How AI Is Affecting Farmed Aquatic Animals. Part 3: Welfare Impacts

AI in Farmed Aquaculture Series

Artificial intelligence (AI) introduces new capabilities to animal agriculture that could alter production methods, economic structures, and animal welfare outcomes. Responding strategically requires an understanding of how quickly such changes will unfold, whether they will benefit or harm animal welfare, and what interventions will remain relevant. In this three-part series, we take a close look at how AI will be used over the next five years in aquaculture, which collectively farms hundreds of billions of animals each year for food.
Part 3: Welfare Impacts
This report analyzes the potential animal welfare impacts of AI innovation and deployment in aquaculture, examining the direct effects of AI tools on individual farmed aquatic animals, the extent to which these tools could drive farm expansion and intensification, the current level of autonomy of AI tools used in the industry, and strategic directions for animal advocates to help mitigate the risks identified. This is the final report in the series.
Key Findings From Part Three of This Series:

• We believe it is plausible that AI tools could bring direct welfare benefits that improve the individual lives of farmed aquatic animals.
• However, positive first-order gains can mask second-order consequences. Tools that deliver direct benefits can also loosen the constraints capping stocking density, farm footprint, and farming cycle length.
• Concerning tools may already be in use. ~64% of the 91 companies in our database manufacture at least one product we would rate medium or high concern for driving expansion and/or intensification.
• Most AI tools are currently advisory. We broadly expect more autonomous tools to be of greater concern for driving expansion and/or intensification.

Executive Summary

This report analyzes the potential animal welfare impacts of AI in aquaculture. In the previous two reports, we identified which AI aquaculture products currently exist (Williamson et al., 2025) and where they are being deployed (Williamson et al., 2026). In this report, we examine how these products could affect farmed aquatic animal welfare, both directly and through their effects on farm expansion and intensification. To conduct our examination, we drew on the academic literature, our existing database of AI-aquaculture companies, and interviews with experts to compile and assess 39 AI example use cases and their welfare implications.

• We separate first-order effects—the direct effect of a tool on an animal, such as earlier disease detection or reduced handling stress—from second-order effects, which follow from how a tool changes farm operations and incentives.
• Direct welfare effects: We believe AI could bring some direct welfare benefits to farmed aquatic animals, but not enough to make their lives net-positive.
• Expansion and intensification: We believe AI could enable the expansion and intensification of aquaculture, which could increase the number of animals subject to suffering. We think there are three main pathways:
  • Enabling higher stocking densities
  • Increasing the number, size, types, or locations of farms, or the range of species farmed
  • Shortening farming cycles
• Potential risk landscape: We gave risk ratings to example AI use cases, depending on the likelihood they could enable negative second-order effects. We found that concerning AI-enabled technologies are already common among companies in our database—~64% of the 91 companies manufacture at least one product we would rate medium or high concern for driving expansion and/or intensification.
• Breadth of concerning technologies: Tools that we think could drive expansion or intensification span multiple product functions—including feed optimization, disease detection and prediction, breeding, and husbandry-parameter optimization.
• Autonomy: Most AI-enabled technologies produced currently play an advisory role in management or husbandry decisions. We expect more highly autonomous tools to broadly be of greater concern, in part due to their potential to enable production in remote, exposed, and offshore sites.
  • Of the 91 companies in our database, we estimate ~70% produce tools that are advisory only, ~21% narrowly autonomous, and ~9% integrally autonomous,^[1] with autonomy lowest in the health and disease product function.
• Strategic directions: We generate some ideas for advocates and funders to help mitigate the risks identified in this report. More data is required to understand specific impacts, but taking a precautionary approach to AI aquaculture tools now could help prevent aggregate welfare from worsening until we are better able to understand the full effects of AI on aquatic animals.
• Our analysis comes with several caveats:
  • Our risk ratings are subjective, informed judgements rather than measurements or estimations. Ratings are not adjusted for species scale, geography, or likelihood of adoption.
  • We do not attempt to rigorously trade off between first- and second-order welfare effects.
  • Data on the welfare effects of AI tools used on farms is scarce, so much of our discussion concerns theoretical impacts rather than well-evidenced effects.
  • See Ways That We Could Be Wrong for more detail.

Background

This report is the third in a three-part series examining how AI is affecting farmed aquatic animals. In Part 1 of this series, we analyzed the current state of AI innovation in aquaculture and found 91 companies with AI-enabled technologies that have direct implications for farmed animal welfare. We found that innovation is concentrated primarily in the US and Norway, predominantly targets high-value species such as salmon and shrimp, and most commonly addresses stock and growth management. In Part 2 of this series, we examined the deployment of these AI-enabled technologies, and found deployment evidence across 71 countries for 66 aquaculture-specialist companies, concentrated in Europe, Latin America, and Southeast Asia, with Norway and Scotland the largest net exporters of aquaculture AI, and salmon and shrimp having the broadest deployment footprints. You can see the database here.^[2] Part 3 lays out the possible welfare implications for aquatic animals from AI innovation and deployment in aquaculture.

Scope

This report draws on three sources of evidence:

1. The academic literature on the welfare impacts of AI-relevant aquaculture practices
2. The database of AI-enabled technologies and deployments built in Parts 1 and 2
3. Interviews with three experts with direct experience of AI in aquaculture

Using these sources, we compiled a list of 39 example use cases, analyzed their potential intended and unintended consequences, and assigned each a risk category. Some examples with different broad product functions are discussed in this report. We then mapped these categories onto the tools in our database to assess how concerning the AI already being produced and deployed is. We also classified the level of autonomy of AI technologies from the 91 companies identified in Part 1, using Claude to make an initial categorization from product descriptions that we then manually checked for 30% of the sample.
See the Appendix for more methodological detail.

Results

The welfare impacts of AI-enabled technologies on farmed aquatic animals can be understood in terms of their direct (first-order) effects on the farmed animal and their downstream (second-order) effects that could follow from changes to farm operations or incentives once a tool is adopted. In other words, a first-order impact makes the life of an average animal better or worse; we are also concerned about second-order impacts, like those that could change the number of animals in the system (see Box 1):

AI Tools Can Have Unintended Consequences

There are myriad examples of how AI-enabled tools could offer first-order welfare benefits: behavioral monitoring supports earlier disease detection; water-quality monitoring lets farmers correct poor conditions before they cause suffering; computer vision enables growth estimation that would otherwise expose animals to handling stress; and continuous automated coverage could catch deterioration that periodic human checks could miss. However, we believe these tools are double-edged: tools that support earlier intervention could have the effect of raising stocking densities, pushing growth rates toward their biological ceiling, or cutting back the buffers built into husbandry parameters such as water-quality control (Ferreira et al., 2026).
We assume farmers’ key incentive is increased productivity, as in agriculture generally (McKinsey & Company, 2024), and, without strong governance, ethical safeguards, and inclusive institutional arrangements, we are concerned that welfare considerations will be overlooked as AI is used to intensify and expand production rather than protect welfare (Ferreira et al., 2026). This would mirror what has been found for precision livestock farming more broadly: systems that reach the market tend to be “those that focus on production efficiency and farmer quality of life,” rather than on welfare (Tuyttens et al., 2022).^[3]
We believe it is plausible that AI tools have direct welfare benefits that could improve the individual lives of farmed aquatic animals. This could be a benefit that applies to only a subset of individuals on a farm—for example, weaker animals who would otherwise get less access to food—or a benefit that applies to all individuals on a farm, such as an improvement to shared water quality. We should also note that a welfare improvement does not mean an absence of suffering: a lice remover could decrease the number of lice a salmon has without removing them entirely. However, to understand whether AI is, in general, net-positive or net-negative for farmed aquatic animals, we need to understand the implications for their aggregate welfare: do the potential improvements to the quality of lives of farmed animals from AI justify the potential second-order effects, such as an increase in farmed animal numbers?
It is possible that an increase in the number of farmed animals does not result in more suffering overall (i.e., if positive experiences outweigh negative ones on farms as a result of the AI-facilitated welfare improvements).^[4] Nevertheless, advocates, including those who have welfare or suffering as their primary focus, may well have reasons to disfavor system-level changes that result in more animals being farmed than under the baseline trajectory. We take this second-order effect seriously, and the concerns we raise below center on it as the core reason advocates should examine AI implementation in aquaculture from angles beyond its directly anticipated impacts. We believe that, on balance, suffering outweighs positive experiences in aquatic farmed animal systems today. We do not believe there is currently enough evidence to suggest that direct welfare benefits that animals could experience from AI tools will swing their lives from net-negative to net-positive. Therefore, we believe that a possible effect of AI in aquaculture will be an increase in the number of farmed animal lives subject to at least some suffering. While there could be positive second-order effects, like fewer animals going into hatcheries due to more precise counting, we worry the negative second-order effect of increasing farmed animal numbers will far outweigh the positive. For this reason, we focus the scope of this report on the potential negative second-order effects of AI tools in aquaculture.

Positive First-Order Effects Can Mask Negative, Second-Order Consequences

We see three main channels through which the second-order effects of AI technologies could negatively impact animal welfare:

You could hypothetically think that if bigger, more efficient farms expand and outcompete smaller ones, this would simply displace production from smaller farms rather than increase the total number of farmed animals. However, this view requires confidence that the increase in production from the expanding farms does not exceed the decrease from the farms they displace—i.e., that the net effect on total farmed animal numbers is zero. We do not believe that demand for farmed animal production has reached its limit; the Food and Agriculture Organization (FAO) predicts that global aquaculture production will rise by 16% between 2024 and 2034 (FAO, 2026, p. 177). Therefore, we believe that if efficiencies enable producers to farm more animals, they will do so, rather than simply reallocating existing production.

