Generative AI And Copyright In Agricultural Extensions

Author: LawSphereX

Category: Technology, Cyber & Privacy

Published: 2026-08-23T19:57:40.363Z

Updated: 2026-08-30T07:10:14.181Z

GENERATIVE AI AND COPYRIGHT IN AGRICULTURAL

EXTENSIONS

Shivranjni, CHRIST (Deemed to be University), Delhi, NCR

ABSTRACT

Generative AI models like GPT-4 and Llama 4, powered by large language models (LLMs), have advanced capabilities across diverse fields. Their application in agricultural extension and advisory services, emphasizing climate-smart practices to enhance food production and reduce greenhouse gas emissions. Despite their potential, generative AI raises significant legal challenges regarding training data and outputs. Many AI systems rely on copyrighted materials, raising concerns about intellectual property (IP) infringement. In agricultural advisory services, LLMs can transform key areas like farm mechanization, crop monitoring, livestock management, and water management by optimizing operations, predicting outcomes, and enhancing resource use. For instance, they can analyze data to recommend planting schedules or improve crop yields.

However, these advancements face hurdles, particularly in managing IP issues related to training data. Copyright laws often do not protect AI-generated outputs unless significant human input is involved. Training generative AI on copyrighted materials sparks debates on "fair use. " Legal uncertainties persist, with lawsuits against AI companies addressing copyright infringement. Until resolved, these issues create ambiguity for adopting LLMs in agriculture. This paper highlights these challenges and explores IP implications globally. Keywords: Advisory services, Agricultural extension, Generative AI, Copyright infringement, Legal concerns

Introduction

Gpt-4 and Llama 4 are the two generative AI which is run by LLMs which has made significant advancements, demonstrating remarkable capabilities across diverse domains and tasks. These models showcase machine intelligence levels surpassing earlier AI models, excelling in fields such as coding, medicine, law, agriculture and psychology, and often operating effectively without specific prompts. In some of the domain it challenges the human expertise, suggesting that they may present precursor of AGI systems. This paper will first explore potential applications of generative AI and LLMs in agricultural extension and advisory services that require assurance of climate-resilient agricultural practices, enhanced food production and reduced greenhouse gas emissions in the atmosphere. Generative AI raises legal issues related to training data and data output (Susarla 2024).

In terms of training data, some generative AI tools are trained using materials scraped from the internet, including copyrighted works, personal information, biometric data and harmful or illegal content. There is ongoing litigation over whether the scraping, downloading and processing of these materials, as well as the trained AI models and their outputs, breach intellectual property (IP) rights, privacy and contract terms. Debates continue regarding the balance of interests between IP owners and AI developers. This working paper aims to address the copyright implications of deploying generative AI in agricultural advisory services.

Advisory services in Agricultural Extension: Agricultural extension and advisory services refer to any organization in the public as well as private sectors, which also includes NGOs (non-governmental organizations), farmers and private companies that basically helps farmers and rural stakeholders to access knowledge, information and technology. These body also provides interaction among different stakeholders and assist in developing technical, organizations and management skills to enhance livelihood and well being (Food and Agricultural of the UN, 2019). The successful sharing of information, provision of resources for modernizing agriculture, promoting rural development and shaping individuals’ perspectives on innovative practices all heavily depend on extension training programs (Kassem et al. 2021).

Agricultural extension services include all institutions and initiatives that help those involved in agriculture tackle challenges and gain knowledge, skills and technologies to improve their quality of life and prosperity (Davis, Babu and Ragasa 2020). These services stem from three main sectors (Jayasingh, Ashish and Das 2024, 121): • Public sector is a type of sector which includes governmental bodies such as ministries and departments of agriculture, as well as agricultural research centres, whereas, Private non-profit sector comprises of local and international NGOs, foundations, community boards, associations, bilateral and multilateral aid projects and other non-profit organizations.

