Thanks for the post, Hazo!
[...] But if the target of the campaign is capable of implementing the change themselves, impact can begin a lot faster (e.g. Shrimp Welfare Project [SWP] giving humane stunners directly to shrimp producers comes to mind).
I agree SWP's Humane Slaughter Initiative (HSI) looks great under short AI timelines.
Nice post, thanks for sharing @Hazo! I think it's very valuable for animal advocates (or anyone trying to help others) to consdider wether they should adjust their approach given that transformative AI may be on the horizon. This is exactly the type of conversation we're trying to encourage at AI for Animals (soon to be 'Sentient Futures'). I'd love to see you join our Slack and contribute to this discussion - I will be sharing your post with our community (provided that's okay with you!).
A few points that might be helpful.
Overall, my personal perspective, and I do work on corporate commitments, they still seems like a good bet. Due to the combination of impact on animals in near/medium term, increasing the competence of campaigning organisations and embedding animal welfare policy within food companies.
Great post, thanks for looking into this! I previously noted four different types of interventions one might want to prioritize given AIxAnimals; I'd love to hear your thoughts on the implications on this intersection from a broader, zoomed out perspective!
It doesn't seem to me like AI could radically change the economics of higher welfare products / alternative proteins without other broader transformative effects, or at least if it does, it seems like it would be for a pretty brief period of time before things get very weird. So I think "past X years" ( to use the framework of your comments) there should be heavy discounting and I wouldn't recommend a save and invest strategy. So to address your four types of interventions:
Executive summary: This exploratory post examines how quickly different animal welfare interventions can deliver tangible benefits to animals, concluding that if we take short transformative AI timelines seriously, the animal advocacy movement may need to prioritize interventions with faster speed to impact—even if that means shifting away from longer-term strategies currently favored in the space.
Key points:
This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.
I have not contemplated deeply the meaning of Rethink Priority's findings on cross cause prioritization, but my perhaps shallow understanding was that despite somewhat high likelihood of AI catastrophe arriving quite soon, "traditional" animal welfare looked good in expectation. I think the point was something like despite quite high chances of AI catastrophe, the even higher chances (but far from 100%) of survival means in expectation animal welfare looks very good. So while it is not guaranteed animal welfare interventions will pay off due to an intervening AI crisis, it is still worth the bet unless you think value growth is extremely high (cubic or logistic) and there are only a very few periods from now until "infinite time" that x-risk will be high (only one or two such periods). I did not read your post carefully, but did you take this into account? That even if there might be a 30% chance of imminent AI catastrophe, the remaining 70% chance of "success" makes animal welfare with longer time horizons still look good in expectation?
Recently, @Lizka and @Ben_West🔸 published A shallow review of what transformative AI means for animal welfare. The main conclusion of this review was that animal welfare interventions should be heavily temporally discounted due to the possibility of transformative AI on short timelines.
A reaction I had when reading this piece was that things tend to happen very slowly in animal agriculture, and even big wins like a corporate welfare commitment can take years before a specific animal is concretely better off. I therefore thought it might be useful to look at some of the main animal welfare interventions and assess how quickly they can help animals in the best case scenario.
A conclusion from this analysis is that animal interventions vary significantly from how quickly they start to have impact, with some interventions having impact almost immediately, some having predictable impact within some period of time (which can be up to a few years), and some having impact at some unspecified point in the future. Optimizing for speed to impact might be a new kind of frame under which animal advocates can evaluate and prioritize interventions.
These are just some preliminary thoughts that I wanted to get out there in the spirit of "shallow reviews" (also given the analysis itself is about the importance of speed). I'd welcome additional thoughts / feedback / pushback.
Many animal welfare interventions achieve impact through lowering demand for animal products. In this category, I include things like starting a plant-based meat company, or doing vegan advocacy.
While it seems clearly possible to have wins in this area extremely quickly, there will often be a delay on having impact because of the structure of the supply chain for animal products.
For the simplest example, a beef cow in the US is generally 18-22 months old when they are slaughtered. This means that, to some extent, the beef supply in the US for the next 18-22 months has already been decided. In theory, marginal reductions in beef demand should decrease the price of beef in the short term, and then only reduce supply in the longer time once the supply chain has had time to adjust. Therefore, it might not be possible for meat reduction advocacy to help cows faster than 18-22 months. In other words, beef consumption reduction within 18-22 months of AI transformation might not have any beneficial effect.
