Front. Vet. Sci. Frontiers in Veterinary Science Front. Vet. Sci. 2297-1769 Frontiers Media S.A. 10.3389/fvets.2025.1645901 Veterinary Science Perspective AI for One Welfare: the role of animal welfare scientists in developing valid and ethical AI-based welfare assessment tools Foris Borbala 1 * Sheng Kehan 2 Dürnberger Christian 3 Oczak Maciej 4 Rault Jean-Loup 1 1Centre for Animal Nutrition and Welfare, Clinical Department of Farm Animals and Food System Science, University of Veterinary Medicine Vienna, Vienna, Austria 2Animal Welfare Program, Faculty of Land and Food Systems, The University of British Columbia, Vancouver, BC, Canada 3Messerli Research Institute, University of Veterinary Medicine Vienna, Vienna, Austria 4Precision Livestock Farming Hub, Clinical Department of Farm Animals and Food System Science, University of Veterinary Medicine Vienna, Vienna, Austria

Edited by: Gabriela Olmos Antillón, Swedish University of Agricultural Sciences, Sweden

Reviewed by: Charlotte Doidge, University of Nottingham, United Kingdom

Lena-Mari Tamminen, Swedish University of Agricultural Sciences, Sweden

 *Correspondence: Borbala Foris, borbala.foris@vetmeduni.ac.at 05 08 2025 2025 12 1645901 12 06 2025 18 07 2025 Copyright © 2025 Foris, Sheng, Dürnberger, Oczak and Rault. 2025 Foris, Sheng, Dürnberger, Oczak and Rault

This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

The increasing use of artificial intelligence (AI) in livestock farming is accelerating the development of automated welfare assessment tools, particularly with advancement in generative AI such as large multimodal models (LMMs). Yet, animal welfare scientists have rarely been involved in the development process of these tools or their subsequent adaptation within the field. Here, we discuss possible roles for animal welfare scientists in the development and validation of AI-based welfare assessment tools. We first examine key uncertainties that emerge during development, including the selection of relevant, valid and reliable welfare indicators and gold standards, hardware and software solutions for data collection, methods for integrating multiple welfare indicators, and the real-world impact of automated welfare assessment tools. Second, we demonstrate the use of LMMs to assess welfare based on a case study using dairy cow cleanliness. Finally, we consider the practical implementation of AI-based welfare assessment and discuss potential tensions around (1) embedded values in LMMs, (2) AI’s influence on decision-making on farms, (3) the integration of AI in current knowledge systems by human-AI collaboration, and (4) the economics of AI-based welfare assessment and improvement. We conclude that LMMs could help automate welfare assessment and communicate results to humans in accessible formats, but outcomes depend on which stakeholders are involved in the development process. We advocate for developing AI-based welfare assessment tools through the One Welfare framework, recognizing that AI deployment affects humans, animals, and the environment simultaneously, and suggest potential pathways for animal welfare scientists to engage in the process.

 generative AI large multimodal model AI alignment Welfare Quality animal interests precision livestock farming section-at-acceptance Animal Behavior and Welfare Introduction

Growing societal concerns about animal welfare (1), coupled with increased farm automation, have accelerated the development of automated welfare assessment tools (2). Previous attempts mostly used precision livestock farming (PLF) to continuously monitor and optimize environmental conditions, production, and health (3). However, PLF development often lacks rigorous validation by independent third parties, particularly across diverse contexts and environmental conditions (4, 5). While previous work emphasizes the potential benefits PLF could bring to animal welfare in theory (3), applications often fail to address the inherent complexity of welfare as a multidimensional concept that probably could not be adequately captured by a single technological solution (6). The threats of PLF for animal welfare are rarely discussed, such as inaccurate predictions stemming from poor reliability (4), reduced human-animal interactions (7, 8), or increased intensification focusing on productivity without considering welfare needs (9).

Artificial intelligence (AI), particularly computer vision, has gained significant momentum in livestock farming applications (10), including detecting diseases or monitoring animal activity patterns. Recently, the emergence of general-purpose generative AI tools, including large multimodal models (LMMs; systems trained on diverse data types such as text, images, and audio that can process and generate content across different modalities) has initiated a widespread “AI revolution” (11). A variety of foundation models pre-trained on vast amounts of unlabeled data are readily available (12). Foundation models can be adapted to diverse applications (12, 13) through fine-tuning (retraining the model on domain-specific data to optimize performance) and prompt engineering (crafting specific input instructions to guide model outputs without retraining the model itself). The development of foundation models and corresponding applications is advancing globally (e.g., USA: ChatGPT-OpenAI, Claude-Anthropic, Gemini-Google, Llama-Meta, EU: Mistral, China: DeepSeek) with the goal of creating AI agents (14) that perform tasks autonomously following human instructions. However, it is unclear if and how generative AI tools could help animal welfare.