And Private for-profit sectors are commercial entities such as input manufacturers and distributors, farmer-operated businesses that both use and share agricultural information, agro-marketing and processing companies, trade associations and private consulting and media firms. Leveraging LLMs for Climate-Smart Agricultural Advisory Services While the adoption and implementation of AI has already been observable in climate smart agricultural practice (Uddin, Chowdhury and Kabir 2024), the application of generative AI in agricultural extension and advisory services encompassing climate smart agricultural practices is still evolving.

It is possible to apply generative AI in various sectors of agricultural extension and advisory services: farm mechanization, the food processing industry, agricultural water management, crop monitoring and improvement and livestock production management (Gaddikeri, Jatav and Rajput 2023). Farm Mechanization: Data can be analysed by LLMs and generative AI from different farm management systems, which includes inventory, financial and labour management. They also contribute in the maintenance of farm machinery and equipment by leveraging sensor data and historical records, which save time and improve efficiency. Moreover, they assist farmers in planning their operations, for example, by identifying the suitable time for planting, harvesting and when to put the fertilisers (ibid. ,22).

Food Processing Industry: Data from sensors and cameras can be analysed by generative AI and LLMs to detect defects in food products and estimate their shelf life. This helps in generating new recipe and meal ideas based on available ingredients and dietary restrictions. In addition, they also can be provided for food safety testing by assessing contamination risks and supports supply chain management in an inventory, logistics and finance area (ibid). In additions this can also provide nutritional information about the food products, answer customer relating quires regarding nutrition and ingredients, create recipes and menu which is specific dietary needs and preference and generate production reports. Time and cost associated with manual data entry are reduced by automating these tasks (ibid).

Agricultural Water Management: Generative AI and LLMs can predict the water needs of the crops by checking historical weather and soil moisture data. These technologies are capable of analysing water samples to detect contaminants, includes heavy metals, pesticides and bacteria. Also, they can provide text-based recommendations for effective water management practices (ibid). Crop Monitoring and Improvement: Data from crop monitoring systems can be analysed by LLMs and generative AI to forecast yields, supporting farmers in harvest planning and operational optimization. These technologies also provide guidance on pest and disease management and contribute to crop breeding experiments by identifying traits that boost yields.

Additionally, by integrating with weather forecasting systems, LLMs and generative AI assess climate change impacts on crop yields and help devise adaptation strategies (ibid. ,23). Application in Livestock Production Management LLMs and generative AI can analyse data from feed management systems to optimize feed usage and reduce waste. They can also assist farmers in identifying issues with their animals so that they can take timely corrective actions. Additionally, they support livestock breeding experiments to identify traits that boost productivity and predicts when equipment, such as feeders and water troughs, needs maintenance based on sensor data. Moreover, they can generate automated reports, including animal health assessments, to help farmers track their operations and identify areas for improvement (ibid).

Copyright Challenges Associated with Generative AI In Output-Related Challenges: Texts generated by an AI language model are generally not protected by copyright laws, as it typically grants to the human creator of an original work. However, if the generated text has significant human input or intervention, it may be taken as original and eligible for the copyright protection. For instance, if a human contributes substantial creative elements on the AI’s responses — by editing, commenting, analyzing or incorporating it— the individual who added such content would hold the copyright for the final work which is been created by mind labour (Lucchi 2023, 7).

LLMs or generative AI in the application to climate-smart agricultural extension and advisory services, copyright protection for the final work (for instance, advice regarding farming or related activities) will not necessarily be an impediment for end user farmers. This is because farmers are meant to receive the advice and suggestions directly from the copyright owner (if the final work qualifies for copyright protection) and use the same by acquiring the prior informed consent of the relevant copyright owner. However, LLMs or generative AI to climatesmart agricultural extension and advisory services in the case of adopting and applying is the actual copyright-related challenges and will arise if copyright-protected materials are used for training relevant AI systems (namely input-related challenges), as described below.

In Input-Related Challenges: A big gap exists in relation to the IP management, specifically, copyright management, of data inputs, meaning the management of IP issues pertaining to the data used for training AI systems (Strowel 2023, 491; Franceschelli and Musolesi 2022; Bonadio, Dinev and McDonagh 2022; Quang 2021, 1407; Sobel 2021; Abbamonte 2021). The main concern in this regard is to determine whether using any copyright-protected material in training an AI system amounts to infringement of copyright (Lucchi 2023, 11). Machine learning depends on extensive training data to deliver accurate results, particularly in areas such as natural language processing, facial recognition, translation and stop sign detection (Brown et al. 2020; Radford et al. 2019, Devlin et al. 2018). This is especially crucial for Chat GPT, which depends on large amounts of data for effective performance.