Chickens live significantly shorter lives than cows - generally 4-6 weeks. But the supply chain is also a little more complicated, muddying the picture of how changes in demand affect supply. Breeders keep purebred lines of genetically optimized chickens that are crossbred over four generations to yield commercial broiler chickens. According to Claude, commercial broiler breeders live to ~16 months old. If there was a large, sudden, decrease in the demand for chicken meat, the size of the breeder flock would remain unchanged, meaning that the same number of broiler chickens would be produced. In the short term, chicken meat prices would be depressed until the entire supply chain had time to adjust to the new market reality. This would happen through breeders electing to retire older, less productive breeders, earlier than they otherwise would. It's not clear to me how long this process would take, given that it would need to happen for parents, grandparents, and great grandparents. It's likely not as slow as the 18-22 months for cows, but it's not an immediate process either.
Egg-laying hens are produced via a similar multi-generation cross breeding process as broilers, although they are themselves kept for up to two years. I would guess that the process of translating shifts in demand to shifts in supply is slower than for broilers, and faster than for cows, but I'm not sure.
Overall, even though there is a delay between reduction in meat demand and reduction in meat supply, it seems like this process is fast enough that meat reduction in the short term will have some impact before AI transformation under short timelines.
Much of the EA animal movement is currently focused on campaigning for corporate commitments around higher welfare practices like cage-free or the Better Chicken Commitment. It's easier for the supply chain to implement some changes compared to others, so even if the movement succeeds in generating demand for better practices through campaigning, the speed at which the practices are implemented may vary based on the ask.
For example, transitioning from caged egg production to cage-free involves retrofitting existing facilities (which can be quite large), or constructing new cage-free facilities. Each of these construction projects can take months, or years to complete. It's likely that egg producers try to anticipate where the market is going, and may start this process before there is confirmed demand (for example, I've heard that there aren't many new caged farms being constructed any more, because it's assumed that most growth is happening in cage-free). However, there might generally be a negative correlation between how counterfactual a campaign is and how long it takes for the supply chain to catch up with demand. For example, if a huge egg buyer suddenly commits to cage-free, the industry may be caught off guard and need to build completely new cage-free facilities to meet this demand. (This phenomenon may hold for any corporate campaign ask.)
Slower growing breeds of broiler chickens may take the longest amount of time to scale up. Because of the multi-generation structure of the broiler supply chain, there is a limit to how quickly capacity for slow-growth broilers can be added. New great-grandparent stock must be added and be given time to reach laying age, then grandparent stock, then parent stock, and only then can additional slow-growth broilers be created. This process could take years.
One the other hand, asks which involve installing a new piece of equipment in an existing facility may have impact relatively quickly. For example, a new controlled-atmosphere stunning machine can be added to a slaughter plant and start to have impact right away. The same holds for an in-ovo sexing machine installed in a hatchery, or a fish / shrimp stunning machine installed on a boat.
These three classes of equipment have an additional benefit of targeting more concentrated parts of the supply chain. Generally, a few hatcheries supply many farms, which then feed back into a few slaughter facilities. Interventions targeted at these concentrated part of the supply chain may be able to scale up quicker than interventions targeted at the farms themselves.
Another aspect that affects the speed to impact is how far removed the target of the campaign is from the part of the supply chain that implements the chain. For example, if a food retailer that buys animal products commits to change, they have to work with their suppliers to implement that change, which may take time. If those suppliers can't implement that change themselves e.g. they're buying cull-free layers from their hatchery, or slower growing breeds from their genetic supplier, then that may take additional time. The amount of time this will take will, of course, depends on the company in question, how easy it is to switch to other suppliers, etc. But if the target of the campaign is capable of implementing the change themselves, impact can begin a lot faster (e.g. Shrimp Welfare Project giving humane stunners directly to shrimp producers comes to mind).
Technological innovation: Developing new technologies and scaling them up generally takes a very long time, especially in food and agriculture. For example, cultivated meat is one of the clearest areas where impact might not be possible under short timelines. Cultivated meat is currently only being sold at the smallest of scales, and there currently isn't any industrial manufacturing anywhere in the world. There's an upper limit on how quickly new industrial processes like cultivated meat manufacturing can grow, and starting from such a small point, it doesn't seem like there will be enough time for cultivated meat to make any significant dent in the meat market, even assuming the underlying technical and regulatory problems can be solved. Many experts now believe the cultivated meat will take significantly longer than previously forecast, and even after it does start displacing some meat demand, it will take even more time for the supply of animal protein to adjust accordingly (see Lowering / shifting meat demand).
Legal / policy advocacy: The legal system generally moves slowly, and new laws often have a phase in period to give supply chains time to adjust. There may be specific opportunities with quick impact, like getting a legal injunction to stop ongoing cruelty, or blocking policies that would have an immediate harmful effect.
Meta: Something like filling high-impact roles, or helping existing organizations move faster / be more effective can have a speed-to-impact in proportion to the speed of the helped organization. Meta interventions around long-term movement health (e.g. cultivating a strong pipeline of early-career talent) look less good under this analysis.
