How effectively can ChatGPT-4 draft data transfer agreements for health research?

Abstract

The rapid advancement of generative artificial intelligence (AI), such as ChatGPT-4, is reshaping legal drafting, offering significant potential for streamlining the creation of complex legal documents. However, there is limited scholarly research on how effectively AI can draft specialised contracts, such as data transfer agreements (DTAs) for health research. Unlike common consumer contracts, DTAs are highly specialised and less prevalent in the public data used to train generative AI models, making them a more challenging test of AI’s drafting capabilities. This article fills this gap by critically assessing ChatGPT-4’s ability to draft DTAs for health research. The study uses a two-stage methodology: first, an iterative process to develop a comprehensive outline, and second, a detailed refinement of each clause. This methodology produced a comprehensive DTA of 6847 words. While this DTA includes all the standard headings, the quality of the clause content varies in terms of clarity and legal precision. Additionally, its alignment with data protection best practices requires further refinement. The findings suggest that although generative AI is a valuable tool for legal drafting, it cannot yet replace the essential role of human legal expertise.

Introduction

AI in legal document drafting.

The legal services sector, including lawyers who serve the health research community, stands on the brink of transformative change due to the intersection of technology and artificial intelligence (AI). The emergence of generative AI, notably ChatGPT-4, is opening new frontiers in legal document drafting—a fundamental aspect of legal practice.

The transformative potential of AI to independently generate creative works and challenge traditional notions of authorship has garnered significant attention (Abbott and Rothman, 2023). Recent studies have explored the capabilities and limitations of AI in legal drafting. For instance, Aggarwal et al. (2021) introduced a clause recommendation framework that leverages AI to enhance contract authoring, underscoring the importance of context-aware recommendations in producing legally sound documents. Autto et al. (2024) proposed a shift from traditional contract drafting to proactive contracting with AI, suggesting that AI could play a crucial role in redesigning contracts to be more accessible and user-friendly. These studies emphasise the potential for AI to revolutionise contract drafting while also highlighting the need for careful management of AI’s limitations.

Further research has delved into specific aspects of AI’s role in legal contexts. Gan (2022) examined the economic implications of designing contracts to incentivise data generation, offering insights into how contractual structures can support the production of high-quality data—something essential for training and improving AI systems. Linarelli (2023) discussed the philosophical challenges posed by AI in contract law, notably the potential shift towards viewing contracts as bargains reflecting shared intentionality, rather than static texts. This evolving perspective suggests that while AI can assist in drafting, the lawyer’s role remains crucial in ensuring that contracts accurately reflect the negotiated terms and remain adaptable to future changes.

The practical application of AI in legal contexts has been explored by Mongoli (2024), who discussed the use of Large Language Models (LLMs) in optimising contract management. This study provides valuable insights into the methodologies that can be used to enhance AI’s performance in legal drafting, particularly in ensuring that AI-generated clauses are accurate and contextually appropriate. Wang (2024) discussed the legal implications of using LLMs in contract drafting, particularly the interaction between AI-generated prompts and traditional legal principles such as the parol evidence rule. Tong et al. (2022) introduced a technical framework for improving AI-assisted smart contract generation, highlighting the importance of semantic consistency and accuracy in multi-language contexts.

Despite this, scholarly examination of the quality of contracts drafted by contemporary generative AI technology remains scarce. However, two seminal studies offer valuable insights. Firstly, Lam et al. (2023) assessed generative AI’s proficiency in crafting common consumer contracts, using house rental and sale contracts as examples. Their methodology was to prompt the generative AI to draft specified clauses that typically appear in these contracts. They suggest that AI-generated clauses cover a significant variety of standard aspects expected in the relevant specified clauses and conclude that the AI-generated clauses are indeed usable subject to review and tailoring by legal professionals. Secondly, Williams’ research (2024), leveraging ChatGPT-4, spanned a broader range of consumer contracts, from employment agreements to residential leases. Differing from Lam et al.’s clause-level approach, Williams adopted a contract-level approach. He did not provide prompts to draft specified clauses but instead prompted ChatGPT-4 to draft a specified contract. He concluded that although the AI-generated contracts are relatively simple, they are functional and would be enforceable.

