The following text is excerpted from the reading.

3.3. Underexamined AI-driven animal welfare risks

3.3.1. Embedded value hierarchies

Embedded value hierarchies have received little attention to date (Werkheiser, 2020), likely due to the problem of AI alignment: creating purpose built AI systems that do the thing that we want them to do, or more technically, building AI systems that optimize for a value function based on human preferences (Russell, 2019). This is a notoriously hard task (Soares and Fallenstein, 2017). Importantly, as currently defined, succeeding at AI alignment would itself pose risks to the animals. Many intensive farming systems already present animal welfare challenges that reflect a prioritization of productivity, such as limited or absent pasture access on some large dairy farms (Robbins et al., 2016). AI systems designed to optimize production according to the same underlying priorities could potentially exacerbate these existing welfare challenges. More generally, current AI alignment efforts largely treat the relevant stakeholders as humans, not animals, thereby omitting any explicit requirement to protect non-human animals as moral patients in their own right (Tse et al., 2025). AI-enhanced PLF is often presented as a win–win that can improve productivity, environmental performance, and animal welfare simultaneously (Tullo et al., 2019). However, it is unlikely that all three goals will be maximized given differences in prioritization that frequently introduce trade-offs (Werkheiser, 2020). The risk lies in situations where animal welfare is systematically deprioritized in favor of the alternative objectives. For instance, if profitability is in strong conflict with animal welfare, any potential welfare gains that initially motivated the technology may be traded off in favor of economic gains. Also, trade-offs implemented by AI-driven farming systems can be experienced as a technical necessity rather than a moral choice by people involved, which risks undermining responsibility and sustained moral attention to animals’ interests (Werkheiser, 2020).

Even if an AI system is aligned to prioritize animal welfare over productivity, negative consequences may still occur if this is not clearly articulated. For instance, an autonomous AI system may be trained, via reinforcement learning (RL), to optimize the average welfare of animals on a farm (as opposed to aggregate welfare). For example, in a situation where the AI system learns that letting animals with lower-than-average welfare die would increase the average welfare of the remaining animals. In practice, this may involve the AI system choosing not to alert farmers about certain welfare problems thus leading to increased chances that animals with lower-than-average welfare die off. To further explain this point, we turn to an example in population ethics, where in the drive to maximize one value (e.g., average human welfare) undesirable outcomes (e.g., reducing the population of humans to just one person who has very high welfare) are prioritized (Parfit, 1984). Unless we directly address and correct this issue, AI systems designed for optimization could have profound adverse impacts on animal welfare.

When considering the risks of embedded value hierarchies, we investigated auditability and behavioral validation as potential mitigation strategies. In “Auditability”, AI systems would be required to generate, for each choice made, a report that constructs and reconstructs the decision process (Albarracin et al., 2024). This would generate a list of decision-points behind each choice, including data-points deemed relevant for the decisions, and explain how the system resolved conflicts between competing values based on these data-points. The generation of this report would ideally make it harder to treat welfare losses as an “unavoidable” technical feature, since it presents multiple points at which alternative decisions could have been made. It also functions as a “hoop test” (Collier, 2011): On initial inspection, the explanation must appear to prioritize animal welfare to a predetermined minimum standard, and it must read as a believable description of the decision process that actually generated the output.

Although auditability is viewed as a minimal indicator, we caution reliance on systems that only utilize “self-reported” priority rules. Independent behavioral validation should test what is in fact prioritized when welfare and other aspects of farming, such as profitability, conflict. Using logs of outcomes, auditors can examine whether the system systematically becomes less willing to recommend welfare-protecting interventions when other values are at play, ideally giving estimates on the difference (i.e., by how much). This “revealed preference” approach allows evaluators to infer the system’s implicit value hierarchy empirically, and to make necessary adjustments.

