animal advocacy movement, and have a general willingness to consider animal welfare. This makes advocacy in this area dramatically more tractable, in my view. I think this is the strongest reason of the four I offer.

2. Possible coalition with small farmers

Installing a PLF system involves large up-front costs, which will be more difficult for small farmers to afford. PLF will therefore be expected to be a force for market concentration, advantaging large corporate operations at the expense of smaller farmers. Perhaps this opens up the possibility of advocating for restrictions on PLF systems in coalition with small farmers.

Plus, people might fear that AI / PLF systems will literally automate away farming jobs, putting farmworkers out of work. Animal advocates allying with small farmers (and people who care about them, or care about preserving traditional ways of life) could make for a powerful coalition.

3. Background public support

There is background public support for both increased restrictions on factory farms and on AI adoption. In a 2023 AI Policy Institute poll, 72% of Americans indicated that they want to slow down AI development and usage. And in a 2019 Johns Hopkins poll, 57% of Americans indicated that they support greater oversight of existing industrial animal farms. These are pretty big numbers.

If a coalitional approach involving small farmers is possible, as discussed above, this might increase public sympathy for the cause. The coalition makes the distinction between factory farms and small farmers inherently salient. And massive factory farms (sinister, industrial) are much more readily typecast as villains than small farmers (familial, wear overalls). This suggests a ready and potent political narrative.

Also: the policy, in slogan form, is easy to convey. ‘Ban AI in factory farms!’ That’s the tweet.

And doesn’t this just seem pretty politically compelling?

Obviously, though, background public support for greater oversight of existing industrial farms has not translated into greater oversight. So this argument alone is far from sufficient.

4. The financial costs of the policy occur in the future

While many PLF systems have been implemented already, the most scary ideas are not yet in operation. Some probably haven’t been invented. Regulation now imposes future opportunity costs on the animal agriculture industry, but this may be more palatable^[5] than regulation that diminishes the industry’s present-day bottom line.

Additionally, restricting technologies that have not been invented or procured does not impose transition costs associated with first-time implementation of the rule. Compare this to a ban on, say, gestation crates: such a ban requires the industry to buy new crates and change other features of the production process, whereas a ban on future technology doesn’t require the industry to adjust its present production process.

Let’s keep talking about AI and animals

As I’ve mentioned, there are a lot of avenues for research and advocacy in this neighbourhood. Please get in touch if you want to discuss! Or even just: if you’re working in AI policy, and writing some document listing risks from AI, toss in a mention about risks to nonhuman animals. That’d be a cool start.

These ideas are inspired in part by the excellent recent conference on AI, Animals, and Digital Minds. Thanks to the organisers of that, and the speakers and participants. I hope my ideas inspire more ideas in turn.

You might also want to check out this piece on ‘Animal Advocacy in the Age of AI’, this website on the same, or join the Hive channel on AI and animals.

Thanks to Ben Stevenson (he writes here), Constance Li, Max Taylor, and Maya Misra for helpful comments on this piece.

This content was originally posted on my blog, Before Porcelain. I expect to write more on related issues there in the future.

^
This could be a relatively smaller proportion of total harm if there are likely to be many at-risk digital minds in the future, or if there will be many more humans but not many more animals, or if AI just affects humans way more profoundly than it affects other species.
^
80 billion large land animals, maybe over 100 billion farmed fish, and hundreds of billions of shrimp and insects.
^
Economists might like to describe this as the difference between the ‘intensive and extensive margin’.
^
Might we expect that AI systems ultimately prove most profitable in the industries focused on farming many small animals, since many small animals means potentially bigger training data sets, and the alternative option of using a human to monitor the individual animals is more costly (since there are more of them)?
^
Here’s another reason / hope for why it might be more psychologically palatable to regulate the future. (I don’t give this hope much credence, which is why I’m banishing it to a footnote.)

Surveys sometimes report that large numbers of people intend to go vegan or vegetarian one day, but eat meat right now. A theory: maybe part of society’s inability to stop torturing animals is that we’re sorta procrastinating? Sure, we’ll go vegan one day, but right now that sounds hard, and we really want McDonald’s. Sure, factory farms are bad, someday we’ll get around to that, but did you hear us about the McDonald’s?

On this theory, restrictions on PLF, because they primarily constrain future meat production rather than present-day meat production, might be more psychologically attractive than other approaches.

I hope? This whole factory farming thing wasn’t supposed to be for forever, right?

Animal advocates should campaign to restrict AI precision livestock farming