Animal welfare science is inherently value-based (15), because farmers, scientists, and the public may hold diverse perspectives on what constitutes a good life for animals. This diversity creates challenges in establishing unanimous standards for animal welfare assessment, a problem not sparing AI-based tools. Different foundation models already vary in how they weigh animal interests (16) and evaluate animal harms (17), with some erasing the reality of intensive livestock farming (18). Developing valid, ethical, and effective AI-based welfare monitoring requires interdisciplinary collaboration (19, 20) and engagement with farmers and society (21). Scientists may underestimate uncertainties outside their expertise (22) and experience false confidence when using AI (23). Thus, without proactive involvement of welfare scientists in AI-based animal monitoring, unexamined assumptions may limit potential benefits or even cause harm to animals (24).

This paper aims to highlight potential roles for animal welfare scientists in the development of AI-based welfare assessment tools and outline how AI could be used for furthering One Welfare (improving the welfare of animals, humans, and the environment), focusing on farm animals. First, we discuss uncertainties around automated welfare monitoring. Second, we present a case study using LMMs to assess welfare indicators. Finally, we consider how generative AI could support welfare monitoring and One Welfare in practice.

 Uncertainties in automated welfare assessment

Automated welfare monitoring is a multidisciplinary process, requiring expertise in animal welfare, engineering, and data science with its practical impact often depending on ethical, psychological, social, and economical factors (Figure 1). The uncertainties in this process are rarely considered systematically during the development or validation of AI-based monitoring tools. How animal welfare should be assessed depends on how it is defined (e.g., (25)). Frameworks behind current welfare assessment schemes such as the five freedoms (26), three spheres (15), five domains (27), and four principles (28) combine different aspects of animals’ life (e.g., nutrition, comfort, behavior, health, naturalness, affective states) to determine welfare. Moreover, animal welfare science has evolved beyond farm-level evaluations to individual-level (29), and emphasizes evaluating affective states (30), and need for positive experiences (31).

Areas (indicated by different colors) involved in the development of automated animal welfare assessment tools. Questions indicate uncertainties associated with the different components of the development process.

Flowchart visualizing the process of assessing animal welfare. It begins with "Animal Welfare Definition" leading to "Valid Indicator Variable" and further to "Measured Indicator Variable(s)". Branches involve "Hardware" and "Software" linking to "Engineering Technology" and "Computer and Data Science Methods," respectively. Questions throughout address sensitivity, accuracy, data completeness, and integration of variables for welfare assessment, concluding with "Communicating Welfare Status," "Action Suggestion," and "Action Execution." Uncertainty of indicators

The validity of welfare indicators may vary depending on the spatial and temporal level of analysis. For example, environmental and animal-based indicators used in the Welfare Quality protocol (WQP; e.g. cleanliness, illness signs, resource provision) can be measured in a standardized manner and have been validated to reflect farm-level welfare. However, these indicators and integration methods would need to be adapted for AI-based individual welfare assessment. It is also unclear how continuous measurements could be aggregated over time to determine animals’ quality of life (32). Current welfare assessment protocols lack validated affective state indicators. Although some sensor technology approaches have been proposed (33), these require validation (34, 35). Research still needs to identify reliable species-specific affective state indicators for on-farm use.

 Uncertainty of gold standard

Gold standards (annotated datasets with correct labels) are the foundation for training AI-based systems. Some indicator variables may be compromised by poor quality gold standards due to low intra- and inter-observer reliability (36) (e.g., lameness, ear position). Lack of diversity in gold standard dataset limits the model’s generalizability across farms, especially in behavior assessment (5, 20). Establishing robust gold standards for validated indicators may require collaboration with animal welfare scientists in curating diverse datasets.

 Uncertainty of hardware

Data quality varies when collecting data on different farms. Automated welfare assessment relies on sensor, video, and audio data, requiring high sensitivity to detect subtle changes in animals within complex environments. Data “noise” from unidentified factors (e.g., management changes) may overshadow welfare-relevant changes (37). Hardware must withstand harsh conditions (e.g., dirt, collisions) and reliably transmit high-resolution data or run local processing algorithms. These factors limit real-world applicability of experimental solutions (3). Standards for sensor accuracy, calibrations and data cleaning are required but lacking.

 Uncertainty of software

Most current AI-based welfare monitoring tools rely on traditional task-specific machine learning approaches, requiring large amounts of training and testing data (13). In agriculture, LMM use will likely increase (38) as LMMs can perform new tasks using prompt engineering without retraining (39). However, high-quality (albeit small) datasets of welfare indicator variables remain essential. Another challenge is balancing cloud-based LMM processing (requiring large data transfers) with edge-deployed smaller models that process data locally (12). However, model distillation techniques (transferring knowledge from larger to compact models) offer promising solutions (40).

 Uncertainty of integration

Although broad consensus exists on which indicators are essential for comprehensive welfare assessment, how to weigh them appropriately (i.e., valence, severity, duration of conditions) remains unresolved (41–43). Furthermore, each indicator carries measurement errors and uncertainties that accumulate when combined. Although attempts exist to establish cumulative welfare assessment frameworks (e.g., Welfare Footprint Project; (44)), differing ethical views among stakeholders may lead to varying opinions about the most important welfare indicators (15). Determining the cumulative impact of single indicators over time may be easier than aggregating different indicators into one measure.