3.3.2. Erosion of traditional farmer skills

It is frequently mentioned that AI could lead to the erosion of traditional farmer skills (Kling-Eveillard et al., 2020, Niloofar et al., 2021, Papakonstantinou et al., 2024, Schillings et al., 2021, Tuyttens et al., 2022). This phenomenon, whereby over-reliance on technological assistance leads to the loss of skills required for some task, is often referred to as “deskilling” (Ferris et al., 2010). Farmer skills, which can develop as a result of both formal learning (e.g. agricultural extension services; Nilsson, 2023) and informal learning (e.g. observing and imitating actions on a family farm; von der Bank et al., 2024), include understanding the physical, nutritional, and hygiene needs of animals, being able to detect potential welfare problems or problematic behaviors (i.e., stereotypic behaviors; Mason and Latham, 2004), and being able to apply this information to address welfare issues (Spoolder and Ruis, 2015). The integration of AI into the day-to-day management of farm animals could lead to deskilling if, for instance, experienced farmers start to only rely on AI systems to detect welfare problems, instead of also relying on their own observational skills and judgment. Indeed, this challenge would be further exacerbated if the AI systems malfunction or are inaccurate, as farmers lacking the skills associated with providing high quality care will be less able to identify and thus respond to potential welfare issues.

We contend that recent developments in AI greatly increase the risk of eroding farming skills. In “traditional” PLF, farmers are still required to interpret the data, at least to some degree, gathered by various sensors and then ideally make an evidenced based data driven decision. With the introduction of LLMs, such skills may no longer be necessary, as the model will interpret the sensor data, convey the information to the farmers in a clearly accessible format, and provide personalized suggestions on how to address the issue (Foris et al., 2025). In this situation, farmers may lose the ability to make data driven decisions that would benefit from the integration of their own experiences with the issue. This could again lead to welfare problems if AI systems malfunction or are inaccurate. Additionally, loss of integrated knowledge from traditional farming skills may also limit the development of AI tools in animal farming. To ensure that AI tools are practical, accurate and useful, it is important to involve farmers in the innovation process (Neethirajan, 2023a). However, farmers are only able to provide useful feedback if they have the skills to comment on the accuracy of these tools given their experiences. Without such feedback, the development of accurate, welfare detecting AI tools will be slow and likely fraught with challenges, including AI technologies proven to benefit the animals.

There are several ways to mitigate the risks of farmer skill erosion directly. One solution is to ensure that farmers are trained in good animal handling and husbandry skills but also trained to use AI systems effectively (Tuyttens et al., 2022). If skilled in the latter areas, we predict a more critical approach when reviewing the AI system outputs, potentially leading to questions about inaccurate welfare assessments and outputs arising from technical failures. Potential strategies focused on preventing “deskilling” can be found in other industries. In the aviation industry, pilots are routinely required to have hands-on flight time (instead of relying entirely on auto-pilot), and practice in simulations where system failure is mimicked (Berzin and Topol, 2025). Similarly, farmers could be routinely required to, for instance, produce their own lameness assessments without using AI, or rehearse scenarios where AI systems fail in delivering feed or regulating the environment. Similar recommendations have been suggested in the medical field to address the risk of eroding clinical skills due to AI use (Berzin and Topol, 2025, Natali et al., 2025). Farm managers and agricultural employers are well positioned to implement this solution by setting strict hiring standards and mandatory training for employees.

Another solution is to keep farmers “in the loop”, so that they maintain a key role in decision making rather than relying entirely on AI-enabled closed-loop systems (Niloofar et al., 2021, Werkheiser, 2020). For instance, instead of relying on AI to make decisions about how a sick animal should be treated, or how the environmental conditions (e.g., temperature, light) should be adjusted, AI tools could be designed to present the options to the farmer, ultimately leaving it up to them to make the final decision. If farmers continue to make these key decisions on a regular basis, then they are less likely to lose their knowledge and skills as a result of cognitive atrophy (Gerlich, 2025).

A key step moving forward will be to ensure that farmers are aware of the potential for inaccuracies within some AI systems. For example, to make farm personnel aware of the possible LLM hallucinations that will no doubt occur (Zhou et al., 2025), technology providers could include disclaimers in their products. Ideally, they would also work with farmer groups to help farmers develop competencies in how to detect likely LLM hallucinations. Farmers could also be given clear instructions on how to verify the outputs produced by their virtual assistant if they suspect that these outputs are mistaken.