We Have Reason to Be Concerned About AI Technologies for Aquaculture

Drawing on the academic literature (see Appendix for more details), our existing database, and interviews with experts, we compiled a list of 39 example use cases for tools in aquaculture.^[5] This list reflects our best effort to identify current and emerging use cases, though others likely exist. Further details can be found as a second tab in our database.
For our list of 39 example use cases, we evaluated how likely each is to increase the number of aquatic animals being farmed (via the routes listed in Box 2) and assigned a subjective low, medium, or high risk rating to each. The methodology and breakdown can be found in the Appendix. These risk levels are not adjusted for scale or geography: a technology that could intensify shrimp farming is treated as no more or less concerning than one intensifying salmon farming, even though far more shrimp than salmon are farmed, and regardless of how likely that technology is to actually be adopted in the countries where the target species is farmed. For more information, see Ways That We Could Be Wrong.
In some cases, we generalized from a specific documented tool to a broader category (e.g., AI-driven genetic optimization of mating combinations → AI-optimized breeding tool). In others, we logged a tool that could be broken into subcomponents (e.g., cannibalism and aggression detection → cannibalism detection and aggression detection, independently). As a result, the overall risk distribution could shift if more technologies were added or broken down further, and the numbers we quote should be interpreted more as informed judgments rather than concrete numbers.
Of the current AI use cases captured, we estimate ~50% are low risk, ~20% high risk, and ~30% mid-range. We intend these figures to be interpreted such that if a new tool were to be developed today, these are the subjective probabilities we would assign to the tool having a certain risk level. In general, tools that could increase the number of farmed aquatic animals through multiple channels—stocking density, farm expansion, and shortened farming cycles—were rated as higher risk than tools that could increase animal numbers through only one channel. This assessment reflects not just that a tool enabling all three is more concerning in itself, but that the probability of an increase in farmed animal numbers rises as producers have more available options through which to expand. We expect these numbers to increase in the mid- and high- risk categories in the near future, as AI autonomy, demand for oversight in remote locations, and AI capabilities are likely to become more developed.
It is not clear to us that there are any AI tools that have no possible negative second-order effects. At minimum, we would expect most tools to generate some financial saving or productivity gain that could be reinvested into farm expansion, but we expect tools to do this to differing degrees.
It should be noted that, in this report, we do not attempt to make the trade-off between positive and negative first- and second-order effects. Given that studies on the implications of AI in aquaculture are still sparse, we do not judge it possible at this time to make a fully informed judgment about whether first-order wins outweigh second-order risks. Furthermore, we believe that evaluating such a trade-off incorrectly could encourage advocates to make poor choices that could be avoided once we have better information.
In the next section we will present the example of AI-enabled removal of sea lice from salmon, as an example of how seemingly positive first-order impacts can result in negative second-order effects.

Lice Laser Delousing as a Case Study of First-Order Effects Masking Second-Order Consequences

In 2023, sea lice cost the Norwegian salmon industry up to 18 billion NOK (Jensen, 2024, as cited in Worm et al., 2026)—approximately 17% of the value of grow-out production that year.^[6] A separate, earlier estimate modeling Norwegian farm data put lice damages at up to 13% of farm revenue (Abolofia et al., 2017). As such, lice directly impact the profitability of salmonid farming, and the industry is highly incentivized to tackle them.
Detecting and removing lice early could alleviate animal suffering. However, because lice act as a constraint on stocking density and farm expansion, removing lice could also increase the number of farmed salmon overall. Below, we detail a non-exhaustive selection of how this might happen in practice.
AI-Enabled Laser Delousing
First-order welfare impacts:

• Sea lice cause direct welfare harm to salmon: they graze on the fish’s skin and underlying tissue, causing wounds, stress, reduced appetite and growth, and greater susceptibility to secondary infection (Abolofia et al., 2017). The treatments for sea lice infestations are themselves also harmful to salmon, causing panic behaviors, injuries, and sometimes mortality (Oliveira et al., 2021; Moulange & Shah, 2025). AI-enabled laser delousing could offer a gentler alternative to chemical and mechanical treatments (Planellas et al., 2026); it could halve the weekly probability of needing reactive treatments, such as harsh mechanical flushing or medicinal baths, and could cut reliance on cleaner fish—certain species of fish, such as wrasse and lumpfish, that remove external parasites from other fish—by 65% (Worm et al., 2026).^[7]

Second-order welfare impacts:

• Stocking density:
  • Lice transmission increases with stocking density (Jansen et al., 2012), so lice are one factor that can act as a ceiling on how densely salmon can be farmed (alongside other factors like regulatory or voluntary caps on density, mortality constraints, environmental concerns). Alleviating pressure on that ceiling, as AI-laser delousing has the potential to do, could enable producers to run higher-density operations and manage lice reactively rather than avoiding them through lower density.
• Farming footprint:
  • Norway’s regulatory Traffic Light System scores each production zone based on modeled lice-induced mortality of wild salmon, determining whether farms in that zone may expand, must freeze production, or are required to cut production (Ministry of Trade, Industry, and Fisheries, 2017). If lice are easier to remove at the farms within a zone, that zone may be more likely to qualify for growth rather than a freeze or reduction.
  • More stable and profitable business operations would, in turn, enable potential farm expansion.

Concerns by Product Function

We have reason to believe that high-risk technology could come from several distinct product functions. We will now discuss, in turn, the tools in each product function category (the management of feed and feed optimization; health and disease; stock and growth; and water quality)^[8] that we believe are most currently concerning for increasing the number of farmed aquatic animals. All examples include a non-exhaustive list of potential impacts. For more information, see Ways That We Could Be Wrong.

AI-Driven Precision Feeding and Real-Time Satiation Detection

AI-driven precision feeding and real-time satiation detection tools sit under our feed optimization product category.
As we determined during the first report (Williamson et al., 2025) of this series, feed purchase is consistently the highest operating cost in an aquaculture facility. Commercial studies and research pilots report that AI-feeding can reduce feed costs by 15–30% compared to feeders operating on fixed or manually adjusted schedules (Ferreira et al., 2026),^[9] meaning more efficient feeding is highly commercially incentivized.
For shrimp specifically, Expert 3 (see Expert Profiles for more detail) pointed to two further incentives: a) it is difficult for producers to reach the same production cost without AI-feeders,^[10] and b) AI-feeding leads to a more uniform size across shrimp, in turn increasing the market value for the batch, since shrimp are priced by size consistency (García-Ballesteros et al., 2021, p. 2, citing Asche et al., 2011; FAO, 2005; Balaban et al., 2008; Zhang et al., 2014).
While we cannot find data on AI-driven feeders specifically, we can look at the growth improvements due to automated feeding over manual feeding as a proxy. Studies lean toward automatic feeders outperforming manual feeding, but other farming factors—such as feeder density, commercial vs. research setting, and feed ration size—vary the results. Automated feeders appear to improve productivity gains over manual ones, but it is unclear whether AI-feeders would further improve upon automatic feeders. For more information on current feeder studies, see the Appendix.
Potential gains from AI-feeders being only an incremental improvement on automatic feeders is one reason we are less worried about AI-enabled feeders than some other applications in this report—even though we still rate it as high risk. Another is that we are uncertain about how much of the cost reductions (from reduced feed use) and the increased growth rates are attributable to feeder choice specifically rather than, for example, genetic selection. Furthermore, feed conversion and genetic selection may not even be separable: how efficiently an animal converts feed into growth is partly genetic, and a genetic line bred to grow fast on one feed composition does not necessarily grow fast on another (Dai et al., 2024). Additionally, Expert 3 suggested that there may be limited room to improve feed conversion ratios in shrimp as they may be approaching their technical limit (but we could not independently verify this). That said, we expect productivity gains from feed optimization to be possible in many species, but there may be biological limits to how far this optimization can go. Therefore, we are not certain how much potential second-order consequences could be attributed to AI-enabled feeders specifically.
That said, it still seems possible that improving feed efficiency could improve welfare by preventing under- and overfeeding and bettering water quality, but could also weaken constraints on stocking density and unlock profits that could be reinvested into farm expansion.
First-order welfare impacts of AI aquaculture tools:

• When access to food is limited, fish or shrimp lower in the social hierarchy get less access to food and are repeatedly threatened or attacked by dominant fish who serve as a source of chronic stress for them (Martins et al., 2012). AI-feeding based on appetite detection could increase the opportunity for all animals to access food—rather than simply the biggest or strongest animals—reducing the risk of competition, starvation, and starvation-induced stress to weaker animals.
• AI-enabled satiation detection could help prevent overfeeding (Georgopoulou et al., 2024), and in turn improve water quality.

Second-order welfare impacts of AI aquaculture tools:

• Stocking densities:
  • More efficient feeding reduces the amount of ammonia and organic waste load entering a pond (Boyd & McNevin, 2024), thereby improving water quality. As water quality degrades with increased stocking density (Xu et al., 2025), improving it could relax some constraints on stocking density.
• Farming footprint:
  • Economic savings from more efficient feeding practices could be reinvested into farm scale-up.
• Farming cycle:
  • More efficient feeding can enable quicker growth, accelerating the rate at which animals reach production weight, which could in turn increase production turnover. Expert 3 attributed significant accelerating growth rates and a decreasing feed conversion ratio (FCR) in shrimp in the last six years to more efficient feeding practices; however, the aquaculture literature cautions that such gains are difficult to attribute to feeding practices alone, since concurrent genetic selection—which can raise growth rate by up to 10% per generation (Zeigler, 2017)—and improvements in culture conditions plausibly contribute to the same trend (Davis et al., 2016).

AI-Enabled Disease Detection and Prediction

AI-enabled disease detection and prediction tools sit under our health and disease product category.
Disease mortality costs the aquaculture industry an estimated $6 billion annually (FAO as cited in World Bank Group, 2014, p. 1). In Norwegian salmon farming, gill disease and delousing injuries are the top two causes of mortality and poor welfare (Moldal et al., 2025, pp. 10, 25), while in shrimp, disease has historically caused major production collapses (Shinn et al., 2018). Disease thus directly affects profitability, strongly incentivizing detection and prediction.
We view AI tools that predict disease risk before onset as more powerful than those that detect disease once present. Detection tools identify an active infection and enable a reactive response; prediction tools forecast outbreak likelihood in advance, giving farmers a window to intervene before losses occur.
Expert 1 argued that disease currently acts as a cap on salmon production, and that to truly alleviate the disease burden, the industry will likely require veterinary solutions rather than advances in monitoring technology. We believe detection tools carry a mildly weaker risk than prediction tools, since they only flag problems after onset, whereas prediction would enable the application of preventative tools as far as they are available, potentially making infectious disease less of a ceiling for stocking density and enabling farmers to avoid the economic losses associated with disease.
AI-generated disease detection and prediction could be used to prevent or reduce the time animals spend suffering, but could simultaneously enable higher density and bigger operations by enabling higher stocking densities and by reducing mortality-related monetary losses that could be reinvested into farm expansion.
First-order welfare impacts:

• AI-enabled disease detection and prediction tools reduce diagnostic delays, support precise medical treatments, and limit the use of precautionary antibiotics (Ferreira et al., 2026).^[11]
• Earlier detection of disease reduces the amount of time animals exposed to disease spend suffering and reduces further propagation (Bonnichsen et al., 2024).
• Predictive health systems may improve survival outcomes (Öz et al., 2025).^[12]

Second-order welfare impacts:

• Stocking densities:
  • Prediction specific: Producers could run higher-risk and/or higher-density operations, using disease predictions as an operational safety net and arguing that strict stocking density limits are outdated because disease will be flagged before it occurs.
• Farming footprint:
  • Disease detection and prediction could better enable producers to turn potential disease deaths into scheduled harvests. This could give rise to mortality statistics that understate the actual disease burden that producers could use to argue for higher stocking density or farm expansion. We expect this risk to be greater with prediction than detection: detection can only convert a case into a scheduled harvest if the disease spreads slowly enough to leave time to slaughter the still-healthy animals before they too become infected.
  • If monitoring of any kind reduces mortality-related losses, producers will be financially better off and could reinvest the savings into farm expansion.
• Farming cycle:
  • Expert 2 stated that health and disease AI monitoring tools are currently used reactively rather than preventively, predominantly to time slaughter and avoid mass mortality events and biomass loss. Normalizing early culling as a production-optimization tool in this way could shorten farming cycles further and increase turnover. This was supported by Expert 1, who suggested that due to the short grow-out cycle in shrimp, AI was informing whether to harvest immediately or continue growing, rather than being used to treat disease.