 Uncertainty of application

The availability of reliable AI-based welfare monitoring does not guarantee system use. Without clear benefits for farmers, uptake may be limited without regulatory requirement or other incentives. Even if these tools are used, improved animal welfare is not guaranteed due to potential “rebound effects” when owners rely excessively on AI-based monitoring and reduce personal observation (45). Human action is still required for corrective actions, and whether this happens may be influenced by how the welfare status is understood and the types of actions required (46) Establishing thresholds that trigger an alarm for action may prove challenging (47), and systems may overlook concerns to reduce alarms to manageable levels. Human factors are key in improving animal welfare (48), emphasizing the need for user-centered design accounting for farmers’ motivations, values, skills, and knowledge.

 AI-based welfare monitoring – a prompt engineering case study

To demonstrate the potential application of LMMs for assessing welfare indicators through prompt engineering, we present a case study. We employed GPT-4o (OpenAI) through Application Programming Interface calls to evaluate cattle cleanliness by categorizing three body parts (hind leg, hindquarter, udder) as either clean or dirty, following the WQP. We used a system prompt instructing GPT-4o to act as an experienced animal welfare scientist with 20 years of farm audit experience. We then provided WQP instructions and implemented few-shot prompting by showing the model two dirty and two clean example images for each body part. Our test dataset comprised eight images per body part (four dirty, four clean) from auditor training materials, with labels by 2 experts (Tierwohltraining, BOKU University, Vienna). We also compared four pre-processing methods: (1) original image (control), (2) bounding box around the body part of interest, (3) segmented image with background removed, (4) segmented image showing only the body part of interest. To account for the probabilistic nature of generative AI, each image was assessed 10 times, with the model explaining its reasoning (Figure 2). Results showed that using segmented images of specific body parts yielded the best performance, with moderate accuracy (hind leg: 0.71, hindquarter: 0.62, udder: 0.52). However, precision was lower (hind leg: 0.63, hindquarter: 0.57, udder: 0.52) and recall of “dirty” images was almost perfect (hind leg: 1, hindquarter: 1, udder: 0.75).

Using GPT-4o to assess cleanliness according to the Welfare Quality protocol for cattle. Columns show image pre-processing approaches. Rows show an example of training and testing images and generated reasoning responses in bubbles (orange false, green correct). The proportion of the given response out of 10 runs is shown at the bottom.

Training and test images display cow hind legs in various stages of cleanliness. Four columns show: 1. Original images, 2. Original images with segments boxed, 3. Segmented images, 4. Segmented body parts. Below, responses label legs as “Dirty” or “Clean,” based on dirt accumulation. Rate of a given response out of 10 rounds is shown underneath.

For the image material used, we achieved a Kappa of 0.5 for segmented hind legs, representing an acceptable level of agreement for human observers. However, the model was biased toward labeling images as “dirty” and removing the background (often containing dirt) brought little improvement. This likely stems from LMMs’ difficulties with spatial reasoning (49) and subsequent challenges in identifying specific body parts for assessment. Focusing the model’s attention using segmentation improved performance. In addition, brown coat pattern and overlapping body parts (e.g., tail over hind leg in Figure 2) may also have biased the model toward classifying images as dirty. We examined only one welfare indicator using a limited image set with varying quality and non-standardized viewing angles; addressing these limitations could potentially enhance performance. However, LMMs sometimes struggle with tasks that humans find trivial (e.g., reading clocks (50)). How well this approach generalizes to other WQP indicators could be further investigated along with a comparison between different LMMs.

LMMs could help scaling up existing protocols, creating new ones (e.g., assessing affective states by decoding vocal and visual signals), and communicating results to humans in a relatable way. However, as our results with cleanliness as one of the “simplest” indicators in WQP demonstrate, current LMMs still require fine-tuning for acceptable accuracy. This could be achieved by providing domain-specific data to open-source LMMs or by utilizing Retrieval Augmented Generation (RAG, see (13) for an example) to provide context-specific materials during prompting.

 Discussion AI for welfare assessment

Due to the outlined uncertainties during development, outcomes from AI-based welfare assessment depend critically on which stakeholders drive the process. On one hand, if development is driven mainly by entities with purely commercial interests, it risks (1) scaling unvalidated metrics that lack animal welfare relevance, (2) further intensification for profit with lower or stagnating welfare standards, (3) not using AI’s potential to improve animal welfare beyond productivity improvements, and (4) restricting farmer and public authority access to data due to proprietary algorithms. On the other hand, if development is driven mainly by entities with unilateral interests in animal protection, it risks (1) scaling up novel assessment methods (e.g., affective states) without sufficient validation, (2) creating solutions disconnected from practical farming realities, leading to frustration and disinterest among farmers, (3) resulting in high-price niche-products, and (4) limiting innovation from commercial companies. The involvement of animal welfare scientists in the tool development and AI regulation process could mitigate some of these risks (e.g., validity concerns) but pathways for engagement are often lacking.