To generate an interactive and reliable output, Chat GPT needs to collect information from publicly accessible websites on the internet which includes text, images and others (Lucchi 2023, 11). Supporting the training of AI algorithms requires the adoption and application of various techniques such as text and data mining (TDM) and generative deep learning approaches (Franceschelli and Musolesi 2022). TDM processes involve extracting and analyzing extensive data sets to reveal valuable insights and patterns (Alpaydın 2004, 2). These insights and patterns can subsequently be used to improve the performance of AI models.

TDM is an essential tool in AI, helping researchers and data scientists analyze large amounts of unstructured data to extract useful information that would be hard to gather manually. By examining these data sets, AI algorithms can identify patterns and make accurate predictions, leading to new content, discoveries and innovations. Without access to large volumes of data, AI would find it challenging to learn and improve. Therefore, the future of AI completely depends on TDM’s ability. However, main challenge is that AI systems require exact copies of artwork in their training data (Lemley and Casey 2021, 743); this means creating a training set with millions of examples by replicating copyrighted images, videos, audio or text-based works (Lucchi 2023, 11).

Therefore, using Chat GPT or similar models for agricultural extension and advisory services will necessitate creating a training set with millions of examples by replicating copyrighted images, videos, audio and text related to agricultural science and education. As a result, AI training necessitates the question as to whether machine copying should be considered fair use or covered by other copyright exceptions. While it is true that some large data sets are purely informational and not protected by copyright, most training data sets contain copyrighted works. For example, the collections used to develop AI algorithms for text, facial recognition and image recognition include copyrighted materials. This raises the question as to whether using these works is legal and what conditions or restrictions could or should be entailed (Lucchi 2023, 12).

This issue is particularly relevant in the adoption and application of LLMs and generative AI in agricultural extension services because input or training data for this kind of generative AI necessarily requires using copyright-protected images of soil conditions, climate conditions, plant conditions and various images connected with the agricultural sector. It is possible to resolve this challenge if uses of copyright-protected materials for training AI systems are considered as fair use, meaning that they do not amount to infringement of the relevant right holders’ copyright. The following two sections of this paper discuss whether the fair use argument is applicable to uses of copyrightprotected materials for training AI systems. The discussion will cover two major jurisdictions in the Western Hemisphere — the United States and the European Union.

Fair Use or Not: The Example of the United States At present, collecting data for TDM is considered as fair use in the United States (Carroll 2019, 894; Lemley and Casey 2021, 746). In the case of Authors Guild v Google, Inc (2015),1 Google

1 Authors Guild v Google, Inc, 804 F Supp (3d) 202 (2015) at 214-15

Books obtained permission for searching entire libraries in order to provide search functions as well as excerpts from books. However, it is not yet clear if these conclusions are applicable to data collection and input for machine learning. Accordingly, there is no guarantee that the courts will extend this precedent to similar technologies, including those involved in agricultural extension or advisory services (Lemley and Casey 2021, 763). In the United States, data collection for TDM may be permitted if it is seen as a transformative use, but it is not always clear when one copyrighted work is being transformed into another different and distinctive copyrighted work. 2 Furthermore, in the Authors Guild v Google, Inc case, the court determined that Google’s digitization of copyrighted books to create a comprehensive index and improve search functionality constituted fair use3.

The digitization process was designed to help users find and access copyrighted books, serving as a valuable resource for researchers and members of the public without competing with original works. In contrast, generative AI allows users to create content that directly competes with original material, producing new works based on existing content. While Google’s indexing complies with fair use, the accessibility of generative AI raises complex copyright challenges as it enables the creation of derivative works that may impact the market for original content (Lucchi 2023, 13). For this reason, numerous court cases are currently under way in the United States to clarify the definitions of “derivative work” and “transformative use” under IP law, especially regarding copyrighted material used to train AI systems4.