AI-Optimized Breeding

AI-optimized breeding tools sit under our stock and growth category.
AI may enable a form of intensification that is harder to regulate or detect, by pushing biological limits rather than visible stocking density.
Selective breeding is an established practice in aquaculture (Sonesson et al., 2023), and there is already precedent for the kind of gains AI-optimized breeding could accelerate, both within aquaculture and outside it.
Specific Pathogen-Free (SPF) P. vannamei Shrimp
SPF P. vannamei were developed to produce broodstock free of specific pathogens, in order to reduce mortality and increase production (Alday-Sanz et al., 2018). SPF stock was introduced into commercial production in the US in 1992, and subsequently spread to Asia (Wyban, 2009). In Thailand, the replacement of P. monodon with SPF P. vannamei meant production systems could operate at 300% higher stocking density and with 280% higher profit per square metre (Wyban, 2009, Table 3). Due to these gains, P. vannamei adoption spread rapidly, and Thai production climbed from 10% of total farmed-shrimp output in 2000 to 75% by 2007 (Wyban, 2009, p. 20).
Commercial broiler chickens
While not an aquatic animal, we think that broiler chickens are also a useful case study here. Between 1957 and 2005, selective breeding increased growth rate in commercial broiler chickens by over 400%, and cut feed conversion by 50% (Zuidhof et al., 2014), at a cost of skeletal defects (Rath et al., 2000; Lilburn, 1994), metabolic disorders (Scheele, 1997), heart conditions (Olkowski, 2007), and altered immune function (Cheema et al., 2003).
There are an estimated 400 (Franks et al., 2021) to 700 (Sonesson et al., 2023) farmed aquatic species—considerably more than the 38 farmed terrestrial animal species—with high genetic diversity within populations that could enable considerable commercial gains through selective breeding (Sonesson et al., 2023). If AI is used strictly to identify and breed for maximum growth or yield without balancing for welfare indicators, it could lead to generations of fish that suffer from metabolic disorders or skeletal deformities due to fast growth, as well as other negative welfare traits (Saraiva et al., 2018)—echoing the same trade-off already documented in broiler chickens. We are concerned that AI-assisted breeding could compress a similar trajectory and its welfare costs, across many species, into a fraction of the time it took in chickens.
Whether AI-assisted breeding can push biological limits significantly further or faster than conventional selective breeding that is already underway remains an open question.
First-order welfare impacts:

• Breeding could be optimized for welfare traits, consequently reducing suffering by reducing susceptibility to diseases and increasing resistance to wounds.^[13]
• Disease-resistant strains could, in principle, reduce farmed animals’ exposure to harmful treatments such as antiparasitic baths and to suffering from ill health.
• Breeding optimized for productivity gains, without balancing welfare indicators, could produce its own direct welfare harms as a side effect (mirroring the broiler chicken trajectory described above) and potentially compressed into far less time by AI.

Second-order welfare impacts:

• Stocking densities:
  • Selecting for calmer and more physically tolerant animals could conceal poor mental welfare rather than fix it. AI-accelerated breeding could yield aquatic animals that physically better tolerate poorer conditions, resulting in lower mortality rates and increased stocking densities, despite continuing to suffer.
• Farming footprint:
  • Breeding for disease resistance and environmental tolerance could reduce the economic risk of scaling up existing farms, encouraging producers to expand into larger or more intensive operations.
  • AI could analyze genotyping data from new species to identify links between genetic markers and observable, commercially favorable traits. This could make it possible, for example, to identify new species that could be commercially farmed.
• Farming cycle:
  • Breeding for faster growth accelerates the rate at which animals reach production weight, which could in turn increase production turnover.

AI Optimization of Husbandry Parameters

AI optimization of husbandry parameter tools sit predominantly in our water quality category, although they often act cross-category. We define AI optimization of husbandry parameters as AI tools that integrate multiple husbandry variables (e.g., feed, temperature, water quality) to identify parameter combinations that meet a defined objective (e.g., growth, survival) and actuate corresponding adjustments.
Species diversification has increasingly become an endorsed strategy for the growth and resilience of the aquaculture sector (Cai et al., 2023), as has expansion into offshore waters, which has gained increased attention in recent years as a major avenue for the growth of the aquaculture sector (Krause et al., 2024). AI-enabled optimization of husbandry parameters could be a key enabler for both: farming currently unfarmed or lesser-farmed species and farming in harsher, less accessible environments.
Moreover, whether AI-optimized husbandry parameters improve or worsen welfare in practice will likely depend on which proxy the optimization targets. Tools that could maintain conditions within welfare-relevant ranges could just as easily be optimized for growth or yield, with welfare maintained only incidentally, where it does not conflict with productivity.
First-order welfare impacts:

• AI-actuated environmental control could maintain farm conditions within welfare-relevant ranges faster than a human response could allow, particularly on isolated, offshore, or low-labor farms.
• Better environmental control enabled by AI could allow for more environmental enrichment to be used. For example, having greater control over water quality could enable producers to add substrates, plants, and other variations into the farming environment.

Second-order welfare impacts:

• Stocking densities:
  • AI-optimized husbandry parameters could enable more intensive stocking of species currently farmed under extensive or semi-intensive systems, resulting in more individuals being farmed, or individuals being farmed in worse conditions (such as in more crowded spaces), or both. As an example, see this intensive indoor vertical marron farm (Kennedy, 2022; Aquatic AI, 2026) , using AI and robotics to automate care and feeding, correlate growth with genetics, and monitor health.
• Farming footprint:
  • Challenges facing farms established in difficult environmental conditions could be alleviated, enabling farm expansion into remote and offshore locations—where welfare oversight is harder to implement—as technological developments such as remote monitoring and feeding are expected to reduce the costs and risks (Knapp, 2013).
• Farming cycle:
  • An AI algorithm could optimize input parameters, such as water quality and feeding regime, against a growth or FCR proxy, identifying parameter combinations that maintain or increase productivity under degraded environmental or welfare conditions.

Concerning Tools Are Already in Use

We find that tools we would consider medium or high risk already make up a large share of AI tools that are currently produced and deployed.
Taking into account all 91 companies in our database, and assigning each company the risk level of its highest-risk tool,^[14] we found that ~64% of the companies developed tools that we would describe as medium or high risk, which indicates that companies are developing products that have a possibility of expanding or intensifying aquaculture, according to our estimations and theorized impacts. We expect the percentage of medium- and high-risk tools manufactured to increase as AI tools become more widespread and autonomous.
Next, we looked at risk levels across the 66 of 91 aquaculture specialist companies for which we have deployment evidence (see the Deployment report for how we defined this). Because we defined companies by their highest risk tools, our method overestimates the true share of medium- and high-risk deployment, and correspondingly underestimates the true share of low-risk deployment: a company’s lower-risk products and functions do not appear in the total, so they cannot dilute the concentration of its higher-risk tools the way they would if we measured risk by product rather than by company. We took this approach to get a picture of the theoretical ceiling of risk today, which we think is most relevant for discussing potential future welfare impacts.
Of the 276 deployments in our database, we found ~25% (68) low risk (likely an underestimate, as per the bias above), ~44% (120) medium risk, and ~32% (88) high risk (the latter two both likely overestimates).
To test whether any individual region’s risk mix differs from this overall pattern, we compared each region’s low/medium/high split to the split we would expect if it matched the full sample. Only four regions had enough deployments to test reliably (see Regional Groupings for countries included): Norway (24); Southeast Asia (37); Latin America, excluding Chile (40); and Europe, excluding Scotland and Norway (60).^[15] Norway and Latin America (excluding Chile) closely follow the average distribution, while Southeast Asia skews further towards low risk and away from medium risk. Europe, excluding Scotland and Norway (not statistically significant p ≈ 0.06), is overrepresented in medium risk (57% vs. the 44% baseline) and underrepresented in low risk (13% vs. 25%). No region reaches statistical significance at p < 0.05. Based on our limited sample, we find no region has a significantly higher proportion of high-risk technologies deployed than others: high-risk technologies are deployed across all regions in our sample.

Autonomous AI Could Expand Where Production Can Physically Occur

We do not believe that AI technologies need to be fully autonomous to pose a welfare risk to farmed aquatic animals. However, we see a real possibility that increasingly autonomous AI could facilitate the expansion of farms into more remote, exposed, and offshore sites—enlarging the geographical footprint of production, and with it, the number of animals within the system. AI autonomy likely enables this expansion because manual labor is hard to sustain at these sites, requiring greater reliance on automation, sensors, robots, and AI to maintain both fish welfare and farming infrastructure (Føre et al., 2026; Fish Focus, 2026).
The following sections describe the current state of AI-autonomy as captured by our database.

Most AI Tools are Advisory

To understand to what extent AI is being used to make welfare-relevant decisions independently of human interaction, we classified the AI technologies from the 91 companies according to their level of autonomy.

To determine the classifications, we gave the product summary descriptions we collected when building up our database to Claude, and then asked it to use these descriptions and an enhanced online check to assign the company’s offerings into an autonomy category. We then manually checked 30% of the sample, including all those labeled integrally autonomous, and found we agreed with almost all the categorizations.
We found that the majority of tools are operating in an advisory capacity only (~70%), while the minority are integrally autonomous ( ~9%), and ~21% are narrowly autonomous.
Figure 1: Distribution of AI technology applications from 91 companies, classified as advisory, narrowly autonomous, or integrally autonomous. Category definitions are provided in Box 3.