 AI for One Welfare

It has long been recognized that the welfare of animals is interconnected with the wellbeing of humans and the environment (24, 51). We argue that AI-based welfare assessment tools should be developed with a One Welfare mindset, a holistic approach reflecting that the spread of AI simultaneously impacts human and animal lives and the environment.

Who tells AI what is right and wrong?

As AI-tools become prevalent, the messages applications convey to users (or the actions AI agents may take) can have a profound impact. ChatGPT can influence users’ moral judgments, despite their awareness of interacting with a chatbot (52), indicating the potential for biased decisions in the context of livestock farming when users engage with AI systems. This highlights not only the need for user education, but also the importance of AI alignment (ensuring that AI reflects and upholds essential values). However, values in society around the use of animals are not homogenous and animal interests are currently not explicitly considered in AI alignment (53). Current LMMs are not autonomous moral agents but their training data and guidelines influence moral judgments. AI-based welfare assessment tools may implicitly or explicitly follow certain ethical guiding principles (e.g., anthropocentric, pathocentric, biocentric, utilitarian (54, 55)). Livestock farming is tied to complex ethical considerations which may vary based on regions and farming systems. Thus, AI-based welfare assessment needs to be transparent about ethical principles in the model and may require a “practical” ethical approach (56), which is bottom-up (i.e., based on real-life observations) rather than top-down (i.e., based on specific theories). Such an approach may help conscientious people, organizations, and policy makers address nuanced, real-life problems that balance animal welfare, rural poverty, environmental sustainability, and food availability (56–58).

 Who decides what to do?

Technology use and human-animal contact vary across farming systems but currently practical improvements to farm animal lives depend on human decisions. However, some barns may be managed without human presence in the future, a kind of “better wilderness” (59) where animals are cared for and looked after by AI. While humans hold responsibility in current farming systems, they are increasingly supported by data. Nevertheless, rapid digital technology adoption, e.g., in the dairy industry (60), has left many farmers struggling to interpret complex data. LMMs could address this challenge by providing simple explanations of technical information and personalized suggestions for improvement. Decisions impacting animal welfare happen at multiple levels and “decision makers” are a diverse group including farm workers, farm owners, veterinarians, company representatives, and private consumers whose decisions may include making tradeoffs between their own interests, animal welfare, and environmental impact (61). Complex psycho-social processes are also involved in decision-making, which may modulate the interaction between AI tools and the human end user, ultimately leading to variable welfare outcomes (62). Sustainable solutions will require ethical AI frameworks and regulations, clarity about accountability (63), data privacy, and a tailored communication approach recognizing diverse cultural and personal values while mitigating bias without propagating unrealistic outcomes (18).

 Where does AI fit in current systems?

Decisions on farms are results of knowledge systems based on farmer experience and formal training, and that often include advisors (e.g., veterinarian, nutritionist, accountant; (64)). Completely outsourcing welfare assessment to AI is unlikely to be beneficial as diversity in “ways of knowing” strengthens decision-making processes (65). However, integrating LMMs to decision-making systems as an “artificial way of knowing” (i.e., a perspective based on high-resolution longitudinal datasets beyond the scale of human comprehension), and using AI as a “copilot” for farmers could bring enormous benefits (see (66) for a human medicine example). Examples for using AI as agricultural advisor already exist (67) and these approaches could be further developed and integrated with “human-in-the-loop” solutions (68) and farm-specific data to provide region- and context-specific advice in multiple languages, independent of literacy.

 Who will pay for it?

Willingness to financially support a good life (69) for farm animals varies greatly among individuals and societies. Replacing livestock farming with alternative proteins is still in its infancy (70) and may not always be culturally accepted or possible. Establishing high-welfare farming systems while ensuring food security across different cultures and income levels is challenging as the willingness or capability of farmers to improve animal welfare may be limited. Improving welfare standards only with price increases had limited success due to the “consumer-citizen gap” (discrepancies between consumer behavior and attitudes, (71)) and the lack of a sizable uptake that would financially support transformative change (72). Realizing animal welfare standards that are aligned with societal values could be financed centrally (73), similarly to other subsidies to animal agriculture. However, this would require strong political will and meaningful ways to translate animal welfare measures to policy, an area requiring further involvement from animal welfare scientists (74). AI-based welfare assessment may also involve social (e.g., job displacement) and environmental costs (e.g., energy usage, rare-earth mining for computer chips) that should be considered through a One Welfare approach. For both data privacy and reduced energy consumption, deploying small-sized LMMs locally on edge devices would be ideal moving forward.

 What could animal welfare scientists do?

Many tools are already marketed for automated welfare assessment without external validation (4, 5) and development in AI applications is quicker than what many animal welfare scientists and legislators are used to. Here we suggest some actions for animal welfare scientists to contribute to the development of reliable AI-based welfare assessment tools: (1) Delivering accessible communication materials for non-experts to understand the strengths and limitations of welfare indicator variables (e.g., level of evidence and confidence for a variable in a species), (2) Establishing rigorous quality standards for sensor data and indicator variables (75), (3) Supporting the creation of certification schemes for commercial AI-based welfare assessment tools, (4) Advocating for the systematic consideration of uncertainties during the welfare assessment tool development and validation process, (5) Publishing high-quality, annotated, and open-sourced indicator variable datasets along with detailed descriptions (e.g., ethograms, relationship between indicator and welfare) to be used as gold standards and for synthetic data generation (76), (6) Consulting and publishing position papers on incorporating animal welfare into AI development and regulation (e.g., EU AI Code of Practice (77)).