Various famous generative AI platforms such as OpenAI are currently facing lawsuits for alleged copyright infringement in connection with training their AI systems using data sets that their creators have not legally acquired. 5 In Tremblay et al v OpenAI, Inc,6 the plaintiffs argue that OpenAI used their copyrighted material without prior authorization in order to train the AI system for Chat GPT. It has been claimed that Chat GPT can effectively summarize a range of books, indicating that the chatbot has fully engaged with and grasped the content of these 2 A transformative use is a kind of use that “alter [s] the first [work] with new expression, meaning or message. “See Camphell v. Acuff-Rose Music, Inc, 510 US 569 (1994).

3 Authors Guild v Google, Inc, 770 F Supp. (2d) at 207-8

4 See, for example, Getty Images (US), Inc v Stability AI, Inc, No. 1:23-cv-00135-GBW (D Del Mar. 29, 2023) [Getty Images]; Silverman v OpenAI, Inc, No. 4:23-cv-03416 (ND Cal 7 July 2023); Tremblay v OpenAI, Inc, No. 4:2023-cv-03223 (ND Cal 7 July 2023). 5 Ibid. It is expected that the resolution of these litigations will depend on the interpretation of the fair use doctrine.

6 Tremblay v OpenAI, Inc, No. 4:2023-cv-03223 (ND Cal 7 July 2023).

works7. In the case of Silverman v Open AI Inc, the plaintiffs similarly allege that OpenAI improperly used copyrighted material, specifically the book The Bedwetter, in order to train Chat GPT. 8The claimants of this class action suit are a group of authors who assert that Chat GPT can generate summaries of their novels when given an appropriate prompt, thereby demonstrating its familiarity with their content, because Chat GPT’s AI system was trained using their copyrighted materials9. In the case of Getty Images (US), Inc v Stability AI, Inc, the prominent photo agency alleges that the developer of the AI art tool Stable Diffusion illegally scraped a significant number of its images. 10This was reportedly done to train the system without obtaining the necessary permissions and/or compensation to Getty Images.

Additionally, Stable Diffusion generated a modified version of Getty’s watermark, aimed at concealing and facilitating copyright infringement. According to the plaintiffs, this modified watermark also violates the Digital Millennium Copyright Act concerning copyright management information. 11 Since these cases are still pending, the legal confusion regarding LLMs’ and generative AI’s ongoing training methods remain unresolved. However, the above-stated cases signify the f irst major legal challenges related to AI and copyright (Lucchi 2023, 14). Until these cases are resolved, possibilities for the adoption and application of LLMs or generative AI in agricultural extension and advisory services will remain uncertain.

A Recent Case of Concern A recent US Supreme Court ruling on a non-technological case, Andy Warhol Foundation for the Visual Arts, Inc v Goldsmith, has raised concerns about negative impacts on the IP rights of AI-generated works. 12

7 Ibid.

8 Silverman, supra note 4.

9 Ibid.

10 Getty, supra note 4 Images (US), Inc. v. Stability AI, Inc., No. 1:23-cv-00135-GBW (D. Del. Mar. 29, 2023) 11 See Digital Millennium Copyright Act, 17 USC, s 1202(b). Specifically, the plaintiffs argued that the defendant’s actions violated the provisions of the Digital Millennium Copyright Act because it altered or removed copyright management information (CMI) embedded in the plaintiffs’ images and instructed the AI system to exclude any CMI from its generated output.

12 Andy Warhol Foundation for the Visual Arts, Inc v Goldsmith, 143 US 1258 (2023)

The petitioner, the Andy Warhol Foundation for the Visual Arts, Inc. (AWF), manages the works of artist Andy Warhol. Respondent Lynn Goldsmith, a photographer, licensed a photo of the musician Prince to the magazine Vanity Fair in 1984. Warhol then used this photo to create an illustration and a series of prints known as the “Prince Series. ” After Prince’s death in 2016, Condé Nast approached AWF to republish the illustration for a special edition but ended up licensing a different piece from the series. Goldsmith later claimed copyright infringement, leading AWF to file a declaratory judgment action, with Goldsmith countersuing. The district court ruled in favour of AWF, stating Warhol’s use was fair use, but the Second Circuit reversed this decision in 2021. AWF then petitioned for certiorari on the fair use factor, which the US Supreme Court granted.