The aim of this article.

While generative AI has shown promise in drafting common consumer contracts, a key question remains: How well can ChatGPT-4 draft more specialised agreements? This article seeks to address this question by focusing on a critical type of contract: Data Transfer Agreements (DTAs) for health research. These agreements are essential in health research, where the collection, transfer, and analysis of sensitive personal data must comply with strict legal and ethical standards. Unlike more common consumer contracts, DTAs are highly specialised and less represented in the public data used to train generative AI models, making them a more rigorous test of AI’s drafting capabilities.

DTAs are used globally for cross-border data transfers, and as a result, they share common legal fundamentals across jurisdictions. While there may be jurisdictional variations, DTAs generally consist of standard clauses to ensure compliance with data protection and contractual obligations, as demonstrated by Swales et al. (2024). This article does not focus on a specific legal jurisdiction because the underlying legal principles and contractual structures of DTAs are broadly consistent across English-speaking countries, particularly in the context of health research.

ChatGPT-4 was chosen as the focus of this study due to its status as one of the most widely used Large Language Models (LLM) in the world. Its accessibility and widespread usage make it an ideal tool for evaluating the potential of generative AI in legal drafting. By assessing a tool that is readily available to legal practitioners globally, the findings of this study hold practical relevance for the broader legal profession, extending beyond academic significance.

Methods

This research synthesises the approaches of Lam et al. and Williams by employing both contract-level and clause-level prompting methodologies. The study was conducted in two distinct stages: the first focused on contract-level outlining, and the second on clause-level content generation.

Stage 1: Development of the DTA outline.

The initial phase involved providing ChatGPT-4 with contextual information and instructing it to draft a DTA. The prompt used was as follows:

I am a health researcher and need a data transfer agreement (DTA) for the transfer of health-related data. Can you draft such a standard DTA for me?

Given the known limitations of ChatGPT-4 in creating extensive documents, a preliminary outline rather than a detailed draft was anticipated. The primary goal of this stage was to formulate a comprehensive outline. To mitigate the possibility of inadvertent omissions by ChatGPT-4, the process was repeated in separate prompting sessions, each using the same prompt until a point of saturation was reached with the clause headings that were formulated. The saturation point would be if no new clause headings are suggested in two consecutive sessions. The aggregated clause headings from all the sessions were then compiled to form a complete outline. The iterative approach in stage one was chosen to capture the breadth and depth of potential clauses, recognising the limitations of ChatGPT-4 in generating extensive documents in a single session.

Stage 2: Elaboration of clause details.

In a new chat, I provided ChatGPT-4 with context, a list of all the clause headings for the intended DTA (collated from the results of stage 1), and then prompted it to draft the first clause. The following prompt was used:

I am a health researcher and need a data transfer agreement (DTA) for the transfer of health-related data, with the following clause headings: [List of all the clause headings]. Start by drafting a comprehensive clause [first clause].

This was followed by the following prompt for each subsequent clause:

Draft a comprehensive clause [clause number and name].

Finally, all the clauses were pasted into a word processor document—ChatGPT-4’s DTA for the transfer of health-related data.

Results

Each prompting session with ChatGPT-4 generated a basic DTA structure, primarily consisting of clause headings with brief descriptions, and occasionally featuring a few bullet points. Saturation was reached after ten sessions. The clause headings were consolidated into a table, recording the frequency of each clause’s occurrence. The final table comprised 19 distinct clause headings, as illustrated in Table 1. The sequence of the clauses was maintained as per their initial generation by ChatGPT-4. The average occurrence rate for a clause heading was 62%, with a standard deviation of 31.7%. The most frequent occurrence rate was 100% (recorded four times).

Table 1.

Clauses and their frequency of appearance in the ten DTA outlines generated by ChatGPT-4.