The eight technologies classified as integrally autonomous take actions without human approval across multiple welfare-relevant variables simultaneously, most commonly involving feeding and environmental control (e.g., water quality, oxygenation, and temperature).^[16] The majority of integrally autonomous AI tools operate within specially designed, enclosed, land-based production environments—including recirculating aquaculture systems (RAS), vertical farming units, and self-contained container systems—rather than open-water or outdoor pond infrastructure.
While RAS has held a limited number of commercial successes to date (Naylor et al., 2021), some industry experts estimate that investment to date will translate into 25% of all salmon production coming from RAS facilities by 2030, up from roughly 1–2% today (van Beijnen, 2025). That said, industry analysts remain skeptical that this scale of growth is achievable: available reports largely reflect theoretical production capacities rather than realized output (van Beijnen, 2025). Since, under some estimations, this type of land-based, self-contained farm could therefore grow substantially, we expect that, in this scenario, autonomous AI tools could also become more widespread. As AI tools become more powerful, they may even make systems like RAS more viable.
We are probably most concerned about AI tools that are both very general and very autonomous, since increases in capability and autonomy are expected to most amplify AI’s impact and risk (Bengio et al., 2024).^[17] We believe this logic could also apply to narrower AI tools, like those in our database; a tool that is both more general (spanning multiple welfare-relevant variables) and autonomous (acting without human approval per action) could affect a larger share of an animal’s environment through more pathways at once and enable more production without human oversight.
Our database contains a number of tools with multiple capabilities, and we would expect these to have a larger impact than a single tool with a single capability. This holds even at a smaller scale: if the data of a single technology could be combined with the data from another, the analysis the AI is able to provide could have far greater impact on aquatic animals than from the data of one tool alone. Expert 3 illustrated this for shrimp farming: water and health data are often recorded on paper or in separate spreadsheets rather than a common platform, and this fragmentation, rather than AI capability specifically, contributes to disease-predictive patterns not being identified. For instance, certain types of algae can be linked to disease or off-flavor problems in shrimp, but the lack of data uniformity and a common platform means AI might not have access to the data to detect these patterns. Expert 3 proposed that these patterns would presumably be easier to identify if all environmental parameters and health controls were recorded in a standardized way, which we may expect producers to start doing if they want to use AI tools. Even where such patterns are found, Expert 3 noted that the response can be limited to harvesting early, since it is difficult to intervene on water conditions in open ponds—a constraint that would not apply to closed systems.
The same holds for autonomy: a tool acting without human approval per action could enable more human-independent production, extending its influence over an animal’s environment without being limited by how much a person can oversee.

AI Autonomy Is Lowest in Health and Disease

Where health and disease capabilities are present, these appear predominantly limited to monitoring and detection rather than autonomous treatment.
AI-enabled tools are increasingly able to recognize a range of health and welfare indicators—including wounds, lice, respiratory frequency, and swimming trajectory—but their main function currently lies in supporting human decision-making, not in delivering measurable welfare gains; there is not yet evidence they meaningfully reduce antimicrobial use or mortality (Ferreira et al., 2026). As far as we could determine, none of the tools marked as integrally autonomous possessed abilities to autonomously actuate against disease. The one exception used lasers to remove lice, and this we classified as narrowly autonomous.
However, autonomy is only one dimension of risk, and AI poses a threat to the welfare of farmed aquatic animals before we reach the point of fully autonomous tools and farms.
In our database, we found just two technologies that used the terminology “disease prediction” to describe their AI technology, and both had some integrally autonomous capabilities in other welfare-related functions.^[18] That said, the experts we spoke with were skeptical that disease prediction was a current tool capability. Expert 1 stated that disease prediction, which could prevent outbreaks, has not yet been achieved; for now, producers can only become more efficient at diagnosis and treatment. Expert 3 was likewise unaware of anyone using AI to predict disease in shrimp farming. Expert 2 added that responses remain reactive rather than preventive—slaughter scheduling or veterinary intervention, rather than earlier intervention—in part because the literature has not yet established how behavioral signatures translate into specific health problems on the farm.

Generalized Concerns of AI-Enabled Tools in Aquaculture

In this report, we have discussed AI-enabled technologies assuming that they work appropriately for the species and in the environments they are designed to target, but we have broader concerns about the implementation of AI tools in general. Welfare interventions do not reliably transfer across species or production systems, even when the underlying problem appears shared (Chiang, 2026). The same could be true of data used to train AI tools to recognize patterns. Training data is often lacking across species and farming contexts, which means that AI tools trained on data from limited contexts may not reliably generalize across species or systems (Ferreira et al., 2026, §5.1). In our database of AI-enabled aquaculture tools produced by 91 companies, 69 tools are applicable to fish (some target only fish types, others shrimp or shellfish alongside fish—see our first report (Williamson et al., 2025) for more details). Of those 69 tools—which cover many different product functions—six appear to target all aquatic species and 22 target fish without specifying the species. Given that tools may not reliably transfer across species or systems, we suggest taking broad applicability claims like these with caution. This caution applies even to functions like water quality monitoring, where the parameters measured—such as oxygen, ammonia, CO₂, pH, temperature, salinity—might seem universal. For example, water quality parameters should be maintained within the range that sustains normal activity and physiology for a given species, and even within a species, requirements vary by life stage (e.g., larvae, juveniles, adults) or physiological status, e.g., metamorphosis, spawning (Villalba et al., 2020, pp. 6–13). Of roughly 400 farmed aquatic species, only 25 (~7% of farmed individuals) are covered by a modest body of welfare literature (five or more publications), and 231 species have no welfare publications at all (Franks et al., 2021). Due to this species-specific data scarcity, we do not expect a tool marketed as applicable to “all aquatic species” to have been calibrated to the specific tolerance ranges that each species, let alone each life stage, requires.
Even on the right species, a tool’s output is not guaranteed to be accurate; thin training data and sometimes an absence of evidence-backed ways to confidently confirm an animal’s actual pain state can produce under-, over-, or misdiagnosis of health, disease, and condition (Ryan & Bossert, 2026, §3.1).
Furthermore, we advise caution when it comes to AI related to regulatory thresholds and reporting. On one hand, AI-generated data could provide a wealth of information that scientists could use to better understand the lives and welfare preferences of aquatic animals. This could help animal advocates lobby industry for regulatory change, both by enabling scientists to understand correlations between farm-level inputs like water quality and feed quantity, and welfare-related outputs like stress, and by giving advocates access to concrete farm-level evidence.
On the other hand, systems could create a digital illusion of safety. We are concerned that, at the most basic level, the mere claim of relying on AI-enabled monitoring could itself function as a form of welfare-washing. For example, a system marketed as objective may still reproduce bias present in its training data (Tuyttens et al., 2022). This could also occur without the underlying data ever needing to be shared, although we expect this to vary between jurisdictions. Even where data is shared, we would not rule out it being gamed by, for example, showing only subsets or skewed samples of data gathered. A more objective-seeming system could also be used to argue for less human oversight: in the agriculture sector, Dutch slaughterhouse operators have already made this case for CCTV monitoring, advocating for “more risk-based targeting and lower [human] inspection frequencies for companies that perform well” (NVWA, 2021). Though none of this is decisive, it speaks to the need to pair AI-enabled monitoring with system-specific standards or regulation to ensure that animal welfare is not hidden or overlooked in the implementation of AI.
We also expect it to be possible for producers to treat predictive capability as an operational buffer. AI tools could generate a false sense of security, leading producers to operate with less margin—e.g., farming with higher stocking densities or less well-controlled environmental parameters—on the assumption that the tool will flag trouble before a mass mortality event occurs. This could lead to the risk that an over-reliance on such tools reduces producers’ vigilance and that any false negatives produced by AI, i.e., where AI declares a parameter in order when it is not, go unnoticed. In at least one survey of agricultural farmers, they reported fearing the loss of their observation skills with dependence on precision technologies (Kling-Eveillard et al., 2020).

Current Strategic Directions Should Set Out to Mitigate Potential Second-Order Risks

Based on the current sparse research landscape, we are uncertain whether AI will be able to enable farmed aquatic animals to lead a net-positive life. However, we see enough reason for concern around negative impacts that we believe advocates should take precautions and could try to mitigate some risks now.
As far as we can determine, current AI technologies represent incremental changes to farming practices, although there are signs of more autonomous and cross-category tools being developed. As far as advocates are concerned, they should keep an eye out for more transformative shifts, such as AI in offshore farms, rather than single-product-category tools—although, as shown in the examples above, even these can come with serious concerns, especially where they alleviate critical production bottlenecks.
Most of the negatives we anticipate come from second-order effects, so we believe advocates should focus on ensuring these effects are best mitigated. We cannot say which of the second-order effects discussed—farm expansion, increased stocking density, shortened turnover cycles—is most urgent to address, nor which responses to them would be most effective.
The following suggestions are not comprehensive, and are a starting point for discussion; we cannot assert that all would work in practice or be devoid of unforeseen side effects. Many of these suggestions could target multiple actors, or be achieved through multiple routes—including certification schemes, retailers, regulators, and insurers. We have not assessed the potential cost-effectiveness, tractability, or relative priority of the following ideas. We present them only as initial ideas based on our assessment of the AI aquaculture landscape.
Make sure welfare scientists are in the conversations

• Build an understanding of aquatic animals’ preferred behaviors in consultation with welfare scientists who could conduct in-situ evaluations of the farmed population itself, preferably in low-stress settings.^[19]

Fund research on developing requirements for ensuring good welfare during AI tool design and implementation
Fund research to establish welfare standards for the design and implementation of AI tools, including but not limited to:

• Camera and sensor placement; for example, making sure that readings take a representative snapshot of on-farm conditions
• Performance on commercial farms, ensuring that tools operate as effectively there as they do in laboratory testing settings
• Mapping and understanding
  • Specific positive-welfare-state indicators that could be measured with AI
  • Welfare-specific behaviors, e.g., work linking physiological or behavioral signals to specific welfare states (Hoyo-Alvarez et al., 2026)
  • Different origins and implications of fatal and non-fatal stress on fish welfare

Ensuring welfare-specific behavior is monitored
Ensure that welfare-specific behavior is monitored by AI tools in a species-relevant way, rather than only production factors.