13The Supreme Court determined that the licensing of Andy Warhol’s “Orange Prince” portrait to Condé Nast did not constitute a transformative use of Lynn Goldsmith’s reference photo because “Goldsmith’s original photograph of Prince, and the Andy Warhol Foundation’s (AWF) copying use of that photograph... share substantially the same purpose, and the use is of a commercial nature. ”14 While the AWF invoked fair use to justify creating derivative works, the Supreme Court ruled that the foundation had no fair use defence for licensing a derivative version of the photograph for commercial purposes. 15 This recent decision could lead to significant limitations on the transformative use doctrine, as the Supreme Court seems to have narrowed its scope (Patry 2023). Hence, it will be intriguing to observe how US courts apply the rules from this case to AI training input licensing.

If a court determines that data ingestion — namely collecting and modifying raw data for AI training — constitutes infringement, it could create significant legal challenges for AI systems. Most data used by generative AI, including text and images, has been obtained without explicit permission from rights holders. The issue at hand is whether using copyrighted works for training data constitutes infringement or if the training purpose is distinct enough to qualify for a fair use defence (Lucchi 2023, 14–15). Fair Use or Not: The Example of the European Union The European Union has adopted a protectionist stance, ensuring accountability for the use of

13 Ibid.

14 Ibid. at 38

15 Ibid. at 2.

training data. Article 4(1) of the Directive on Copyright in the Digital Single Market16 offers a broad exception for TDM. This enables individuals, such as commercial AI developers and educators, to create copies of works or databases for information extraction and retain them as necessary for AI training. 17However, rights holders can exclude TDM exemptions in their contracts with miners (entities involved in TDM) to protect their commercial interests. 18 This provision has garnered considerable criticism for providing a copyright exception that is viewed as being too restrictive (Lucchi 2023, 15). Unlike the traditional understanding of copyright as primarily protecting original expression, this provision appears to cover factual information and data, which has resulted in significant backlash (Margoni and Kretschmer 2022).

Nonetheless, the implementation of this option to opt out, and the extent to which AI developers will adhere to it, are still uncertain (Lucchi 2023, 14–15). Another issue related to data aggregation involves the implementation of EU data protection laws (Hacker, Engel and Mauer 2023, 1112–23). Data aggregation is crucial for training and improving generative AI models (Lucchi 2023, 15). It involves collecting and combining large amounts of data from various sources to enhance a model’s performance and capabilities (ibid. ). However, processing personal data in the European Union must follow strict rules set by the General Data Protection Regulation. These challenges have not been thoroughly examined in both legal and policy contexts and require further investigation and resolution (ibid. ).

Anti-trust/Competition Issues Large, established companies such as Google, Facebook, Amazon and OpenAI have extensive collections of language and image data for AI applications (Foster 2019, 1). This access gives them a significant competitive edge in the AI industry. Consequently, they can leverage their data sets to train and develop more advanced AI models, improving their products and services. For new entrants into the industry, this poses legal challenges, as the ownership and licensing of data sets can be intricate and are subject to IP rights, privacy regulations and other legal 16 EC, Directive 2019/790 of the European Parliament and of the Council of 17 April 2019 on copyright and related rights in the Digital Single Market and amending Directives 96/9/EC and 2001/29/EC, [2019], OJ L 130/92, art 4(1), online: .

17 Ibid. at arts 4(1) and 4(2).

18 Ibid. at art 4(3)

factors. Moreover, the expense of creating or licensing a data set “from scratch” can be substantial, making it challenging for smaller companies to compete with established players (Lucchi 2023, 12). Furthermore, anti-trust concerns may arise if dominant market players control access to the data sets essential for developing AI models, potentially stifling innovation and competition. Therefore, ensuring fair and open access to training data is a vital legal issue in the progress and deployment of AI technology. Recommendations • Data-sharing agreement: Creating clear data-sharing agreements with data providers can be an effective step for any LLM- or generative AI-based service provider, including agricultural extension service providers (cf. Leistner and Antoine 2022).