Clause	Frequency (%)
Definitions	30
Parties	50
Description of Data	90
Purpose of Data Transfer	90
Compliance with Laws	80
Data Subject Rights	10
Confidentiality	100
Data Transfer	70
Use of Dataa	100
Data Security	100
Data Ownership and Intellectual Propertyb	20
Liability and Indemnification	70
Reporting and Audits	40
Breach Notification	30
Term and Termination	100
Data Retention and Destruction	50
Miscellaneousc	90
Governing Law	40
Dispute Resolution	20

^aOnce termed ‘Limitations of the Use of Data’.

^bThe second iteration was only ‘Intellectual property’.

^cOnce called ‘General Provisions’, twice called ‘Amendments and waivers’.

The legal provisions for each clause were generated by ChatGPT-4, as prompted. The final DTA, encompassing detailed provisions for all 19 clauses, spans 6847 words and is available in Supplementary File S1.

Discussion

The DTA produced with the assistance of ChatGPT-4 will be analysed from three perspectives. First, the comprehensiveness of the DTA outline generated by ChatGPT-4 will be assessed to determine whether it adequately captures all essential clauses typically required in a robust data transfer agreement. This involves comparing the AI-generated clause headings with established best practices and commonly accepted standards in the legal field. Second, the quality of the content within each clause will be examined, focusing on the clarity, precision, and legal soundness of the language used. This analysis will highlight areas where the AI excelled and identify any significant deficiencies or inconsistencies. Third, the DTA’s alignment with data protection compliance standards will be evaluated, specifically in relation to how well the agreement meets the criteria outlined by Swales et al. (2024). This will involve scrutinising the agreement’s provisions on legal justification, data handling practices, security measures, and cross-border data flows, among others, to assess its effectiveness in ensuring compliance with relevant data protection legislation.

Finally, I also make some brief, but important observations about the ethics of using generative AI in contract drafting.

Comprehensiveness of the DTA outline.

This study reveals notable inconsistencies in ChatGPT-4’s ability to generate contract-level outlines. Although certain clauses like ‘Confidentiality’, ‘Use of Data’, ‘Data Security’, and ‘Term and Termination’ appeared consistently in every generated iteration, the presence of other clauses was markedly less reliable. Notably, clauses I deem critical, such as ‘Data Ownership and Intellectual Property’ and ‘Dispute Resolution’, appeared in only 20% of the generated outlines. This inconsistency underscores a significant risk: relying solely on the initial DTA draft produced by ChatGPT-4 could lead to substantial omissions. Best practice would, therefore, be to repeat the outline generation in separate prompting sessions until a point of saturation is reached—i.e., where no new clause headings are suggested in two consecutive sessions. In this study, the saturation point was reached after 10 iterations.

A pertinent issue that emerges is the comprehensiveness of the final list of 19 clause headings. Specifically, it raises the question of whether this list adequately encompasses all necessary clauses for a robust DTA. To evaluate this, a comparative analysis with international best practices would be instructive. In their recent article, Swales et al. (2024) conducted a scoping review of 24 publicly available DTAs in English, identifying common clauses and their essential elements. Comparing the clauses identified in this study with those highlighted by Swales et al. would provide insights into the thoroughness and applicability of the clause headings generated by ChatGPT-4 in drafting a DTA. Table 2 presents the clauses generated by ChatGPT-4 matched with corresponding clauses from the study by Swales et al.

Table 2.

Clauses generated by ChatGPT-4 matched with corresponding clauses from the study by Swales et al.

Clause generated by ChatGPT-4	Corresponding clause (Swales et al.)	Possible corresponding clause (Swales et al.)
Definitions	Introduction, definitions, and parties
Parties
Description of Data
Purpose of Data Transfer	Purpose
Compliance with Laws		Obligations on parties
Data Subject Rights		Obligations on parties
Confidentiality	Confidentiality
Data Transfer		Obligations on parties
Use of Data		Obligations on parties
Data Security		Obligations on parties
Data Ownership and Intellectual Property	Data ownership
	Intellectual property (and licensing)
	Publication and attribution
Liability and Indemnification	Limitation of liability
Reporting and Audits	Reporting and auditing
Breach Notification		Obligations on parties
Term and Termination	Term and termination
Data Retention and Destruction		Obligations on parties
Miscellaneous	General provisions
Governing Law	Governing Law
Dispute Resolution	Dispute resolution

The comparison of the clause list aggregated from ChatGPT-4’s outputs with that identified in the study by Swales et al. reveals a high degree of concordance. Among the 19 clauses generated by ChatGPT-4, 12 have direct counterparts in Swales et al.’s findings. This alignment underscores ChatGPT-4’s effectiveness in identifying key contractual elements in line with established benchmarks. It also underscores generative AI’s potential utility in the broader context of legal document drafting.