• Push producers to monitor more welfare-specific behaviors that they might not choose to monitor on their own.
• Push stress data to be tracked as a minimum welfare indicator, e.g., in certification schemes. Stress correlates with mortality, but not all stress is fatal. So, monitoring could ideally cover all stress—not just the mortality-linked subset that producers already have an incentive to track.
• Reward transparency itself, not just good welfare outcomes: producers should not necessarily be penalized for sharing data that reveals poor welfare, since accurate data is more valuable to advocates and regulators than unrepresentative data.
• If a producer is using AI, they could be required by, for example, certification schemes, to also monitor indicators of non-lethal suffering (where species-specific metrics are available), and/or they could be required to share their data with auditors.

Apply welfare scrutiny to AI-assisted breeding

• Advocates could require disclosure of what trait is actually being selected for whenever an AI tool is used to accelerate or scale up selective breeding, since traits that make animals more tolerant of poor conditions, or that are pursued mainly for productivity, can trade off with welfare directly.
• Certification schemes and regulators could require that any welfare claim about an AI-assisted breeding program—whether framed as a welfare intervention or not—be verified across the animal’s full life cycle, rather than taken on the strength of a single indicator at a single life stage.

Push for farming designs with less barren environments

• As farms automate, advocates could push for AI-enabled environmental control to be used for enrichment—such as lighting, substrates, and bubble curtains—rather than achieving environmental control by using barren settings. Camera-based monitoring systems typically need a clear, uniform background to accurately detect and track animals, which tends to favor barren, sterile housing (Tuyttens et al., 2022, p. 7). Without deliberate design choices, AI-enabled environmental control could risk reinforcing rather than improving the bareness of environments like recirculating aquaculture systems (RAS).
• To the extent that tools cannot easily generalize between farming environments, advocates should encourage tool designers to test tools in enriched environments, so as to avoid producers being locked into farming in barren environments in the future simply because those are the conditions in which AI tools were validated.

Improving non-AI-specific regulations to set the groundwork for AI-specific regulations

• Extend existing governance. For example,
  • Fold AI-driven delousing into Norway’s existing Traffic Light zone scoring system, such as by ensuring farms are only allowed to expand when salmon are not repeatedly suffering from lice, rather than only when delousing is frequent enough that lice rates appear low but salmon may still be suffering.
  • Where it exists, legislation could require that AI tools, where proven safe and better for welfare than other alternative options, should be used as a first-choice treatment (though better information around the second-order effects of better delousing should be examined first).
• Improve the quality of stocking density regulations—for example, basing them on peak rather than average stocking densities, and excluding theoretical water volume that animals would not inhabit (such as deep water) from the calculations. We remain uncertain what stocking density is optimal for individual welfare, but at the population level, stocking density regulations could still serve to prevent the overall number of farmed animals from significantly increasing.
• Build, or build upon, comprehensive, binding welfare standards used in aquatic farming to cover AI applications.

Preventing major intensification

• Introduce regulations on major stocking density increases. In the absence of species-specific, good quality data, advocates could try to tie AI adoption to relative, not absolute, stocking density limits—e.g., producers should not raise stocking densities more than ~10% above a pre-AI-adoption baseline. This depends on the underlying stocking density regulation being well-specified (as mentioned above) and stocking density being well-measured prior to AI adoption.
• AI systems should be required to give ethical and welfare considerations significant, explicit weight relative to economic factors, through a structured, standardized process before acting on a prediction. A systematic review of 38 aquaculture AI studies found only three had integrated any standardized risk framework (e.g., ISO 31000) at all, and these were research studies rather than commercially deployed tools (Gkikas et al., 2026).

Using governance to embed AI-specific regulations

• Welfare proxies could be built directly into the AI system’s optimization target rather than monitored as a separate output; embedding welfare indicators at the design stage keeps welfare from being relegated to a side effect as the system optimizes purely for productivity metrics such as growth or yield (Adekoya, 2026).

Lobby tool designers and producers for the release of key data

• Secure public and regulator access to welfare data before producer-owned AI data streams become the norm, e.g., through mandatory data-sharing clauses in aquaculture certification schemes.
• Push for disciplined, purpose-bound data-sharing models so that welfare-monitoring data collected for early intervention cannot be freely repurposed toward optimization ends without oversight.
• Use producer data and independent trials to confirm whether tools actually deliver the productivity gains they are anticipated to. For example, if AI-enabled feeders do not meaningfully move growth, FCR, yield, or other productivity metrics, the tool’s expected threat level, and the urgency for advocates to act, decrease accordingly.

Further concerns and suggested directions that may hold relevance to the AI in aquaculture context can be found in precision livestock farming for animal welfare literature (e.g., Tuyttens et al., 2022).

Conclusions

This third and final installment of How AI Is Affecting Farmed Aquatic Animals examines the potential welfare implications of the AI innovation and deployment mapped in Parts 1 and 2. We distinguish between the first-order, direct effects of AI tools on individual animals and their second-order, indirect effects such as increasing the number of animals being raised in the farming system. We find that while AI-enabled technologies can plausibly offer direct welfare benefits, we do not yet believe these are likely to be sufficient to make farmed aquatic animals’ lives net-positive. Instead, we worry AI is more likely to enable the expansion and intensification of aquaculture—through higher stocking densities, farm expansion, and shortened farming cycles—thereby increasing the number of animals subject to suffering.
These risks are not confined to a single type of technology and span multiple product functions. Furthermore, AI use cases we would rate as medium or high risk could already be widespread among currently deployed tools. Most AI tools today remain advisory rather than autonomous, but we expect a shift towards greater autonomy in the near future, potentially enabling the expansion of farming into more remote, exposed, and offshore locations where human oversight is harder to sustain.
Given these findings, we think advocates, funders, and welfare scientists should be concerned about the potential spread of AI tools in aquaculture. More data is required to understand specific impacts, but taking a precautionary approach to AI-aquaculture tools now could help prevent aggregate welfare from worsening until we are better able to understand the full effects of AI on aquatic animals.

Acknowledgements

This report is a project of Rethink Priorities (RP)—a think-and-do tank dedicated to informing decisions made by high-impact organizations and funders across various cause areas. Sophie Williamson managed the database, did the research, conducted the expert interviews, and wrote the report. Hannah Moulange managed the project and reviewed the research. William McAuliffe oversaw the early stages of the project, while Alyse Spiehler oversaw the latter stages. Thanks to Kevin Xia, Natasha Boyland, and Aaron Boddy, who each gave input on the draft, and to Rey Edison for input on topics related to AI-optimized breeding. Thanks to Shane Coburn for copyediting, to Thais Jacomassi for bibliography support, and to Elisa Autric for publishing the report online and assisting with dissemination. This report was produced by Rethink Priorities between June and August 2026. The project was supported by a Movement Grant from Animal Charity Evaluators (ACE). The views expressed are those of the authors and do not necessarily reflect those of Animal Charity Evaluators.

Parts of this content were prepared with the assistance of AI tools (such as Claude and Gemini), which we use to improve efficiency and readability. All outputs are supervised, reviewed, and fact-checked by Rethink Priorities’ staff, who remain responsible for the final content.

If you are interested in RP’s work, please visit our research database and subscribe to our newsletter.

Appendix

Methodology for the Literature Review

We spent approximately 4.5 hours searching and screening peer-reviewed literature on the welfare impacts of AI in aquaculture using Claude (Sonnet 4.6), using its web search tool to query the open web, including Google Scholar, PubMed/PMC, ScienceDirect, Springer, Wiley, Frontiers, MDPI, arXiv, and Preprints.org. We stopped when successive queries across new angles consistently returned only papers already identified. The titles, abstracts, and authors of all papers found were checked manually to confirm the accuracy and relevance of the papers found by Claude. This research was then supplemented by one hour of using the keywords “AI”/”Artificial Intelligence,” “Welfare” and “Aquaculture” on Wiley and ScienceDirect, manually. In total, we discovered 36 articles (30+6) for which we could access the full text.^[20]

Several limitations apply. The search was restricted to English-language literature and to sources accessible without subscription. The search did not cover specialist databases (e.g., ASFA, CAB Abstracts), non-English-language literature, and conference proceedings not indexed in the above sources, nor papers behind hard paywalls. We also excluded literature on AI welfare impacts in slaughter contexts. A structural feature of the literature itself further constrained retrieval; because welfare is rarely framed as a measured dependent variable in AI aquaculture papers, searches on welfare-outcome terms tended to retrieve the same conceptual reviews repeatedly rather than surfacing new empirical work.
In full, we read four papers that we believed to be highly welfare relevant (Planellas et al., 2026; Fitzgerald et al., 2025; Dawkins, 2025; Ferreira et al., 2026). Then, when looking for mentions of AI technologies and their welfare implications, we used Claude and NotebookLM to scan for relevant information from that selection of papers, which we then read manually. Where information from papers has been used to support our reasoning, we have cited the literature in-line in this document.

Methodology and Evaluation of AI Tool Risk Level

For the 39 example use cases identified from the literature (see the Welfare Impacts tab of the database), we made an informed guess on how likely each one is to be able to expand or intensify aquatic animal farming based on:

• How likely the application is to increase stocking density (likely = 2, possibly = 1, unlikely = 0)
• How likely the application is to increase the number or size of farms (likely = 2, possibly = 1, unlikely = 0)
• How likely the application is to increase the total number of animals farmed via quicker turnover, such as shortening the farming cycle—e.g., quicker growth leading to a younger slaughter age and the next cohort being brought into the system sooner (likely = 2, possibly = 1, unlikely = 0)
• Whether the application is highly commercially incentivized (yes = 1, no = 0)
  • We add this as an additional risk criterion on the assumption that a tool only becomes particularly concerning if producers wish to use it.

For each application, we totaled these together to give us a score out of 7, a concern rating, with 0 the least concerning and 7 the most concerning. Of these, we grouped them such that:

• Concern rating 1–2 corresponds to the category low risk
• Concern rating 3–5 corresponds to the category medium risk
• Concern rating 6–7 corresponds to the category high risk

Figure 2 shows one researcher’s distribution of concern ratings across the 39 example use cases we identified, informing the stance we laid out in the main report. This detail is included for informational purposes only.
Figure 2: Distribution of our concern ratings across the 39 example use cases we identified. These evaluations are highly subjective, likely to change, and based on only a limited review of the landscape. We group tools with ratings 1–2 as low risk, 3–5 as mid-risk, and 6–7 as high risk.
This distribution comes with many caveats: the technology list reflects only a limited, subjective review of what we could identify, rather than an exhaustive survey, and the concern ratings themselves have not been independently verified. However, it should give a general sense of how concerned we think animal advocates should generally be about AI in aquaculture.
The reason we use low, medium, and high risk levels in the main body of the report rather than Concern Rating is that we believe the granularity of Concern Rating projects more confidence in our evaluations than is warranted, as these are ultimately subjective categorizations.
We rated these technologies as a median 3 out of 7 for concern. Of the 39 example use cases, ~51% (20) received a rating of 2 or below, and ~26% (10) received a rating of 5 or higher.
We also broke this down by product function, mapping onto the categories used in our first report: the AI-technology list we compiled included tools in every product function category identified there—water quality, feed and optimization, health and disease, stock and growth, operations and planning, and, additionally, cross-functional.^[21]
Only three of six categories had more than seven technologies: operations and planning, health and disease, and stock and growth. Given the small sample size in most categories, we cannot argue that there is any meaningful correlation between average concern rating and product function overall. The tentative exception is operations and planning, which clearly diverged from the overall average, with a mean concern rating of 2, and which had no tools that we classified as high risk. This category includes tools such as robot net cleaners, predation deterrents, satellite farm monitoring, and market forecasting. For more information, see Ways That We Could Be Wrong.