Finally, AI farming assistants could be designed to function as an educational tool rather than just an advisory one (Kuska et al., 2024). For instance, they could provide an explanation as to why there might be a welfare problem (e.g., “The cow’s behavior suggests she may have mastitis”) or explain why a certain recommendation is made (Kate and Neethirajan, 2026). These types of tools could also improve inaccurate welfare assessment due to “farm blindness” (Brunt et al., 2025, Mee, 2020), or prompt farmers with relevant questions (tailored to their individual skill level) to reinforce their reasoning abilities (Natali et al., 2025). This would allow farmers to retain or improve their own skills in diagnosing animal welfare problems and attending to them. AI developers are best positioned to build such tools, with the help of animal welfare scientists. Policymakers such as governmental agencies could aid in shaping the incentive structure for AI developers, as such tools are unlikely to be incentivized freely by the market.

3.3.3. Shift of traditional farmer identities

The risk of eroding traditional farming skills is closely related to the risk of eroding traditional farming identities. In small-scale farming, having meaningful relationships with animals and having knowledge of individual animals may be an important component of a farmer’s identity (Doidge et al., 2023, Vigors et al., 2023). However, forming close relationships with animals or having knowledge of individual animals may not be as important for farmers if the benefits of doing so (such as being able to identify individual animals at risk) can be achieved using AI. This can lead to a shift in farming identities, where being a “good farmer” is associated with data observation, gathering, and validating (Doidge et al., 2023), rather than traditional stockmanship skills. In situations where human intervention would be needed but is no longer a part of standard animal care practices, animal welfare may be compromised.

More recent developments in AI farming tools, such as incorporating LLMs in the form of virtual farm assistants (Ferreira and Dórea, 2025), could also lead to a shift in farming identities. While farmers may engage less with data as they might do with traditional PLF, they may increasingly see their role to involve following instructions and recommendations provided by their chatbots. The identity of a “good farmer” may therefore be increasingly associated with successfully following these recommendations, rather than having meaningful relationships with, or personal knowledge of, individual animals. Again, this could lead to poor animal welfare outcomes if AI systems are not working for the benefit of animals. To mitigate this risk, we draw on the nursing field which can provide some “lessons learned” on the risk of eroding traditional identities due to AI (Bodur et al., 2025, Ramadan et al., 2024). Similar to the risk of eroding farming identities, some have expressed concern that AI could shift the nursing role “from compassionate professionals to machine operators” (Bodur et al., 2025, p. 10).

This risk can be mitigated by first identifying the unique role that nurses play. In the context of nursing, forming meaningful human connection with patients through compassion and empathy is essential, and cannot be easily replicated using AI (Mohanasundari et al., 2023). Nurses can then be encouraged to see AI as a complementary tool, rather than as a replacement for their own unique role. A similar strategy could be employed to mitigate the risk of eroding farming identities. For instance, farmers may be uniquely placed to cultivate meaningful, positive human-animal relationships. These positive relationships can reduce fear in animals, an essential component when striving to improve animal welfare (Rault et al., 2020). With this unique role in mind, farmers can be encouraged to see AI as a complementary tool rather than as a replacement for their own position. In the medical literature, it has been argued that a nurse in the workplace will become “the information integrator, health coach, and deliverer of human caring, supported by AI technologies, not replaced by them” (Robert, 2019, p. 38). Likewise, a farmer can be identified as the information integrator, health coach, and deliverer of animal caring, supported by AI technologies, not replaced by them.

3.3.4. Objectification, instrumentalization, and speciesism

The rapid integration of AI into animal farming is predicted to pose a significant risk of deepening the objectification and instrumentalization of animals, a process where sentient beings are viewed primarily as tools or data points rather than individuals with intrinsic value (Rollin, 1995, Tuyttens et al., 2022). This process may also reinforce speciesism, understood as the assignment of different moral worth based solely on species membership (Singer, 2015), as some have argued that treating animals merely as resources (i.e., objectification) entails speciesism (Francione, 2000). While all animal farming arguably involves some level of instrumentalization and speciesism, PLF and AI may reinforce or intensify existing tendencies to conceptualize animals primarily in terms of measurable productive functions.