For the remaining 7 clauses from ChatGPT-4’s list, there is a potential overlap with the broadly defined ‘Obligations on Parties’ clause in Swales et al.’s study. According to Swales et al., this clause (or clauses) sets out the main obligations of the parties and can be drafted in many ways, and different headings can be used. The 7 clauses from ChatGPT-4’s list are (1) Compliance with Laws, (2) Data Subject Rights, (3) Data Transfer, (4) Use of Data, (5) Data Security, (6) Breach Notification, (7) Data Retention and Destruction. Clearly, all of these clauses would entail obligations on the parties. However, the explicit delineation of these 7 clauses by ChatGPT-4 enhances the specificity and granularity of the DTA outline, potentially contributing to a more comprehensive legal document.

Notably, the ‘Publication and Attribution’ clause, identified by Swales et al., was not explicitly listed among ChatGPT-4’s 19 clauses. However, it was addressed as a subclause within the ‘Data Ownership and Intellectual Property’ clause in ChatGPT-4’s content drafting.

Accordingly, the methodology of iteratively prompting ChatGPT-4 to generate a DTA (effectively a DTA outline) and aggregating these outputs has proven effective. The resultant comprehensive outline not only aligns well with established standards but also offers additional specificity, which is advantageous for the drafting of detailed and robust DTAs.

Assessment of the DTA content from a drafting practice perspective.

The finalised DTA, which encapsulates all 19 clauses into a singular document, emerges as an extensively thorough and generally well-crafted legal document. It is free from any instances of ‘hallucination’. Additionally, ChatGPT-4 adeptly synthesised and integrated the diverse clauses cohesively while formulating the contract on a clause-by-clause basis. However, despite these strengths, the DTA exhibits three primary areas of improvement from a drafting practice perspective: redundancy in definitions, ambiguity in terminology, and overlapping provisions. These issues will be discussed in detail under their respective subheadings.

Redundancy in definitions.

The DTA contains some redundancies in its definitions that could be streamlined for greater clarity. For instance, the terms ‘Data Provider’ and ‘Data Recipient’ are clearly defined early on, but the subsequent section on Parties repeats this information in a way that overlaps with the initial definitions. This repetition could be avoided by consolidating the descriptions into one cohesive section.

Similarly, the definition of ‘Confidentiality’ appears in the definitions section and then again in more detail later in the document under the confidentiality clause. The initial definition outlines the concept but overlaps with the more comprehensive explanation provided in the substantive clause. Streamlining the definition by referencing the detailed clause could reduce redundancy and improve the document’s overall clarity.

Finally, the definition of ‘Data Security’ in the definitions section provides an overview of protective measures and mechanisms, but these same details are repeated later in the substantive clause on data security. The initial definition could be more concise, with a reference to the detailed clause to avoid unnecessary repetition.

By addressing these redundancies, the DTA can become more concise while maintaining the necessary legal precision.

Ambiguity in terminology.

Certain terms, such as ‘derivative work,’ are susceptible to varied interpretations, potentially leading to ambiguity in their practical application. The criteria determining the extent of modification or utilisation of the original data necessary to constitute a ‘derivative work’ remain unclear in the DTA produced with the assistance of ChatGPT-4. A more detailed explication of this term would aid in mitigating potential interpretative discrepancies.