AI Example Use Cases by Risk Level

We list how we categorized example use cases by risk level for transparency, but caveat that we appreciate there may be very reasonable arguments for moving use cases up or down in risk level. Further details supporting our reasoning can be found in the database. For more information, see Ways That We Could Be Wrong.
High risk

• Cross-functional
  • AI optimization of husbandry parameters
• Feed and optimization
  • AI-driven precision feeding and real-time satiation detection
• Health and disease
  • AI-mortality prediction
  • Disease detection
  • Disease prediction
• Stock and growth
  • AI-optimized breeding
• Water quality
  • AI-driven environmental control (a standalone, as opposed to a subset of AI-enabled husbandry)

Medium risk

• Cross-functional
  • AI-monitored behavioral depth-steering and submerged precision feeding
• Feed and optimization
  • Real-time satiation detection (monitoring)
• Health and disease
  • AI laser delousing (salmon only)
  • AI-mortality detection
  • Cannibalism/aggression detection
  • Lice detection (salmonid only)
• Operations and planning
  • Labor and process AI-automation
  • Market/supply-demand forecasting and financial risk tools
• Stock and growth
  • AI-informed harvest timing recommendation
  • AI-sorter (grow out)
  • Biomass estimation
  • Molting detection

Low risk

• Feed and optimization
  • Feed waste quantification
• Health and disease
  • AI-guided vaccination
  • Behavioral monitoring
  • Biologger-based physiological welfare monitoring
  • Passive acoustic monitoring
  • Ventilation rate detection (papers on salmon only)
  • Wound/injury detection
• Operations and Planning
  • AI-enabled predation deterrents
  • Aquaculture LLMs
  • Farm management data logging and advisory analytics
  • Fish robot net cleaner
  • Fish robot net fixer
  • Net damage detection
  • Satellite pond/farm monitoring
• Stock and growth
  • AI hatchery screening and embryo sorting
  • Animal counter
  • Animal grader
  • Individual fish image-identification
• Water quality
  • Microbiome health analysis
  • Water quality monitoring

Productivity Gains Due to Feeders

We are uncertain to what extent AI-enabled feeders will increase productivity metrics compared to existing manual and automated systems. Given that conventional automated feeders have already driven productivity gains over manual ones,^[22] as outlined below for P. vannamei,^[23] additional improvements from AI may be incremental.

• Non-acoustic automatic (timed) feeders versus manual feeders:
  • Non-acoustic automatic feeders beat manual feeding on at least one of yield, growth, or weight in all four trials found (Valle et al., 2026; commercial trial), (Jescovitch et al., 2018; research trial), (Liang et al., 2025; research trial), (Valle et al., 2023; commercial trial), with effects ranging roughly from +6% to +22%. None of these trials reported finding that automatic feeders performed worse on yield, growth, or weight than manual feeding.
• Acoustic automatic demand-feeders^[24] versus manual feeders:
  • Acoustic demand-feeders significantly outperformed manual feeding on growth, yield, and weight (~+50% across each) in one trial (Jescovitch et al., 2017; research trial), but came out roughly equal in another (Valle et al., 2023; commercial trial).
• Acoustic automatic feeders versus non-acoustic automatic (timed) feeders:
  • One trial found acoustic automatic feeders increased growth by 25% over timed automatic feeders, but used 63% more feed (Jescovitch et al., 2017; research trial). Another found timed feeding beat acoustic by ~10% on yield (Valle et al., 2023; commercial trial).

Expert Profiles

Table 1: Overview of the background and experience of the three experts interviewed for this report

Ways That We Could Be Wrong

The lists below are non-exhaustive.
Generalization of AI use cases

• We may expect a type of tool to be able to increase stocking density, but only a couple of tools in the category may be good enough to do so.
• Larger welfare implications may come from the same type of tool within one category, or different types of tool between two categories.

Tool applications

• Producer descriptions of tools may not reflect their actual capabilities.
• A tool’s stated function may not match how it is actually used.

Moral considerations

• The large number of trade-offs between improved direct welfare conditions and more farmed animals on aggregate animal welfare
• The extent to which early mortality due to disease, predation, or early slaughter is worse than a longer life and eventual slaughter
• The extent to which animals sorted out from pens to enable differently timed slaughter (e.g., due to sex, growth rate, maturation), are better or worse off
• The extent to which escape from net pens leads to an increase or decrease in welfare for the individual escaped farmed animal
• The extent to which the moral weight of different farmed species should factor into how concerning a technology is rated
• The extent to which early slaughter to avoid disease causes better welfare than longer rearing and disease onset

Risk levels

• Risk levels are assigned per tool type, and not differentiated for species or farm type. For example, we would expect water quality tools used in RAS shrimp farms to have a much bigger impact than those used in open-sea salmon farming.
• Risk levels may depend heavily on species—many of which we may be unfamiliar with (due to the diversity of aquaculture species) and so could not accurately judge—meaning there may be risky use cases we have overlooked or misjudged.

Autonomy

• Whether removing human oversight (increasing autonomy) is net better or worse for individual-level welfare, independent of the second-order expansion risks
• Whether interactions with AI hardware are more or less stressful than interactions with humans or non-AI alternatives
• The current levels of tool autonomy as suggested by the language used by technology manufacturers may be over- or understated compared to actual capabilities

Regional Groupings

• Africa: Ghana, Kenya, Lesotho, Nigeria, Rwanda, Tunisia, Uganda, Zambia
• East Asia: China, Japan, South Korea, Taiwan
• Europe: Bulgaria, Croatia, Cyprus, Denmark, England, Faroe Islands, Finland, France, Germany, Greece, Iceland, Ireland, Italy, Lithuania, Moldova, Netherlands, Norway, Poland, Portugal, Romania, Russia, Scotland, Slovakia, Spain, Sweden, Switzerland
• Latin America: Argentina, Brazil, Chile, Colombia, Costa Rica, Ecuador, El Salvador, Guatemala, Honduras, Mexico, Nicaragua, Panama, Peru
• Middle East: Armenia, Georgia, Israel, Saudi Arabia, Turkey
• North America: Canada, US
• Oceania: Australia, New Zealand
• South Asia: Bangladesh, India
• Southeast Asia: Brunei, Cambodia, Indonesia, Malaysia, Myanmar, Philippines, Singapore, Thailand, Vietnam

Bibliography

Abolofia, J., Frank Asche, & James E. Wilen. (2017). The Cost of Lice: Quantifying the Impacts of Parasitic Sea Lice on Farmed Salmon. Marine Resource Economics. https://doi.org/10.1086/691981

Adekoya, K. (2026, March). Beyond Human-Centered AI: Animal Welfare and Governance in the Global South. Effective Altruism Forum. https://forum.effectivealtruism.org/posts/BDMuEnZXJMTgvynRu/beyond-human-centered-ai-animal-welfare-and-governance-in-3

Alday-Sanz, V., James Brock, Timothy W. Flegel, Robins McIntosh, Melba G. Bondad-Reantaso, Marcela Salazar, & Rohana Subasinghe. (2018). Facts, truths and myths about SPF shrimp in Aquaculture. Reviews in Aquaculture, 12(1). https://doi.org/10.1111/raq.12305

Aquatic AI. (n.d.). Aquaculture for the Age of AI. Retrieved August 27, 2026, from https://caribou-iguana-zg87.squarespace.com/

Asche, F., Lori Snyder Bennear, Atle Oglend, & Martin D. Smith. (2011). U.S. Shrimp Market Integration. Duke University Environmental Economics Working Paper. http://dx.doi.org/10.2139/ssrn.1954082

Aung, T., Rafiza Abdul Razak, & Adibi Rahiman Bin Md Nor. (2024). Artificial intelligence methods used in various aquaculture applications: A systematic literature review. Journal of the World Aquaculture Society. https://doi.org/10.1111/jwas.13107

Balaban, M. O., Sencer Yeralan, & Ymir Bergmann. (2008). Determination of Count and Uniformity Ratio of Shrimp by Machine Vision. Journal of Aquatic Food Product Technology. https://doi.org/10.1300/J030v03n03_04

Bengio, Y., Geoffrey Hinton, Andrew Yao, Dawn Song, Pieter Abbeel, Trevor Darrell, Yuval Noah Harari, Ya-Qin Zhang, Lan Xue, & Sören Mindermann. (2024, May). Managing extreme AI risks amid rapid progress. Science. http://doi.org/10.1126/science.adn0117

Bonnichsen, R., Glenn Gunner Brink Nielsen, Jeppe Seidelin Dam, Dorte Schrøder-Petersen, & Kurt Buchmann. (2024). AI-Driven Realtime Monitoring of Early Indicators for Ichthyophthirius multifiliis Infection of Rainbow Trout. Journal of Fish Diseases, 48(1). https://doi.org/10.1111/jfd.14027

Boyd, C. E. & Aaron A. McNevin. (2024). Resource Use and Pollution Potential in Feed-Based Aquaculture. Reviews in Fisheries Science & Aquaculture, 32. https://doi.org/10.1080/23308249.2023.2258226

Cai, J., Hing Ling Chan, Xue Yan, & PingSun Leung. (2023). A global assessment of species diversification in aquaculture. Aquaculture, 576. https://doi.org/10.1016/j.aquaculture.2023.739837

Cheema, M. A., M A Qureshi, & G B Havenstein. (2003). A comparison of the immune response of a 2001 commercial broiler with a 1957 randombred broiler strain when fed representative 1957 and 2001 broiler diets. Poultry Science. http://doi.org/10.1093/ps/82.10.1519