 Conclusion

AI-based welfare assessment tools may bring individualized welfare assessment, which could help providing a good life for farm animals under human care. However, critical uncertainties persist regarding the validity of measurements, methodological approaches, and how end-users interpret and implement the resulting information to improve animal lives. Thus, it is time for animal welfare scientists to get involved in the development and validation of AI-based tools for welfare assessment and improvement in the spirit of One Welfare. This will ensure systems that respect animals along with the wellbeing of farmers, consideration of consumers, and environmental impact.

 Data availability statement

Publicly available datasets were analyzed in this study. This data can be found here: https://tierwohltraining.boku.ac.at/home/uebungszentrm/quiz-uebersicht-rinder/. Code for the case study can be found under the following link: https://github.com/skysheng7/welfare_assess_GPT4o.git.

 Author contributions

BF: Conceptualization, Methodology, Supervision, Visualization, Writing – original draft, Writing – review & editing, Data curation, Investigation. KS: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Visualization, Writing – original draft, Writing – review & editing. CD: Conceptualization, Writing – original draft, Writing – review & editing. MO: Conceptualization, Writing – original draft, Writing – review & editing. J-LR: Conceptualization, Writing – original draft, Writing – review & editing.

 Funding

The author(s) declare that financial support was received for the research and/or publication of this article. KS was supported by a Natural Sciences and Engineering Research Council (NSERC) Discovery Grant (RGPIN-2021-02848; Ottawa, ON, Canada) awarded to Dr. Marina von Keyserlingk. KS also received funding from Elizabeth R Howland Fellowship (Vancouver, BC, Canada), Pei-Huang Tung and Tan-Wen Tung Graduate Fellowship (Vancouver, BC, Canada) and Wilson Henderson Fellowship (Vancouver, BC, Canada). Open Access funding provided by University of Veterinary Medicine Vienna.

We would like to thank Christoph Winckler (BOKU University, Vienna) for providing image material for the case study.

 Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The author(s) declared that they were an editorial board member of Frontiers, at the time of submission. This had no impact on the peer review process and the final decision.

 Generative AI statement

The author(s) declare that no Gen AI was used in the creation of this manuscript.

 Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

 References Busch G Spiller A. Consumer acceptance of livestock farming around the globe. Anim Front. (2018) 8:1–3. doi: 10.1093/af/vfx005, PMID: 32002207 Molina FM Marín CCP Moreno LM Buendía EIA Marín DCP. Welfare quality® for dairy cows: towards a sensor-based assessment. J Dairy Res. (2020) 87:28–33. doi: 10.1017/S002202992000045X, PMID: 33213579 Berckmans D. General introduction to precision livestock farming. Anim Front. (2017) 7:6–11. doi: 10.2527/af.2017.0102 Gómez Y Stygar AH Boumans IJMM Bokkers EAM Pedersen LJ Niemi JK . A systematic review on validated precision livestock farming technologies for pig production and its potential to assess animal welfare. Front Vet Sci. (2021) 8:492. doi: 10.3389/FVETS.2021.660565 Stygar AH Gómez Y Berteselli GV Dalla Costa E Canali E Niemi JK . A systematic review on commercially available and validated sensor Technologies for Welfare Assessment of dairy cattle. Front Vet Sci. (2021) 8:634338. doi: 10.3389/fvets.2021.634338, PMID: 33869317 Weary DM von Keyserlingk MAG. Using animal welfare to frame discussion on dairy farm technology. Animal. (2023) 17:100836. doi: 10.1016/j.animal.2023.100836, PMID: 37793707 Tuyttens FAM Molento CFM Benaissa S. Twelve threats of precision livestock farming (PLF) for animal welfare. Front Vet Sci. (2022) 9:727. doi: 10.3389/fvets.2022.889623 Werkheiser I. Technology and responsibility: a discussion of underexamined risks and concerns in precision livestock farming. Anim Front. (2020) 10:51–7. doi: 10.1093/AF/VFZ056, PMID: 32002203 Cronin GM Rault JL Glatz PC. Lessons learned from past experience with intensive livestock management systems. Rev Sci Tech. (2014) 33:139–51. doi: 10.20506/rst.33.1.2256, PMID: 25000786 Fuentes S Viejo CG Tongson E Dunshea FR. The livestock farming digital transformation: implementation of new and emerging technologies using artificial intelligence. Anim Health Res Rev. (2022) 23:59–71. doi: 10.1017/S1466252321000177, PMID: 35676797 Haque MA Li S. Exploring ChatGPT and its impact on society. AI Ethics. (2025) 5:791–803. doi: 10.1007/s43681-024-00435-4 Li J Xu M Xiang L Chen D Zhuang W Yin X . Foundation models in smart agriculture: basics, opportunities, and challenges. Comput Electron Agric. (2024) 222:109032. doi: 10.1016/j.compag.2024.109032 Ferreira REP Dórea JRR. Leveraging computer vision, large language models, and multimodal machine learning for optimal decision-making in dairy farming. J Dairy Sci. (2025) 108:7493–510. doi: 10.3168/jds.2024-25650 Qiu J Lam K Li G Acharya A Wong TY Darzi A . Llm-based agentic systems in medicine and healthcare. Nat Mach Intell. (2024) 6:1418–20. doi: 10.1038/s42256-024-00944-1 Fraser D. Understanding animal welfare. Acta Vet Scand. (2008) 50:S1. doi: 10.1186/1751-0147-50-S1-S1, PMID: 19049678 Ghose S Tse YF Rasaee K Sebo J Singer P. The case for animal-friendly AI. arXiv. (2024). doi: 10.48550/arXiv.2403.01199 [Preprint]. Kanepajs A Basu A Ghose S Li C Mehta A Mehta R . What do large language models say about animals? Investigating risks of animal harm in generated text. arXiv. (2025). doi: 10.48550/arXiv.2503.04804 [Preprint]. Sheng K Tuyttens FAM von Keyserlingk MAG. The erasure of intensive livestock farming in text-to-image generative AI. arXiv. (2025). doi: 10.48550/arXiv.2502.19771 [Preprint]. Zhang L Guo W Lv C Guo M Yang M Fu Q . Advancements in artificial intelligence technology for improving animal welfare: current applications and research progress. Anim Res One Health. (2024) 2:93–109. doi: 10.1002/aro2.44 Siegford JM Steibel JP Han J Benjamin M Brown-Brandl T Dórea JRR . The quest to develop automated systems for monitoring animal behavior. Appl Anim Behav Sci. (2023) 265:106000. doi: 10.1016/J.APPLANIM.2023.106000 Kelly BL Proudfoot KL Da Silva TE Costa JHC Ventura BA. Animal welfare information frames US public perceptions of precision technology use on dairy farms. J Dairy Sci. (2025) 108:6327–39. doi: 10.3168/jds.2024-25603, PMID: 40252759 Han Y Dunning D. Metaknowledge of experts versus nonexperts: do experts know better what they do and do not know? J Behav Decis Making. (2024) 37:e2375. doi: 10.1002/bdm.2375 Messeri L Crockett MJ. Artificial intelligence and illusions of understanding in scientific research. Nature. (2024) 627:49–58. doi: 10.1038/s41586-024-07146-0, PMID: 38448693 Coghlan S Parker C. Harm to nonhuman animals from AI: a systematic account and framework. Philos Technol. (2023) 36:25. doi: 10.1007/s13347-023-00627-6 WOAH-World Organization for Animal Health. Chapter 7.1. Introduction to the recommendations for animal welfare. In Terrestrial animal health code. (2018). Available online at: https://www.woah.org/fileadmin/Home/eng/Health_standards/tahc/2018/en_chapitre_aw_introduction.htm (Accessed May 23, 2025). Webster J. Assessment of animal welfare: the five freedoms In: Animal welfare: a cool eye towards Eden. Editor J. Webster. Oxford, UK: Blackwell Science (1994). 