Recent advancements in generative AI and the deployment of digital twins threaten to accelerate this risk through the mechanism of datafication (Kling-Eveillard et al., 2020, Neethirajan, 2026, Neethirajan, 2023a). Digital twins create a virtual representation of the animal based on continuous sensor inputs. While this allows for granular monitoring, it risks reinforcing a feedback loop where the stockperson interacts primarily with the data rather than the individual animal. If a dashboard reports optimal production metrics, physical signs of negative affective states, such as boredom or frustration, that are not captured by sensors may be overlooked. Furthermore, the integration of LLMs as decision-support assistants acts as a semantic filter (Ferreira and Dórea, 2025). An LLM trained exclusively on production performance measures is likely to describe an animal’s state in functional terms rather than welfare-centric terms. By framing welfare issues as technical inefficiencies, these systems effectively obscure the ethical implications of care decisions, eroding the human-animal bond that serves as a critical safeguard for welfare (Werkheiser, 2020).

To mitigate these risks, AI systems must be designed to facilitate, rather than replace, human engagement. One approach is the implementation of ethical interface design, where user interfaces are explicitly designed to highlight animal sentience rather than just productivity. For example, dashboards can be programmed to represent animal welfare on a continuum from negative through neutral to positive (Foris et al., 2025), reporting on specific behaviors associated with positive affective states, such as social play, or use natural language describing individual behavioral traits to re-individualize the animal for the user (Schillings et al., 2021). In theory, AI could even be used to help decode animal communication (Fieldhouse, 2025, Kate and Neethirajan, 2026), potentially giving farmers the ability to understand individual animals on a more personal level. Additionally, protocols should ensure that AI decision support is paired with mandatory physical observation, ensuring the human remains the ultimate moral agent in the care relationship (Kelly et al., 2025, Niloofar et al., 2021). Finally, AI systems must be transparent regarding their objective functions.

Recommendations should explicitly state if a management decision prioritizes productivity over individual welfare, allowing decisionmakers to make an informed ethical choice rather than defaulting to a black box optimization (Ghavi Hossein-Zadeh, 2025).

3.3.5. Humane washing

Humane washing is defined as a deceptive marketing tactic where an organization disseminates disinformation to convince people that the organization takes responsibility for animal welfare (Bjørkdahl and Syse, 2021). For instance, an organization may label their products with misleading language that gives the impression of high welfare but which has no legal definition; a common label frequently used in this way is “humanely raised” (Franklin, 2015). Animal products are often marketed to consumers using images of animals kept under extensive agrarian conditions, such as dairy cows with calves at foot grazing on pasture (Keyserlingk et al., 2013). The intent is to convey messages that farm animals are kept under conditions that align with societal values, often highlighting naturalness and freedom of movement; images that have been exacerbated with generative AI (Sheng et al., 2026). In reality, of course, many farm animals do not have good welfare and live lives that are worse than what is conveyed to the public. If the public are unaware of the actual welfare of these farm animals, then it follows that there will be less societal pressure on farmers and producers to improve animal welfare.

Concerns raised by the public and the desire by farmers to “do the right thing” (Schuppli et al., 2023) have resulted in many businesses wishing to integrate AI into farm management. However, humane washing remains a central challenge. A recent rapid systematic review by Weary and von Keyserlingk (2023) reported that the issue is also prevalent in the scientific literature, with 27 of 28 research articles justifying their research by framing the PLF tool under investigation as providing benefits for animal welfare but using outcome measures that are not directly related to animal welfare. Worrisome is that the humane washing approach will potentially be used to convince some farmers to purchase AI-driven technologies which, in some instances, provide no insights into animal welfare. We see this challenge continuing with the introduction and widespread use of generative AI. Developers of AI tools could make use of text and even images to deceive downstream actors along the value chain about the welfare of farmed animals. If the downstream actors, including citizens and consumers, believe that these animals have good welfare as a result of humane washing, then there will be less pressure on farmers to improve the welfare of their farm animals.