Chiang, C. (2026, March). Aquaculture fragmentation is a portfolio problem CEAs miss. Effective Altruism Forum. https://forum.effectivealtruism.org/posts/gs9b8zdZ9oEgsxs9N/aquaculture-fragmentation-is-a-portfolio-problem-ceas-miss

Dai, P., Xiangyan Zhu, Sheng Luan, Juan Sui, Xianhong Meng, Jiawang Cao, Jian Tan, & Jie Kong. (2024). Quantitative Genetics of Feed-Efficiency-Related Traits for the Pacific Whiteleg Shrimp Penaeus vannamei in a Plant-Based Diet Environment. Biology, 13(12). https://doi.org/10.3390/biology13121012

Davis, A., Elkin Amaya, Jesus Venero, Oscar Zelaya, & David B. Rouse. (2016, October). Feed Management for Improving Production, Economic Returns for the Semi-Intensive Pond Production of Litopenaeus vannamei. The Fish Site. https://thefishsite.com/articles/feed-management-for-improving-production-economic-returns-for-the-semiintensive-pond-production-of-litopenaeus-vannamei

Dawkins, M. S. (2025). Smart farming and Artificial Intelligence (AI): How can we ensure that animal welfare is a priority? Applied Animal Behaviour Science, 238. https://doi.org/10.1016/j.applanim.2025.106519

FAO. (2005). Seafood Price Indices. https://openknowledge.fao.org/items/0a0d613f-18f8-4ae1-a1c1-bb554b9129aa

FAO. (2026). The State of World Fisheries and Aquaculture 2026. Food and Agriculture Organization of the United Nations. https://doi.org/10.4060/cd8357en

Ferreira, J. C., Marco Esperança, & Øystein Klakegg. (2026, May). AI-Enabled Aquaculture Beyond Performance: A Review of Sustainability, Welfare and Inclusion Impacts. IntechOpen. https://www.intechopen.com/online-first/1247759

Fiocco, D., Vasanth Ganesan, Ana Luiza Mokodsi, Franziska Alesso, & Otto Gryschek. (2024, October). Global Farmer Insights 2024. McKinsey & Company. https://perma.cc/Z3MP-56CH

Fish Focus. (2026, June). Offshore Fish Farming Systems Poised for Rapid Expansion. https://fishfocus.co.uk/offshore-fish-farming-systems-poised-for-rapid-expansion/

Fiskeridirektoratet. (2023). Key figures from Norwegian Aquaculture industry 2023. https://www.fiskeridir.no/english/aquaculture/statistics-for-aquaculture/booklets/_/attachment/inline/de986fdf-1e3e-4dee-b734-f8f5d5a6d4f1:3e1a99a01b179d4013145330ea85abc827290cab/nokkeltall-havbruk-2023-eng.pdf

Fitzgerald, A., Christos C. Ioannou, Sofia Consuegra, Andrew Dowsey, & Carlos Garcia de Leaniz. (2025). Machine Vision Applications for Welfare Monitoring in Aquaculture: Challenges and Opportunities. Aquaculture, Fish, and Fisheries, 5(1). https://doi.org/10.1002/aff2.7003

Føre, H. M., Trine Thorvaldsen, Andreas Misund, & Hans V. Bjelland. (2026). Future of Intensive Pen Aquaculture. Digital Twin for Marine Fish Farming. https://doi.org/10.1007/978-3-032-10362-8_9

Franks, B., Christopher Ewell, & Jennifer Jacquet. (2021a). Animal welfare risks of global aquaculture. Science Advances, 7(14). https://doi.org/10.1126/sciadv.abg0677

Franks, B., Christopher Ewell, & Jennifer Jacquet. (2021b). Animal welfare risks of global aquaculture. Animal Welfare Risks of Global Aquaculture, 7(14). https://doi.org/10.1126/sciadv.abg0677

García-Ballesteros, S., Beatriz Villanueva, Jesús Fernández, Juan Pablo Gutiérrez, & Isabel Cervantes. (2021). Genetic parameters for uniformity of harvest weight in Pacific white shrimp (Litopenaeus vannamei). Genetics Selection Evolution. https://link.springer.com/article/10.1186/s12711-021-00621-6

Georgopoulou, D., Charalabos Vouidaskis, & Nikos Papandroulakis. (2024). Swimming behavior as a potential metric to detect satiation levels of European seabass in marine cages. Frontiers in Marine Science. http://doi.org/10.3389/fmars.2024.1350385

Gkikas, M. C., Michele Thornton, Dimitris C. Gkikas, Spyros Sioutas, & John A. Theodorou. (2026). Artificial Intelligence in Aquaculture Risk Management: A Systematic Review by PRISMA. Applied Sciences. https://doi.org/10.3390/app16042032

HATCH Aquaculture Accelerator. (2019). Feed management and administration. https://perma.cc/47XU-5VWA

Hoyo-Alvarez, E., María J. Cabrera-Álvarez, Margalida Vanrell-Valls, Arjan P. Palstra, & Pablo Arechavala-Lopez. (2026). From signals to states: Biologger-based classification of seabass welfare states in sea-cages. Animal Biotelemetry, 14. https://doi.org/10.1186/s40317-026-00461-5

Jansen, P. A., Anja B. Kristoffersen, Hildegunn Viljugrein, Daniel Jimenez, Magne Aldrin, & Audun Stien. (2012). Sea lice as a density-dependent constraint to salmonid farming. Proceedings of the Royal Society. https://doi.org/10.1098/rspb.2012.0084

Jensen, B.-A. (2024, July). Poor fish health could cost fish farmers at least NOK 30 billion. IntraFish. https://www.intrafish.no/fiskehelse/darlig-fiskehelse-kan-koste-oppdretterne-minst-30-milliarder-kroner/2-1-1672834

Jescovitch, L. N., Carter Ullman, Melanie Rhodes, & Donald Allen Davis. (2017). Effects of different feed management treatments on water quality for Pacific white shrimp Litopenaeus vannamei. Aquaculture Research. https://doi.org/10.1111/are.13483

Kennedy, J. (2022, March). Aquatic AI’s $2M capital raise to scale high-tech marron system. AgriFutures. https://www.growag.com/article/aquatic-ais-2m-capital-raise-scale-high-tech-marron-system

Kling-Eveillard, F., Clément Allain, Xavier Boivin, Valérie Courboulay, Pauline Créach, Aurore Philibert, Yannick Ramonet, & Nathalie Hostiou. (2020). Farmers’ representations of the effects of precision livestock farming on human-animal relationships. Livestock Science, 238. https://doi.org/10.1016/j.livsci.2020.104057

Knapp, G. (2013). The development of offshore aquaculture: An economic perspective. FAO. https://www.fao.org/fishery/docs/CDrom/P24/i3530e/root/12.pdf

Krause, G., Jenny Weitzman, Megan E. Rector, Ramon Filgueira, Sander van den Burg, Dorothy Jane Dankel, Marit Schei Olsen, & Tonje C. Osmundsen. (2024). The social science of offshore aquaculture: Uncertainties, challenges and solution-oriented governance needs. Hypothesis and Theory, 3. https://doi.org/10.3389/faquc.2024.1384037

Liang, Q., Gang Liu, Yazhi Luan, Jiangbo Niu, Yasong Li, Huwei Chen, Ying Liu, & Songming Zhu. (2025). Impact of Feeding Frequency on Growth Performance and Antioxidant Capacity of Litopenaeus vannamei in Recirculating Aquaculture Systems. Animals. https://doi.org/10.3390/ani15020192

Lilburn, M. S. (1994). Skeletal Growth of Commercial Poultry Species. Poultry Science, 73(6). https://doi.org/10.3382/ps.0730897

Limsuwan, C. & Carlos A. Ching. (2013, March). Automatic feeding in shrimp aquaculture. Global Seafood Alliance. https://www.globalseafood.org/advocate/automatic-feeding-shrimp-aquaculture/

Martins, C. I. M., Leonor Galhardo, Chris Noble, Børge Damsgård, Maria T. Spedicato, Walter Zupa, Marilyn Beauchaud, Ewa Kulczykowska, Jean-Charles Massabuau, Toby Carter, Sònia Rey Planellas, & Tore Kristiansen. (2011). Behavioural indicators of welfare in farmed fish. Fish Physiology and Biochemistry, 38. https://doi.org/10.1007/s10695-011-9518-8

Ministry of Trade, Industry and Fisheries. (2017). Regulations on production areas for aquaculture of food fish in the sea of salmon, trout and rainbow trout (production area regulations). LovData. https://lovdata.no/dokument/SF/forskrift/2017-01-16-61

Moldal, T., Jannicke Wiik-Nielsen, Victor Henrique Silva de Oliveira, & Julie Christine Svendsen. (2025). Norwegian Fish Health Report. Norwegian Veterinary Institute. https://www.vetinst.no/rapporter-og-publikasjoner/rapporter/2025/norwegian-fish-health-report-2024/_/attachment/inline/6b11b72c-ee8f-4529-921f-1a3d85dc419e:2d59843d7c1e34e9200669ae47f2974d8ee51b6a/Fish%20Health%20Report%202024.pdf

Moulange, H. & Sagar Shah. (2025, November). Mapping salmon welfare: Sea lice treatments. Rethink Priorities. https://rethinkpriorities.org/research-area/mapping-salmon-welfare-2/

Naylor, R. L., Ronald W. Hardy, Alejandro H. Buschmann, Simon R. Bush, Ling Cao, Dane H. Klinger, David C. Little, Jane Lubchenco, Sandra E. Shumway, & Max Troell. (2021). A 20-year retrospective review of global aquaculture. Nature. https://doi.org/10.1038/s41586-021-03308-6

Netherlands Food and Consumer Product Safety Authority. (2021). Camera surveillance of animal welfare in slaughterhouses. https://www.nvwa.nl/documenten/dierenwelzijn/slachthuizen/publicatie/cameratoezicht-dierenwelzijn-in-slachthuizen

Oliveira, V. H. S., Katharine R. Dean, Lars Qviller, Carsten Kirkeby, & Britt Bang Jensen. (2021). Factors associated with baseline mortality in Norwegian Atlantic salmon farming. Scientific Reports. https://doi.org/10.1038/s41598-021-93874-6

Olkowski, A. A. (2007). Pathophysiology of Heart Failure in Broiler Chickens: Structural, Biochemical, and Molecular Characteristics. Poultry Science, 86(5). https://doi.org/10.1093/ps/86.5.999

Öz, M., Enes Üstüner, Sümmani Çifci, Furkan Budak, Emin İleri, & Suat Dikel. (2025). Artificial intelligence for fish disease diagnosis and management: Innovations, challenges, and One Health implications. Aquaculture International. https://doi.org/10.1007/s10499-025-02352-7