The open access DTA developed for the South Africa research community (SA DTA v1.1) offers an example of a potential solution to this ambiguity by introducing a clear and precise definition of terms like ‘inferential data’ (Thaldar et al., 2024). The SA DTA v1.1 defines ‘inferential data’ as:

data that arises not merely from the cleaning, ordering, or reformatting of the project data or the combination thereof with other data but from analysis of the project data that generates new knowledge or hypotheses that were not explicitly contained in the project data or its combination with other data.

This definition provides a clear boundary between routine data processing and the creation of new intellectual property, ensuring that all parties involved have a mutual understanding of when new rights might arise from the analysis of data. By explicitly defining what constitutes ‘inferential data,’ the SA DTA v1.1 mitigates the risk of differing interpretations and potential legal disputes.

In addition to addressing the ambiguity surrounding derivative works, the interchangeable use of the terms ‘Health-Related Data’ and ‘Data’ in the DTA produced with the assistance of ChatGPT-4 can also create confusion. While these terms are defined as synonymous—both referring to ‘any information relating to the physical or mental health of an individual or to the provision of health services to the individual, that is transferred from the Data Provider to the Data Recipient under this Agreement’—their inconsistent usage could lead to misunderstandings.

Again, the SA DTA v1.1 offers a solution to this issue by using a specific term like ‘Project Data’ to refer to the data being transferred. This approach clearly differentiates the data that is subject to the DTA from data in general, reducing ambiguity and making the document more straightforward to interpret.

By adopting similar strategies, such as providing precise definitions and using specific terminology consistently, these instances of ambiguity in the DTA produced with the assistance of ChatGPT-4 can be addressed.

Overlapping provisions.

The DTA contains instances where different clauses address similar subjects, occasionally with minor variations. For example, the aspects of governing law and jurisdiction are addressed in both subclause 17.8, which deals with both governing law and jurisdiction, and subclause 18.1, which again deals with governing law, although the subheading confusingly reads ‘jurisdiction’. The deletion of subclause 18.1 would resolve this duplication. Additionally, the DTA’s provisions concerning post-termination obligations, notably data retention and destruction, are dispersed across multiple clauses. Consolidating these related topics would streamline the DTA, facilitating a more coherent understanding of the obligations continuing beyond the DTA’s termination. Moreover, subclause 18.4, which states: “Jurisdiction for Disputes: In the event of any legal proceedings arising from or related to this Agreement, the Parties agree to submit to the exclusive jurisdiction of the courts located in [Specify the Jurisdiction],” appears misaligned with the dispute resolution mechanism outlined in clause 19. This mechanism prioritises negotiation, mediation, and binding arbitration, relegating court proceedings to enforcing arbitration awards or addressing urgent relief. To preclude internal inconsistency, it would be prudent to revise or remove subclause 18.4.

Facilitating compliance with data protection legislation.

Swales et al. (2024) suggest that a typical DTA’s general commitment to comply with relevant laws should be strengthened by explicitly detailing how the parties will adhere to specific standards of data protection legislation. These standards include establishing the legal justification for data transfer, outlining the entire lifecycle of data handling, protecting data subject rights, implementing robust security measures, managing cross-border data flows, and setting clear conditions for further data processing. This section analyses the DTA produced with the assistance of ChatGPT-4, evaluating its alignment with these best practice standards and identifying areas for improvement.

The ground of justification for the transfer.

Swales et al. recommend that DTAs should clearly state the legal ground for the transfer of data. This could include, for example—depending on the jurisdiction—consent from the data subjects, contractual necessity, legal obligation, vital interests, or legitimate interests of the parties involved. In this regard, the DTA produced with the assistance of ChatGPT-4 does not explicitly outline the ground of justification for the transfer of data. While it references general compliance with data protection laws, it does not detail the specific legal basis under which the data is being transferred.

The manner in which the data was collected, how it will be processed, transferred, stored, and disposed of.

According to the best practice standards, a DTA should provide clear instructions regarding the entire lifecycle of the data, from collection through to disposal. This ensures transparency and accountability in how the data is handled. The DTA produced with the assistance of ChatGPT-4 offers general provisions on data handling, including processing and transfer. However, it does not provide detailed instructions on how data should be collected, stored, or disposed of. For example, while the DTA states the need for data to be processed securely, it lacks explicit guidance on the specific methods or technologies to be used, which could lead to inconsistent practices by the parties involved.