These agreements can help navigate the complexities of using protected content for AI training while ensuring compliance with copyright laws and safeguarding content owners’ rights (European Commission 2020). The agreements would define the scope of data usage, outline limitations, specify necessary permissions and arrange the required licences for utilizing copyrighted material in AI training (Kop 2021). Data-sharing agreements offer AI developers a legally binding framework for overseeing access to and use of protected content during AI training. These agreements specify authorized data usage and ensure adherence to agreed terms. They also help to identify nonviable data and establish criteria for acceptable content use, including allowed AI algorithms and guidelines 8 Mahatab Uddin for data timing and duration.

In addition, these agreements often include clauses that restrict data usage, preventing developers from retrieving, repurposing or monetizing protected material beyond the initial scope (Lucchi 2013, 18). • Adopting remuneration programs: Another effective strategy for protecting AI training data is to implement remuneration programs, such as revenue sharing or royalty payments, to ensure that creators of copyrighted materials used in AI systems receive fair compensation (Senftleben 2023; Frosio, forthcoming 2025). This strategy is essential for acknowledging the value of copyrighted content and ensuring that creators receive fair compensation for their works used in AI systems.

By adopting revenue-sharing or royalty structures like any other generative AI developers, the LLM- or generative AI-based agricultural advisory and extension service providers can directly connect financial gains to the use of copyrighted materials. This creates a strong incentive for content creators to provide their work as training data, allowing them to benefit from the system’s success. In a revenue-sharing arrangement, creators would receive a predetermined share of the revenue generated by the AI system based on their contribution to the training data, ensuring that they are rewarded for the value that their work brings to the AI’s functionality (Lucchi 2023, 18).

• Royalty-based compensation model: LLM- or generative AI-based agricultural extension and advisory service providers can also introduce a royalty-based compensation model, which will allow content creators to receive a set fee for each instance that an AI system uses their copyrighted works. This fee structure may consist of a fixed amount per use or a percentage of the revenue generated by the AI. Such a model ensures that creators are fairly compensated for the entire duration that their works are utilized, directly tying the fee to their usage (ibid. ). Implementing revenue sharing or royalty structures necessitates clear agreements between AI developers and content creators that define the terms of compensation. These agreements must specify how revenue sharing or royalties are calculated, as well as their frequency and duration.

Transparent remuneration mechanisms safeguard the interests of both parties, fostering a fair and sustainable ecosystem for the use of copyrighted works in AI systems. Overall, compensating content creators through these models recognizes the value of their contributions and ensures they receive fair rewards for their role in the success of generative AI tools, promoting a mutually beneficial relationship and equitable practices (ibid. ). In short, the adoption of any or all of these three suggestions would help in removing copyright-related legal disputes and open the scope for a wider-scale application of LLMs or generative AI in climatesmart agricultural extension and advisory services.

Conclusion: The successful deployment of generative AI in agricultural advisory and extension services relies heavily on addressing copyright-related provisions, particularly those concerning TDM for generative AI. However, a comprehensive legal framework in this area has yet to be established. While the European Union has adopted a protectionist stance by ensuring accountability for the use of training data, the situation in the United States remains uncertain, pending the outcomes of several ongoing court cases. Once these cases are resolved, the extent to which investors and stakeholders in the generative AI sector can rely on the fair use defence for using copyrighted materials to train their AI systems will become 9 Generative AI’s Copyright Challenges in Agricultural Extension clearer.

Regardless of the viability of this defence, generative AI developers can mitigate legal risks by considering strategies such as establishing data-sharing agreements, adopting remuneration programs and implementing royalty-based compensation models. These approaches can help safeguard against copyright infringement claims and support the effective deployment of generative AI in agricultural advisory and extension services in the face of evolving copyright laws.