Pasquet, A. (2018, November). Effects of Domestication on Fish Behaviour. IntechOpen. http://dx.doi.org/10.5772/intechopen.78752

Planellas, S. R., João L. Saraiva, Eliane Gonçalves-de-Freitas, Pablo Arechavala-Lopez, Bernice Bovenkerk, Michael Breen, Steven J. Cooke, Martin Føre, Lene Northwood, & Lars Helge Stien. (n.d.). Fish welfare in a changing world: New developments and current challenges. Journal of Fish Biology. https://doi.org/10.1111/jfb.70423

Rath, N. C., Huff G.R., Huff W.E., & Balog J.M. (2000). Factors Regulating Bone Maturity and Strength in Poultry. Poultry Science, 79(7). https://doi.org/10.1093/ps/79.7.1024

Ryan, M. & Leonie Bossert. (2026). Facing the pain: Ethical considerations of AI-based pain detection of farmed animals. AI and Ethics, 6. https://doi.org/10.1007/s43681-025-00905-3

Saraiva, J. L., Maria F. Castanheira, Pablo Arechavala-López, Jenny Volstorf, & Billo Heinzpeter Studer. (2018, April). Domestication and Welfare in Farmed Fish. IntechOpen. https://doi.org/10.5772/intechopen.77251

Scheele, C. W. (1997). Pathological changes in metabolism of poultry related to increasing production levels. Veterinary Quarterly, 19(3). https://doi.org/10.1080/01652176.1997.9694756

Shinn, A., J. Pratoomyot, D. Griffiths, T.Q. Trong, N.T. Vu, P. Jiravanichpaisal, & M. Briggs. (2018). Asian Shrimp Production and the Economic Costs of Disease. Asian Fisheries Science. https://www.asianfisheriessociety.org/publication/downloadfile.php?id=1223&file=Y0dSbUx6QTJOekUyTURVd01ERTFORGMzTWpVek5EZ3VjR1Jt

Shrimp Welfare Project. (n.d.). Shrimp Welfare Project. Retrieved August 26, 2026, from https://www.shrimpwelfareproject.org/

Sonesson, A. K., E. Hallerman, F. Humphries, A. W. S. Hilsdorf, D. Leskien, K. Rosendal, D. Bartley, X. Hu, R. Garcia Gomez, & G. C. Mair. (2023). Sustainable management and improvement of genetic resources for aquaculture. Journal of the World Aquaculture Society, 54(2). https://doi.org/10.1111/jwas.12968

Tuyttens, F. A. M., Carla F. M. Molento, & Said Benaissa. (2022). Twelve Threats of Precision Livestock Farming (PLF) for Animal Welfare. Frontiers in Veterinary Science, 9. https://doi.org/10.3389/fvets.2022.889623

Valle, J. C., César Molina-Poveda, & Miguel Jover-Cerdá. (2023). The Effects of Manual, Time and Sound Feeding Systems on the Growth and Production of White Shrimp (Litopenaeus vannamei) in Semi-intensive Farming Systems in Ecuador. International Journal of Aquaculture, 13. https://www.aquapublisher.com/index.php/ija/article/view/3796

Valle, J. C., Juan Carlos Ayala, Yorvi Reyes, Cesar Molina-Poveda, & Miguel Jover-Cerdá. (2026). Productivity and profitability of Penaeus vannamei production in semi-intensive shrimp farms in Brazil using manual and timed automatic feeding. Aquatechnica. https://doi.org/10.33936/at.v8i1.8252

van Beijen, J. (2025, March). Recirculating aquaculture systems: Money makers or dream breakers? The Fish Site. https://thefishsite.com/articles/recirculating-aquaculture-systems-money-makers-or-dream-breakers

Villalba, M. T., Andrea Fabris, Niels Henrik Henriksen, Leonidas Papaharisis, Yannis Pelekanakis, Michail Pavlidis, & Hans van de Vis. (2020). Guidelines on Water Quality and Handling for the Welfare of Farmed Vertebrate Fish. EU Platform on Animal Welfare Own Initiative Group on Fish. https://feap.info/wp-content/uploads/2020/06/fish-guidelines-june-2020-rev-12.6.2020.pdf

Williamson, S., William McAuliffe, & Hannah Moulange. (2025, December). How AI is affecting farmed aquatic animals. Part 1: Innovation. Rethink Priorities. https://rethinkpriorities.org/research-area/how-ai-is-affecting-farmed-aquatic-animals-1/

Williamson, S., William McAuliffe, & Hannah Moulange. (2026, June). How AI is Affecting Farmed Aquatic Animals. Part 2: Deployment. Rethink Priorities. https://rethinkpriorities.org/research-area/how-ai-is-affecting-farmed-aquatic-animals-2/

World Bank Group. (2014). Reducing disease risk in aquaculture (English). https://documents.worldbank.org/en/publication/documents-reports/documentdetail/110681468054563438

Worm, M., Jan Bulla, John Harald Pettersen, & Benedikt Frenzl. (2026). Control of salmon lice through optical delousing in commercial salmonid aquaculture. Aquaculture. https://doi.org/10.1016/j.aquaculture.2025.742910

Wyban, J. (2009, January). World Shrimp Farming Revolution: Industry Impact of Domestication, Breeding and Widespread Use of Specific pathogen-free Penaeus vannamei. Reseach Gate. https://www.researchgate.net/publication/308052439_World_Shrimp_Farming_Revolution_Industry_Impact_of_Domestication_Breeding_and_Widespread_Use_of_Specific_pathogen-free_Penaeus_vannamei

Xu, W., Bin Zhang, Yongzhen Zhao, & Yucheng Cao. (2025). Effect of Stocking Density on Water Quality, Harmful Nitrogen Control, and Production Performance of Penaeus vannamei in Biofloc-Based Systems with Limited Water Exchange. Fishes. https://doi.org/10.3390/fishes10070326

Zeigler, T. R. (2017, January). Genetics key to maximum growth rate for shrimp. Global Seafood Alliance. https://www.globalseafood.org/advocate/genetics-key-to-maximum-growth-rate-for-shrimp/

Zhang, D., Kirt D. Lillywhite, Dah-Jye Lee, & Beau J. Tippetts. (2014). Automatic shrimp shape grading using evolution constructed features. Computers and Electronics in Agriculture. https://doi.org/10.1016/j.compag.2013.11.009

Zuidhof, M. J., B.L. Schneider, V.L. Carney, D.R. Korver, & F.E. Robinson. (2014). Growth, efficiency, and yield of commercial broilers from 1957, 1978, and 2005. Poultry Science, 93(12). https://doi.org/10.3382/ps.2014-04291

1. See Box 3 for definitions. ↑
2. Since the release of Parts 1 and 2, Shrimp Welfare Project shared a database with us (publication expected at the end of year) containing smart technology and AI aquaculture products based on Chinese-language searches that contains substantially more innovation in the China region than we were able to find. ↑
3. The authors acknowledge that these factors alone do not guarantee commercial uptake. ↑
4. A reader could reasonably disagree with this if aggregate welfare is not the main metric they are concerned with. For more information, see Ways That We Could Be Wrong. ↑
5. It should be noted that we have since found another table of AI applications and technologies in aquaculture (Aung et al., 2025, Table 2), but we decided not to revise our list due to time constraints and the preliminary understanding that a lot of their functions overlap with ours, some are more detailed breakdowns of categories we already capture, and some fall outside the welfare-relevant functions this report focuses on. ↑
6. Grow-out production of Atlantic salmon in 2023 was valued at NOK 106.54 billion (Fiskeridirektoratet, 2023, p. 14) ↑
7. It should be noted that 3 of the 4 authors on the study cited are affiliated with the product manufacturer. ↑
8. We do not discuss tools with the Operations and Planning product function. ↑
9. We could not independently verify these percentages: Ferreira et al. cites 5 sources of which 2 we could not access, and we cannot find supporting information in the other 3. ↑
10. We did not follow up concretely what the comparison was to, but we understood it was with respect to fixed-schedule feeders. ↑
11. Authors caveat that the long-term antibiotic-reduction effects are not yet field-validated, and the findings are currently based on limited data. ↑
12. Authors caveat that the improvement of survival outcomes is context-dependent and is not yet validated under commercial farming conditions. ↑
13. We are skeptical that breeding for welfare traits independently of productivity benefits is incentivized, given productivity gains as a farming prioritization (McKinsey & Company, 2024) ↑
14. So a company using both an AI-enabled predation deterrent and AI-driven precision feeding and satiation detection—the former considered lower risk than the latter for the expansion and/or intensification of farming—is assigned the risk level of the latter. ↑
15. We separated Chile (20 deployments) from the rest of Latin America (40), and Scotland (14) and Norway (24) from the rest of Europe (60), because each country’s deployment count is large enough relative to its region that including it would skew the regional figure toward that one country’s pattern rather than the region as a whole. For more details, see Part 2 (Williamson et al., 2026). ↑
16. We have since found that at least one of these eight companies had already been liquidated prior to the compilation of our database. ↑
17. For more information, see Ways That We Could Be Wrong. ↑
18. The extent to which these tools functionally do predict disease was not verified. Additionally, a specific search was not carried out to decipher whether the tools predicted, monitored, or detected disease, but relied on descriptions already logged in past searches (tool description; information used to log tools under health and disease; information used to assign tool autonomy). ↑
19. We believe that basing aquatic animal preferences on observations from intensive environments—where animals may have developed specific behavioral adaptations to stressful conditions—could lead to ineffective or even detrimental welfare measures. However, we also appreciate that farmed and wild populations can come to have distinctly different needs (Saraiva et al., 2018; Pasquet, 2018), in which case we would expect that basing welfare measures solely on observations of wild conspecifics could be counterproductive. ↑
20. Two articles were not reviewed because we could not access the full texts. ↑
21. It should be noted that many products could be used in combination, potentially increasing their concern ratings, or moving them into a different product category (for example, from water quality monitoring to environmental control). ↑
22. It is our understanding that automatic feeders are somewhat widely used (HATCH, 2019). For example, as early as 2012, more than 60% of shrimp farmers in Thailand were using automated feeders (Limsuwan & Ching, 2013). Unfortunately, we could not find current statistics on coverage. ↑
23. It should be noted that even within the trials being compared in specific reports, not all variables were always held constant (like feeder density). ↑
24. Feeders that use underwater microphones to detect the sound of shrimps eating and release feed according to feeding intensity ↑