10–4. Mellor DJ Beausoleil NJ Littlewood KE McLean AN McGreevy PD Jones B . The 2020 five domains model: including human–animal interactions in assessments of animal welfare. Animals. (2020) 10:1870. doi: 10.3390/ani10101870, PMID: 33066335 Welfare Quality. Assessment protocols for cattle, pigs and poultry. Lelystad, The Netherlands: Welfare Quality Consortium (2009). Richter SH Hintze S. From the individual to the population – and back again? Emphasising the role of the individual in animal welfare science. Appl Anim Behav Sci. (2019) 212:1–8. doi: 10.1016/j.applanim.2018.12.012 Reimert I Webb LE van Marwijk MA Bolhuis JE. Towards an integrated concept of animal welfare. Animal. (2023) 17:100838. doi: 10.1016/j.animal.2023.100838, PMID: 37612226 Rault J-L Bateson M Boissy A Forkman B Grinde B Gygax L . A consensus on the definition of positive animal welfare. Biol Lett. (2025) 21:20240382. doi: 10.1098/rsbl.2024.0382, PMID: 39837489 McMillan FD. Quality of life in animals. J Am Vet Med Assoc. (2000) 216:1904–10. doi: 10.2460/javma.2000.216.1904, PMID: 10863585 Neethirajan S. Affective state recognition in livestock—artificial intelligence approaches. Animals. (2022) 12:759. doi: 10.3390/ani12060759, PMID: 35327156 Dawkins MS. Smart farming and artificial intelligence (AI): how can we ensure that animal welfare is a priority? Appl Anim Behav Sci. (2025) 283:106519. doi: 10.1016/j.applanim.2025.106519 Browning H. Validating indicators of subjective animal welfare. Philos Sci. (2023) 90:1255–64. doi: 10.1017/psa.2023.10 Tuyttens F de Graaf S Heerkens J Jacobs L Nalon E Ott S . Observer bias in animal behaviour research: can we believe what we score, if we score what we believe? Anim Behav. (2014) 90:273–80. doi: 10.1016/j.anbehav.2014.02.007 Guarino M Norton T Berckmans D Vranken E Berckmans D. A blueprint for developing and applying precision livestock farming tools: a key output of the EU-PLF project. Anim Front. (2017) 7:12–7. doi: 10.2527/af.2017.0103 Shaikh TA Rasool T Veningston K Yaseen SM. The role of large language models in agriculture: harvesting the future with LLM intelligence. Prog Artif Intell. (2024) 14:117–64. doi: 10.1007/s13748-024-00359-4 Trad F Chehab A. Evaluating the efficacy of prompt-engineered large multimodal models versus fine-tuned vision transformers in image-based security applications. ACM Trans Intell Syst Technol. (2025) 16:1–22. doi: 10.1145/3735648 Hinton G Vinyals O Dean J. Distilling the knowledge in a neural network. arXiv. (2015). doi: 10.48550/arXiv.1503.02531 [Preprint]. Botreau R Bracke MBM Perny P Butterworth A Capdeville J Van Reenen CG . Aggregation of measures to produce an overall assessment of animal welfare. Part 2: analysis of constraints. Animal. (2007) 1:1188–97. doi: 10.1017/S1751731107000547, PMID: 22444863 Botreau R Bonde M Butterworth A Perny P Bracke MBM Capdeville J . Aggregation of measures to produce an overall assessment of animal welfare. Part 1: a review of existing methods. Animal. (2007) 1:1179–87. doi: 10.1017/S1751731107000535, PMID: 22444862 Browning H. Assessing measures of animal welfare. Biol Philos. (2022) 37:36. doi: 10.1007/s10539-022-09862-1 Alonso W Schuck-Paim C. Cumulative pain: an evidence-based, easily interpretable and interspecific metric of welfare loss. Preprints. (2022). doi: 10.20944/preprints202208.0247.v1 [Preprint]. Dürnberger C. Das gläserne Tier. Ethische Fragen zur Digitalisierung in der Nutztierhaltung In: Smart Barning – Digitalisierung in der Nutztierhaltung. München: Hg Internationale Gesellschaft für Nutztierhaltung (IGN). Editor Christoph Winckler, (2021). 70–3. Whay HR. The journey to animal welfare improvement. Anim Welf. (2007) 16:117–22. doi: 10.1017/S0962728600031134 Dominiak KN Kristensen AR. Prioritizing alarms from sensor-based detection models in livestock production - a review on model performance and alarm reducing methods. Comput Electron Agric. (2017) 133:46–67. doi: 10.1016/j.compag.2016.12.008 Hemsworth PH Coleman GJ. Opportunities and challenges to enhance the welfare of livestock through positive human-animal relationships. Appl Anim Behav Sci. (2025) 287:106656. doi: 10.1016/j.applanim.2025.106656 Shiri F Guo X-Y Far MG Yu X Haffari G Li Y-F. An empirical analysis on spatial reasoning capabilities of large multimodal models. arXiv. (2024). doi: 10.48550/arXiv.2411.06048 [Preprint]. Saxena R Gema AP Minervini P. Lost in time: clock and calendar understanding challenges in multimodal LLMs. arXiv. (2025). doi: 10.48550/arXiv.2502.05092 [Preprint]. Pinillos RG Appleby MC Manteca X Scott-Park F Smith C Velarde A. One welfare – a platform for improving human and animal welfare. Vet Rec. (2016) 179:412–3. doi: 10.1136/vr.i5470, PMID: 27770094 Krügel S Ostermaier A Uhl M. Chatgpt’s inconsistent moral advice influences users’ judgment. Sci Rep. (2023) 13:4569. doi: 10.1038/s41598-023-31341-0, PMID: 37024502 Hagendorff T Bossert LN Tse YF Singer P. Speciesist bias in AI: how AI applications perpetuate discrimination and unfair outcomes against animals. AI Ethics. (2023) 3:717–34. doi: 10.1007/s43681-022-00199-9 Rigley E Chapman A Evers C McNeill W. Anthropocentrism and environmental