In the following theoretical example using generative AI, we show how the risks concerning animal welfare could unfold (see Figures. S1–S3). An agribusiness company convinces their farmer clients to use a modified LLM which functions as a virtual farm assistant. To ensure that farmers buy into the technology (and continue to support the company) this LLM is intentionally designed to give overly positive descriptions of the farm animal’s welfare status and avoids using language that suggests extremely poor welfare. An example of how this is done could include the following communication by the AI virtual farm assistant: “Cow 4 requires extra feed” which gives no additional information relating to the potential welfare status of this cow. Alternatively, if the LLM were to say, “Cow 4 has not eaten for 2 days; the animal must be inspected visually immediately to determine whether it is ill,” then the farmer is more likely to respond immediately to investigate the situation.

To retain trust in farming, we suggest increases in traceability (i.e., the ability to track a product’s history “through all stages of production, processing, and distribution”) and transparency (i.e., the characteristic of “being visible and open, particularly with respect to information about business practices that may influence economic decisions to support or interact with some company,” Manning et al., 2022, p. 36). In the context of animal welfare, there is tremendous potential to increase traceability using AI if these systems can gather and interpret large amounts of continuous data that track the welfare of farmed animals (Majewski et al., 2026, Neethirajan, 2023a). These data could be used, in theory, to provide an accurate description of farming practices and the welfare of farm animals to the farm owners and the buyers of animal products. With this information, consumers (who have expressed concern for animal welfare; Hyland et al., 2022, Jose et al., 2025) and citizens (Clark et al., 2016) can judge for themselves whether the animals coming through a particular farm and supply chain have been reared and slaughtered using proven animal welfare practices. Obviously, for such a solution to work, the associated AI tools need to be accurate and reliably track the actual welfare of farm animals. As such, risks relating to poor validation and vague welfare measurements will need to be addressed before utilizing AI as a means of increasing traceability. When considering transparency, the ease of access, degree of completeness and accuracy of information, as well as the purpose(s) for sharing provided by the business or product supply chain is important (Turilli and Floridi, 2009). Farmers utilizing AI systems can increase transparency by making relevant information about their farming practices, including but not limited to data collected by their AI systems, more accessible to post farm gate users.

We do recognize that there are also challenges associated with increasing transparency. First and foremost is the issue of privacy (Drewry et al., 2019). When assured at the farm level, increased transparency may lead to some farmers choosing not to adopt welfare-enhancing AI on their farms. A potential solution would be to take a step-by-step approach, where in the case of dairy production the milk processor is privy to the individual farm’s animal welfare status and they in turn ensure their customers that they only include milk from farms that achieve a minimum standard of animal welfare. Although some have suggested the use of data exchange systems which are privacy-preserving (Neethirajan and Kemp, 2021), it seems unlikely that full transparency and privacy can be achieved simultaneously. We therefore strongly encourage approaches to increasing transparency in the animal farming sector by including all actors along the value chain in these discussions, including the farmers and the community, who ultimately buy the products for consumption (Bolton et al., 2024).

Another solution that could aid in circumventing humane washing, that also addresses the risk of “algorithmic inaccuracy and poor validation”, is the development of certification schemes for commercial AI-based welfare assessment tools (Foris et al., 2025). Similar schemes could be developed to ensure that LLM-based virtual farm assistants do not deceive farmers about the welfare of their animals. A similar example can be found in medicine, where LLMs used to provide clinical decision support are required to be approved as medical devices under EU and US law (Gilbert et al., 2023). Similar regulations could be applied to modified LLMs in animal farming, requiring these tools to give accurate outputs to be approved for use. Discussions are needed surrounding minimum standards for language use such as requiring AI tools to use certain words with clear animal welfare relevance (i.e., “hunger”, “distressed”, “fear”) when assessing if particular thresholds are met.