Work Cited:

1 Alpaydın, Ethem. 2004. Introduction to Machine Learning. Cambridge, MA

MIT Press.

2 Abbamonte, Giuseppe B. 2021. “The Rise of the Artificial Artist

AI Creativity, Copyright and Database Right.” European Intellectual Property Review 43 (8): 702–

10. https://tind.wipo.int/ record/45277?ln=en.

3. Bonadio, Enrico, Plamen Dinev and Luke McDonagh. 2022. “Can Artificial

Intelligence Infringe Copyright? Some Reflections.” In Research Handbook on Intellectual Property and Artificial Intelligence, edited by Ryan Abbott, 245–57. Cheltenham, UK: Edward Elgar.

4. Brown, Tom B., Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla

Dhariwal, Arvind Neelakantan et al. 2020. “Language Models are Few-Shot Learners.” Preprint, arXiv, May 28. https://doi.org/10.48550/arXiv.2005.14165.

5. Bubeck, Sébastien, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric

Horvitz, Ece Kamar, Peter Lee et al. 2023. “Sparks of Artificial General Intelligence: Early Experiments with

GPT-4.”

Preprint, arXiv, March 22. https://doi.org/10.48550/arXiv.2303.12712.

6 Carroll, Michael W. 2019. “Copyright and the Progress of Science

Why Text and Data Mining Is Lawful.” UC Davis Law Review 53: 893–964. https://lawreview.law.ucdavis.edu/sites/g/files/

7. dgvnsk15026/files/media/documents/53-2_Carroll.pdf.

8. Davis, Kristin E., Suresh Chandra Babu and Catherine Ragasa, eds. 2020. Agricultural

Extension: Global Status and Performance in Selected Countries: Synopsis. Washington, DC: International Food Policy Research Institute.

9 Devlin, Jacob, Ming-Wei Chang, Kenton Lee and Kristina Toutanova. 2018. “BERT

Pre-training of Deep Bidirectional Transformers for Language Understanding.” Preprint, arXiv, October 11. https://doi.org/10.48550/arXiv.1810.04805.

10. European Commission. 2020. Towards a European strategy on business-to-government

data sharing for the public interest. Final report prepared by the High-Level Expert Group on Business-to-Government Data Sharing. Luxembourg: Publications Office of the European Union. https://op.europa.eu/en/publication-detail/- /publication/d96edc29-70fd-11eb-9ac9 01aa75ed71a1/language-en.

11. Food and Agriculture Organization of the United Nations. 2019. The international Code

of Conduct for the sustainable use and management of fertilizers. Rome, Italy: Food and Agriculture Organization of the United Nations.

12 Foster, David. 2019. Generative Deep Learning

Teaching Machines to Paint, Write, Compose and Play. Newton, MA: O’Reilly.

13. Franceschelli, Giorgio and Mirco Musolesi. 2022. “Copyright in generative deep

learning.” Data & Policy 4: e17. https://doi.org/10.1017/dap.2022.10.

14. Frosio, Giancarlo. Forthcoming 2025. “Should We Ban Generative AI, Incentivise It or

Make It a Medium for Inclusive Creativity?” In A Research Agenda for EU Copyright Law, edited by Enrico Bonadio and Caterina Sganga. Cheltenham, UK: Edward Elgar.

15. Gaddikeri, Venkatesh, Malkhan Singh Jatav and Jitendra Rajput. 2023.

“Revolutionizing Agriculture: Unlocking the Potential of ChatGPT in Agriculture.” Food and Scientific Reports (3): 20–25. www.researchgate.net/publication/369977080_Revolutionizing_Agriculture_ Unlocking_the_Potential_of_ChatGPT_in_Agriculture.

16. Hacker, Philipp, Andreas Engel and Marco Mauer. 2023. “Regulating ChatGPT and

Other Large Generative AI Models.” In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, 1112–23. New York, NY: Association for Computing Machinery.

17. Jayasingh, Debi Kalyan, Anand Ashish and Kiran Sourav Das. 2024. “Artificial

Intelligence in Agricultural Extension.” In Innovative Agricultural Strategies and Concepts in Extension, edited by Jyotishree Anshuman and Biswajit Mallick, 115–29. New Delhi, India: AkiNik Publications.