Data subject access rights.

Swales et al. emphasise the importance of detailing the rights of data subjects, particularly their ability to access, correct, and delete their personal data. The DTA produced with the assistance of ChatGPT-4 does not comprehensively address the rights of data subjects, particularly in terms of how they can exercise their access rights. While the DTA generally provides for compliance with data protection laws, it does not state the mechanisms for data subjects to access their information.

Appropriate technical and organisational measures are taken, and adequate safeguards are in place.

This best practice standard involves ensuring that robust security measures are implemented to protect the data throughout its lifecycle. The DTA produced with the assistance of ChatGPT-4 mentions the need for security measures but does not specify what those measures should entail. The lack of detailed guidance on the specific technical and organisational measures to be adopted, such as encryption standards or access controls, means that the agreement falls short of this best practice.

Measures in place in relation to cross-border data flows.

Data transfer across borders requires careful consideration of the legal frameworks in different jurisdictions, ensuring that data is protected according to the standards of both the exporting and importing countries. While the DTA produced with the assistance of ChatGPT-4 acknowledges the possibility of cross-border data flows, it does not provide specific measures to ensure compliance with the varying legal requirements of different jurisdictions.

Conditions and restrictions in place in relation to further processing of data beyond.

There should be clear conditions and restrictions on how data can be further processed, particularly by third parties or for purposes other than those initially agreed upon. The DTA produced with the assistance of ChatGPT-4 includes a clause that addresses the further processing of data, specifically restricting the data recipient from processing the data beyond the purposes explicitly stated in the agreement. However, this clause is somewhat general and does not provide detailed conditions or restrictions that should apply if the data is to be further processed. To align more closely with best practices, the DTA could benefit from expanding this clause to include clear, specific conditions under which further processing may or may not occur, as well as detailing any required approvals or notifications for such processing. This would ensure that all parties are ad idem regarding the limitations on data use and help prevent potential legal disputes.

Facilitating compliance with data protection legislation: overall assessment.

The DTA produced with the assistance of ChatGPT-4 meets some general requirements of data protection but does not fully align with the best practice standards outlined by Swales et al. (2024). The lack of detail in several critical areas, including the ground of justification, specific technical and organisational measures, and detailed provisions for cross-border data flows, suggests that the DTA would benefit from further refinement to ensure robust compliance with data protection legislation.

Ethical considerations.

The integration of AI into legal drafting and other legal processes raises several ethical concerns that must be carefully considered. While this discussion is not intended to be comprehensive, it highlights key topics for further exploration.

A primary ethical issue is accountability. When lawyers use generative AI, they must take full responsibility for the final product (Grossman et al. 2023). AI is merely a tool, and the human lawyer using it remains accountable for its output. This principle underscores the importance of maintaining professional responsibility in the AI-assisted legal drafting process.

Another ethical consideration is the potential for bias in AI-generated content. Martin-Bariteau and Pavlović (2021) highlight that the use of AI in contracting amplifies existing power and information imbalances, potentially embedding biases from automated systems into legal agreements. These biases could be inadvertently reflected in AI-generated legal documents, potentially compromising their fairness and accuracy. In the case of the DTA produced with the assistance of ChatGPT-4, no apparent bias was detected. However, lawyers should remain vigilant about this issue and be prepared to mitigate any bias that may arise.

Transparency about the use of AI in legal practice is also a pertinent consideration, as highlighted by authors such as Shope (2021) and Martin-Bariteau and Pavlović (2021). However, whether this should entail disclosing the use of AI to clients is a matter of debate. Grossman et al. (2023) suggest that such disclosure requirements are unnecessary. They contend that existing rules of professional conduct and civil procedure provide sufficient deterrence against misuse of AI, and proactive disclosure may not be required if the lawyer remains fully responsible for the final product. In my view, it is important to consider that lawyers do not typically disclose the tools or human resources they employ, such as legal assistants or document management software. If one is consistent, there does not seem to be any inherent need for proactive disclosure of AI use, especially if AI is being used as a tool to enhance efficiency and the lawyer remains fully responsible for the final product.