Lastly, farmers should have access to and retain ownership of logs of raw sensor data in addition to the LLM outputs. Having access to such logs would allow farmers to verify the claims made by the LLM; thus, disincentivizing AI producers from customizing these tools to welfare wash. This solution would of course require the raw sensor data to be reliable, and farmers have the appropriate skills or have access to trained support to interpret these data.

4. Remaining challenges and opportunities

The feasibility of the solutions proposed in this paper is heavily dependent on the surrounding socio-political and regulatory environment. While interventions such as mandatory labelling or transparency indexes rely on active civil oversight and democratic institutional structures, these mechanisms may be ineffective or entirely absent in non-democratic contexts. In regions where state transparency is limited and legal protections for whistleblowers are weak, the top-down enforcement of animal-centric AI alignment faces significant barriers. There is also a systemic disconnect between the actors held responsible for animal welfare and those with the practical agency to implement changes. At the individual farm level, farmers are expected by society to take seriously their “duty to care” but in practice may lack the economic or, in the case of AI, the technical autonomy to alter the optimization parameters set by technology providers. Conversely, companies possess the capability to design and deploy welfare-aligned systems, yet they often operate outside the jurisdiction of traditional national animal welfare legislation (i.e., in the EU). This discrepancy means that those most capable of acting on the technology are not always the ones legally or ethically accountable for the results. Navigating these challenges requires a coordinated effort between international standards organizations and policymakers to ensure that responsibility is paired with the technical and economic capacity to act, regardless of the political landscape.

Public policies can play an important role in ensuring that the increasing use of AI on farms translates to positive outcomes for animals. Existing AI regulation tends to focus primarily on the impact that AI may have on humans, devoting little consideration to the interests of animals. The EU AI Act, for instance, makes no mention of animal welfare (McLaughlin, 2025, Regulation (EU), 2024, Val, 2026). Ensuring that AI regulation takes the interests of all sentient beings into consideration could be essential for mitigating AI-driven risks to farm animals. For example, the EU AI Act uses a risk framework to classify AI systems: those classified as “high-risk” are subject to strict requirements, including documenting and analyzing the risks posed by the system, ensuring sufficient transparency to deployers, and ensuring an appropriate level of accuracy and robustness. At present, AI systems are classified as high-risk if they pose a high risk to human rights or safety (Val, 2026). AI systems that pose risks to farm animal welfare may therefore not qualify as being high-risk, despite potentially impacting a large number of animals, unless they also pose sufficient risks to humans. The Act could therefore be amended to ensure that AI systems that pose risks to farm animal welfare are also subject to stricter requirements. Provisions such as these could also be included in existing or future AI regulation outside of the EU, and could significantly impact the trajectory of AI use in animal farming going forward.

Finally, it is important to emphasize that even if AI in animal farming carries many risks for animals, it also has the potential to improve their welfare (Zhang et al., 2024). An AI-enhanced farm which is designed primarily for the benefit of animals could help to reduce welfare problems that stem partly from limited time and resources of human caretakers. In addition, these systems may also monitor and facilitate positive experiences that are increasingly recognized as an important component of welfare (Rault et al., 2025). As we have argued, some uses of generative AI could also lead to improved animal welfare, such as by functioning as an educational tool to improve farmer skills, or by using natural language to describe individual behavioral traits to re-individualize animals for users. Given ongoing developments in general AI capabilities, it is highly plausible that the accuracy and robustness of welfare enhancing AI tools such as these will improve over time, thus increasing their potential to benefit animals. Not using AI to benefit animals may therefore constitute a welfare risk in itself (Bossert, 2023). For this reason, it may be important to ensure that regulation does not completely prohibit the use of AI in animal farming. Establishing ethical guidelines for the use of AI in animal farming (Coghlan et al., 2026, Foris and Rault, 2025) and the means to assure specific AI environments that prioritize the interests of animals may also be a good strategy at this stage, leading to a shift away from farms that either use AI solely to prioritize productivity or make no use of AI that is likely to improve welfare.