Research: It can be hard to know how quickly research can pay off because we often don't know what we don't know. Research can certainly have a fast path to impact, especially if that's something it explicitly optimizes for, and if the research itself can be done quickly. That said, short timelines definitely seem to suggest moving from an "explore" posture to an "exploit" posture on the margin. Also, it may make more sense to favor scrappier, BOTEC-style research over slower, academic-style research.
Under short AI timelines, the lack of time to have an impact can be a tough pill to swallow (I, for one, found this analysis unpleasantly sobering). It's possible that the animal movement is better served by thinking of itself as placing a bet on longer timelines. Or, it might be worth thinking about if there are positive effects that might survive an AI-transformation. But insofar as we want to take short timelines seriously, and we believe that the effects of successful intervention post-transformation should be discounted to zero, it may be worth shifting focus to interventions with faster payoffs, at least on the margin.
Thanks for looking into this, I would be excited for more people to investigate interventions which make sense on short timelines.
One implication of this is: if we are (say) ten years away from farming becoming irrelevant and it takes five years to construct a farm, then delaying construction for five years is equivalent to preventing the farm's construction. So delaying tactics may be (relatively) more valuable in short-timelines worlds. (Thanks to @Jakub Stencel for this point.)
Thanks, Ben. You may be interested in this thread, @Jakub Stencel.
Realistically, I doubt farmers would want to build a farm if farming became irrelevant at the end of its construction. However, are you implying that delaying the construction of a farm for a given number of years decreases farm-years more if farming becomes irrelevant sooner? I do not think this is the case. A farm which is delayed will last for "time until farming becomes irrelevant" - "time to build the farm" - "delay" assuming this expression is shorter than its lifetime. So I agree it will last less for a shorter time until farming becomes irrelevant. However, the decrease in farm-years is equal to the delay regardless of when farming becomes irrelevant. In your example where farming becomes irrelevant in 10 years, there would be 5 farm-years without the delay, and 0 with the delay. If farming became irrelevant in 20 years, there would be 15 farm-years without the delay, and 10 with the delay. In both cases, the delay would cause a decrease of 5 farm-years. I may be missing something.
It is also worth noting that farming becoming irrelevant sooner could affect the time to build the farm, and the welfare per farm-year, which may become positive for huge economic growth. In addition, I think increasing animal farming is beneficial even if farmed animals have negative welfare due to increasing the welfare of soil nematodes, mites, and springtails more than it decreases the welfare of farmed animals.
Thanks, good question - I am assuming here that you have some positive discount rate such that you care more about reducing farming in 10 years than you do in 15.
Thanks. I still do not get it. For a construction starting in year 1, and lasting 5 years, and a delay of 5 years:
In both cases, the delay would eliminate the farming from years 6 to 10.
Yes, but the earlier 5 years are more valuable!
Given some discount rate r, the value of preventing 5 years starting at t0 is ∑t0+5t=t0rt. Here is a plot with r=0.8:
You can see that increasing values of t0 (horizontal axis) result in less valuable outcomes.
I maintain that the (net present) value of the delay does not depend on when farming becomes irrelevant. The counterfactual reduction in farming starts after the construction time of 5 years. So I think t0 in your expression should be equal to 5 years regardless of whether farming becomes irrelevant in 10 or 20 years.
Nitpick. Your formula includes 6 years from t0 to t0+5. The sum should start at t0+1 such that it only covers 5 years, which is the duration of the delay.
Fair point. Here is a notebook showing that, under 20 year timelines, saving 4 animals/year indefinitely is better than saving 5/year for 10 years, but that the order is reversed under 10 year timelines. Does this make sense now?
Thanks. Those values make sense. At the same time, I do not think they show "delaying tactics may be (relatively) more valuable in short-timelines worlds". Holding the construction time constant, the net present value (NPV) of the delay would be the same because it would counterfactually reduce farming in the same years regardless of when farming becomes irrelevant.
I think we might be talking past each other. I'm just trying to make the point in the above table: the delaying tactic is not the most effective in a long timelines world, but it is the most effective in a short timelines world. (I think you agree?)
Thanks for being so patient! I understand what you mean now. You agree the cost-effectiveness of the delay is the same in both scenarios, but are pointing out that the difference between the cost-effectiveness of the delay and that of other tactics decreases if farming becomes irrelevant sooner. I got confused because whether tactics involve delays or not is not what really matters for how their cost-effectiveness is affected by the time when farming becomes irrelevant. What matters is that their (counterfactual) effects materialise soon such that they are not heavily discounted. An intervention delaying the construction of a farm for a super long time would be a delaying tactic, but short timelines would decrease the vast majority of its value. In contrast, buying beef does not involve any delays, but arguably helps soil animals via increasing agricultural land for a few years after the beef is bought, so it would not be affected by short timelines.