Saudi Arabia’s AI Training Exception: A Practical Guide to Article 26

Introduction: What AI Training Means for Copyright

AI training involves feeding large volumes of text, images and other data into computer systems so they can learn patterns and generate outputs. That process typically involves copying the works in the training material. Saudi Arabia’s newly introduced Article 26(4) ¹ permits certain copying for the development of AI products and algorithms, subject to conditions that protect rights holders the people or organisations that own or control copyright. It is a copyright exception: a legal permission to use a protected work without a separate licence, but not a blanket permission. The provision provides a clearer route for innovation while raising practical questions about which works can be used, how they are obtained and how much may be copied. The next section breaks those questions into the statutory conditions and the safeguard that applies alongside them.

Article 26(4): Three Conditions for Copying in AI Development

Article 26(4) sets out a specific exception for copying an original work to develop AI products and algorithms. Its text is: “Copying the original work for the purposes of developing artificial intelligence products and algorithms, provided that the work has been published legitimately, that the ownership of the original copy is legitimate, and that the copying is within the limits of what meets the purpose.” ¹ the provision is best understood as imposing three cumulative conditions:

1. Legitimate publication. The work must have been published legitimately. In practical terms, developers should examine whether the work was made available through a lawful route; mere availability online does not by itself establish this condition. The provision does not say that the work must be registered as copyright, and Article 26(4) should not be read as imposing a registration requirement. ¹
2. Legitimate ownership of the original copy. The ownership of the original copy must be legitimate. A developer should be able to identify how it obtained the original copy for example, through a lawful purchase or other authorised transfer and keep records supporting that route. This focuses on the legitimacy of the copy and its acquisition, not simply on whether the material was easy to download. ¹
3. Copying limited to the development purpose. The copying must remain within the limits of what meets the purpose. Developers should be able to explain why the work was copied and why the amount copied was reasonably connected to developing the relevant product or algorithm. Article 26(4) does not expressly limit the exception to non-commercial projects, but an AI-development purpose does not justify unrestricted reproduction, dissemination or other uses outside that purpose. The broader rights and licensing framework in Articles 7–12 remains relevant to any use not covered by the exception. ³

Article 37(1): An Additional Safeguard on the Use of Works

These three Article 26(4) conditions are not the end of the analysis. Article 37(1) operates as an additional safeguard on top of them: in the cases covered by Articles 26 to 36, third-party use must not contradict the work’s normal exploitation that is, the ordinary market in which the work is commercially used and must not cause undue harm to the legitimate interests of the rights holders, meaning their legitimate economic or other copyright-related interests.² In our view, this means that a developer may satisfy the three Article 26(4) conditions and still need to assess market impact and harm. The practical effect is likely to depend on the material used, the way it was obtained and the particular development activity.

Practical Example: Training an Arabic Research Tool

Consider a developer training a specialist Arabic research tool on a collection of published articles. First, it should verify that each article was published legitimately¹; being available online is not enough. Second, it should verify that it has legitimate ownership of the original copy, including a lawful route of acquisition, and keep evidence of that route. Third, it should assess whether copying each article in full is within the limits of what the tool’s development requires, or whether more limited copying would serve the same purpose. Finally, even if those three conditions are met, Article 37(1) requires the developer to consider whether use of the articles could interfere with the publisher’s normal market or cause undue harm to the rights holders’ legitimate interests. ² the exception’s application therefore depends on the facts, and a licence may be the clearer route where any step remains uncertain.

Saudi Arabia, the UAE and the EU: Different Routes for AI-Related Copying

Saudi Arabia’s approach differs from the UAEs. Article 22 of UAE Federal Decree-Law No. 38 of 2021 on Copyright and Neighbouring Rights sets out limitations and exceptions, but those exceptions do not expressly address AI training. ⁴ That does not by itself resolve every possible use; the relevant activity and rights involved still need to be assessed. Saudi Arabia is more direct because Article 26(4) expressly refers to copying for the development of AI products and algorithms, while also imposing conditions on the work used and the extent of copying. ¹

The EU takes a different route through text and data mining (TDM), meaning automated techniques that analyse large amounts of digital text and data to identify patterns, trends or other information. Directive (EU) 2019/790, the Digital Single Market Directive, gives qualifying organisations a TDM exception for scientific research in Article 3, where the relevant conditions are met.⁵ Article 4 provides a broader TDM exception, but allows rights holders to reserve or opt out of that use.⁵ Separately, Article 53(1)(c)–(d) of Regulation (EU) 2024/1689 (the EU AI Act) requires providers of general-purpose AI models to maintain a copyright-compliance policy and publish a sufficiently detailed summary of training content; those transparency duties are separate from the copyright exception.⁶

Conclusion: What Developers Should Do Next

Saudi Arabia’s approach gives AI developers a clearer potential route to use copyrighted works, but it is not a free pass. Its significance lies in pairing an express AI-development exception with conditions concerning legitimate publication, legitimate ownership of the original copy and purpose-limited copying, together with Article 37(1)’s additional protection for normal exploitation and rights holders’ legitimate interests.¹ ² Developers should build those questions into dataset governance: document where each dataset came from, how the original copies were obtained, why the amount copied is needed, and whether the use could affect an existing market or unduly harm rights holders. Where the position remains uncertain, seek a licence or narrow the use. As Saudi Arabia’s approach is tested in practice, developers that can show disciplined, purpose-limited and well-documented use will be better placed to support innovation while respecting the works on which AI development depends.