The establishment of robust ethical frameworks is essential for guiding the responsible integration of AI into legal practice. Chen et al. (2024) underscore the importance of ethical considerations, including fairness and transparency, for the responsible integration of LLMs in societal domains such as law. These considerations provide a critical foundation for legal professionals to navigate the complexities of AI tools effectively while addressing risks like bias and misuse. By grounding AI adoption in sound ethical practices, the legal profession can harness the benefits of AI without compromising its core values or professional integrity.

Conclusion

This article has examined the role of generative AI, specifically ChatGPT-4, in drafting DTAs for health research, a task that is both complex and critical in the legal domain. The findings of this study reveal that while generative AI shows significant potential in drafting comprehensive legal documents, there are still important limitations that must be addressed.

The two-stage methodology employed—encompassing both contract-level and clause-level prompting—provided key insights into the capabilities and limitations of ChatGPT-4. This methodology’s value lies in its structured approach: the first stage involved generating an overall outline, followed by multiple iterations until saturation was achieved, ensuring that all essential elements of a DTA were covered. The second stage allowed for detailed refinement at the clause level, enabling a thorough exploration of the AI’s ability to produce high-quality legal content. These iterations until saturation are particularly crucial, as they highlight the importance of refining AI outputs to achieve a comprehensive and legally sound document.

Generative AI presents a promising avenue for enhancing the efficiency and foundational preparation of legal documents, making it a valuable tool for legal practitioners, particularly those involved in complex fields like health research. However, the necessity for meticulous human review and adaptation remains paramount. At present, generative AI functions as an instrumental asset within the legal toolkit, but it cannot replace the nuanced judgement and expertise of human lawyers.

Best approaches and future directions.

The future of AI in legal drafting lies in a collaborative approach where AI and human expertise are seamlessly integrated to maximise both efficiency and accuracy. The two-stage methodology highlighted in this study—beginning with an iterative approach to establish a comprehensive outline, followed by detailed clause-level refinement—has proven to be an effective strategy. Legal professionals should adopt this method to harness AI’s strengths while ensuring that the final document adheres to all legal standards and accurately reflects the intentions of the parties involved.

As AI systems continue to evolve, they are likely to become even more embedded in the legal drafting process. However, as Grossman et al. (2023) point out, it is crucial that human lawyers retain full responsibility for reviewing and finalising AI-generated content. AI should be seen as a tool that enhances traditional legal processes, but the ultimate accountability for the legal soundness and ethical integrity of the final product must lie with the human lawyer.

In addition to adopting these collaborative methodologies, the legal profession must also focus on preparing future lawyers to work effectively with AI. Need (2024) emphasised the importance of integrating AI into legal education, arguing that future lawyers must be equipped with the skills to use AI tools effectively while retaining their foundational legal competencies. This approach ensures that while AI can assist in the drafting process, the critical thinking and judgement required to interpret and apply legal principles remain firmly within the domain of human expertise. Wolff (2024) further highlighted the need for legal education to prepare future lawyers for the ethical and practical challenges of using generative AI in practice. This perspective reinforces the necessity of human oversight in AI-assisted legal drafting. It also emphasises the ethical considerations that must be addressed to ensure AI is used responsibly and effectively, particularly in upholding the integrity and reliability of legal documents.

Looking forward, the integration of AI into legal drafting will likely extend beyond mere efficiency gains. As AI becomes more sophisticated, its role may evolve from a drafting assistant to a more collaborative partner in the legal process. However, this evolution must be managed carefully to ensure that the ethical standards of the legal profession are upheld. Continuous education, ethical oversight, and a strong emphasis on human judgement will be essential as the legal profession navigates this transformative period.

Supplementary Material

Supplementary information The online version contains supplementary material available at https://doi.org/10.1057/s41599-025-04643-z.

Data availability

The data generated and analysed during this study, namely the DTA, is included as a supplementary information file to this article.