18. Kassem, Hazem S., Bader Alhafi Alotaibi, Muhammad Muddassir and Ahmed Herab.

2021. “Factors influencing farmers’ satisfaction with the quality of agricultural

extension services.” Evaluation and Program Planning 85: 101912. https://doi.org/10.1016/ j.evalprogplan.2021.101912.

19. Kop, Mauritz. 2021. “The Right to Process Data for Machine Learning Purposes in the

EU.” Harvard Journal of Law & Technology 34: 1–40. https://jolt.law.harvard.edu/ assets/digestImages/Mauritz-Kop_The-Right-to-Process-Data-for-Machine-Learning Purposes-in-the-EU.pdf.

20. Leistner, Matthias and Lucie Antoine. 2022. IPR and the use of open data and data

sharing initiatives by public and private actors. Brussels, Belgium: European Parliament. www.europarl.europa.eu/RegData/etudes/STUD/2022/732266/IPOL_STU(2022)7322 66_ EN.pdf. Lemley, Mark A. and Bryan Casey. 2021. “Fair Learning.” Texas Law Review 99 (4): 743–78. https://texaslawreview.org/fair-learning/.

21 Lucchi, Nicola. 2023. “ChatGPT

A Case Study on Copyright Challenges for Generative Artificial Intelligence Systems.” European Journal of Risk Regulation 1–

23. https://doi.org/10.1017/ err.2023.59.

22. Margoni, Thomas and Martin Kretschmer. 2022. “A Deeper Look into the EU Text and

Data Mining Exceptions: Harmonisation, Data Ownership, and the Future of Technology.” GRUR International: Journal of European and International IP Law 71 (8): 685 –701. https://doi.org/10.1093/grurint/ikac054.

23. Patry, William. 2023. “Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith:

did the U.S. Supreme Court tighten up fair use?” Journal of Intellectual Property Law & Practice 18 (9): 628–41. https://doi.org/10.1093/jiplp/jpad060.

24. Quang, Jenny. 2021. “Does Training AI Violate Copyright Law?” Berkeley Technology

Law Journal 36 (4): 1407–50. https://doi.org/10.15779/Z38XW47X3K.

25. Radford, Alec, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei and Ilya

Sutskever. 2019. “Language Models are Unsupervised Multitask Learners.” OpenAI. https://cdn.openai.com/ better-languagemodels/language_models_are_unsupervised_multitask_learners.pdf.

26. Senftleben, Martin. 2023. “Generative AI and Author Remuneration.” International

Review of Intellectual Property and Competition Law 54: 1535–60. https://doi.org/10.1007/s40319-023 01399-4.

27 Sobel, Benjamin. 2021. “A Taxonomy of Training Data

Disentangling the Mismatched Rights, Remedies, and Rationales for Restricting Machine Learning.” In Artificial Intelligence and Intellectual Property, edited by Jyh-An Lee, Reto Hilty and Kung- Chung Liu, 221–42. Oxford, UK: Oxford University Press.

28 Strowel, Alain. 2023. “ChatGPT and Generative AI Tools

Theft of Intellectual Labor?”

International Review of Intellectual Property and Competition Law 54 (1): 491–94. https://doi.org/10.1007/s40319-023-01321-y.

29. Susarla, Anjana. 2024. “Generative AI could leave users holding the bag for copyright

violations.” The Conversation, March 22. https://theconversation.com/generative-aicould-leave-users holding-the-bag-for-copyright-violations-225760.

30. Touvron, Hugo, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine

Babaei, Nikolay Bashlykov et al. 2023. “Llama 2: Open Foundation and Fine-Tuned Chat Models.” Preprint, arXiv, July 18. https://doi.org/10.48550/arXiv.2307.09288.

31. Ataharul Chowdhury and Ashad Kabir. 2024. “Legal and ethical aspects of deploying

artificial intelligence in climate-smart agricultural technologies.” AI & Society 39 (3): 221–34. https://doi.org/10.1007/s00146-022-01421-2