wellbeing in AI ethics standards: a scoping review and discussion. AI. (2023) 4:844–74. doi: 10.3390/ai4040043 DG EPRS. The ethics of artificial intelligence: Issues and Initiatives. (2020). Available online at: https://policycommons.net/artifacts/1337278/the-ethics-of-artificial-intelligence/1944981/ (Accessed June 10, 2025). Fraser DA. “Practical” ethic for animals. J Agric Environ Ethics. (2012) 25:721–46. doi: 10.1007/s10806-011-9353-z Anthony R. The ethical implications of the human-animal bond on the farm. Anim Welf. (2003) 12:505–12. doi: 10.1017/S0962728600026099 Rault J-L Sandøe P Sonntag Q Stuardo L. Positive animal welfare: bridging the gap or raising inequalities worldwide? Front Anim Sci. (2022) 3:825379. doi: 10.3389/fanim.2022.825379 Dürnberger C. A better wilderness? Ethical questions and social ambivalences of precision livestock farming In: Kallhoff A Liedauer E, editors. Greentopia: Utopian thought in the Anthropocene. Cham: Springer International Publishing (2024). 193–205. von Keyserlingk MAG Mills KE Weary DM. Attitudes of western Canadian dairy farmers toward technology. J Dairy Sci. (2024) 107:933–43. doi: 10.3168/jds.2023-23279, PMID: 37709035 Ammann J Mack G El Benni N Jin S Newell-Price P Tindale S . Consumers across five European countries prioritise animal welfare above environmental sustainability when buying meat and dairy products. Food Qual Prefer. (2024) 117:105179. doi: 10.1016/j.foodqual.2024.105179 Rault J-L Holyoake T Coleman G. Stockperson attitudes toward pig euthanasia1. J Anim Sci. (2017) 95:949–57. doi: 10.2527/jas.2016.0922, PMID: 28380586 Cheong BC. Transparency and accountability in AI systems: safeguarding wellbeing in the age of algorithmic decision-making. Front Hum Dyn. (2024) 6:1273. doi: 10.3389/fhumd.2024.1421273 Mills KE Koralesky KE von Keyserlingk MAG Weary DM. Social referents for dairy farmers: who dairy farmers consult when making management decisions. Animal. (2021) 15:100361. doi: 10.1016/j.animal.2021.100361, PMID: 34560439 Robinson CJ Macdonald JM Perry J Bangalang N Nayinggul A Nadji J . Coproduction mechanisms to weave indigenous knowledge, artificial intelligence, and technical data to enable indigenous-led adaptive decision making: lessons from Australia’s joint managed Kakadu National Park. Ecol Soc. (2022) 27:436. doi: 10.5751/ES-13747-270436 Letterie G. Three ways of knowing: the integration of clinical expertise, evidence-based medicine, and artificial intelligence in assisted reproductive technologies. J Assist Reprod Genet. (2021) 38:1617–25. doi: 10.1007/s10815-021-02159-4, PMID: 33870475 De Clercq D Nehring E Mayne H Mahdi A. Large language models can help boost food production, but be mindful of their risks. Front Artif Intell. (2024) 7:1326153. doi: 10.3389/frai.2024.1326153, PMID: 39525499 Mosqueira-Rey E Hernández-Pereira E Alonso-Ríos D Bobes-Bascarán J Fernández-Leal Á. Human-in-the-loop machine learning: a state of the art. Artif Intell Rev. (2023) 56:3005–54. doi: 10.1007/s10462-022-10246-w Farm Animal Welfare Council. (2009). Animal Welfare in Great Britain: Past, Present and Future. FAWC, 17 Smith Square, London. http://www.fawc.org.uk. Malila Y Owolabi IO Chotanaphuti T Sakdibhornssup N Elliott CT Visessanguan W . Current challenges of alternative proteins as future foods. NPJ Sci Food. (2024) 8:53. doi: 10.1038/s41538-024-00291-w, PMID: 39147771 Schäufele I Janssen M. How and why does the attitude-behavior gap differ between product categories of sustainable food? Analysis of organic food purchases based on household panel data. Front Psychol. (2021) 12:5636. doi: 10.3389/fpsyg.2021.595636, PMID: 33679515 Christensen T Denver S Sandøe P. How best to improve farm animal welfare? Four main approaches viewed from an economic perspective. Anim Welf. (2019) 28:95–106. doi: 10.7120/09627286.28.1.095 Busch G Spiller A. “Why we need an animal welfare tax-the citizen-consumer gap,” Diskussionspapiere - Department für Agrarökonomie und Rurale Entwicklung, Universität Göttingen. (2020). p. 13. Available online at: https://www.cabidigitallibrary.org/doi/full/10.5555/20203337013 (Accessed June 6, 2025). Fischer B. Beyond ordinal scales: making animal welfare count in policy analysis. Front Vet Sci. (2025) 12:6475. doi: 10.3389/fvets.2025.1556475, PMID: 40115833 Mason GJ. Animal welfare research is fascinating, ethical, and useful—but how can it be more rigorous? BMC Biol. (2023) 21:302. doi: 10.1186/s12915-023-01793-x, PMID: 38155349 Gaind S Hockings N Oczak M. Evaluating 2D vs 3D neural network efficacy in sow posture detection with synthetic data. In: Roth Peter M. (Institute of Computational Medicine VV), Langs Georg (Computational Imaging Research Lab MV), Licandro Roxane (Computational Imaging Research Lab MV), editors. OAGM Workshop. Vienna, Austria: Austrian Association for Pattern Recognition (OAGM) (2023). European Commission. General-purpose AI code of practice | shaping Europe’s digital future. (2025). Available online at: https://digital-strategy.ec.europa.eu/en/policies/ai-code-practice (Accessed June 6, 2025).