Saudi Arabia's new Copyright Law took effect on August 12, 2026, replacing a 2003 statute that had nothing to say about machine learning. Buried in the replacement is a provision that will matter more to the Kingdom's AI ambitions than almost anything else in the text: a statutory exception permitting reproduction of lawfully published copyrighted works to train AI systems, without the rightsholder's permission and without compensation. Saudi Arabia has effectively legislated its way into having one of the most permissive AI-training regimes in the world — and unlike the jurisdiction it most resembles, the European Union, it did so without giving creators any way to say no.
What the Law Actually Permits
The exception, detailed in the law's new-uses provisions, is not unconditional. Legal analysis of the text from Baker McKenzie identifies five safeguards: the work must have been lawfully published, the copy must have been lawfully acquired, the reproduction must be limited to what serves the AI-development purpose, the use must not conflict with normal exploitation of the work, and it must not cause unreasonable prejudice to rightsholders' legitimate interests (Baker McKenzie). Gowling WLG's summary of the same provision frames the operative test more simply: lawful publication, lawful acquisition, and copying "limited to what is necessary for the intended purpose" (Gowling WLG). Either reading leaves the same open question — what counts as "necessary" — for the Saudi Authority for Intellectual Property (SAIP) to answer in an implementing regulation that, as of this writing, still hasn't been published, even though the law itself is already in force (SAIP).
The Case Against the Exception, Fairly Stated
Rightsholders have a real grievance here, and it deserves more than a dismissal. A regime with no opt-out mechanism means a Saudi novelist, journalist, or software vendor has no lever to keep their published work out of a competitor's training set — not even the limited, machine-readable reservation that EU law provides. The European Union's Copyright in the Digital Single Market Directive built exactly that lever into its own text-and-data-mining exception: Article 4 lets rightsholders reserve their works from TDM use, provided the reservation is expressed in an appropriate, typically machine-readable, format. A. Its scope has been contested — a recent ruling found that the AI Act's own reference to Article 4 confirms EU lawmakers intended AI training to count as TDM — but the reservation right itself is not in dispute. Saudi Arabia's law contains no equivalent, a gap A&O Shearman's analysis flags directly, noting the framework provides no opt-out or contractual waiver mechanism for rightsholders at all (A&O Shearman). For a Kingdom trying to build a domestic publishing, media, and software economy alongside its AI sector, that's not a small omission — it hands foreign and domestic AI labs a resource local creators can't withhold.
Why the Bet Is Still Defensible
The steelman for creators is real, but the EU's alternative isn't obviously working better. Two years into implementation, the opt-out mechanism has generated exactly the kind of interpretive litigation Riyadh appears eager to avoid: what counts as a machine-readable reservation, whether it binds AI training the same way it binds conventional data mining, and how a developer scraping billions of documents is supposed to verify compliance at scale. Saudi Arabia, in choosing a clean statutory carve-out over a contestable reservation right, is making the same wager Singapore and Japan made with their own TDM exceptions: that legal certainty for developers is worth more, in growth terms, than a control right for rightsholders that is hard to enforce and easy to litigate around anyway. Baker McKenzie's analysis calls this one of the first statutory AI-training exceptions in the Gulf Cooperation Council — a first-mover claim that lines up with the Kingdom's decision to designate 2026 the "Year of Artificial Intelligence," a designation the Ministry of Communications and Information Technology's own site reflects in its current AI programming (MCIT). That timing is not incidental. SDAIA's national AI strategy and the tens of billions of dollars in committed data-center and compute investment only pay off if models can actually be trained on Arabic-language and regional content at scale, and licensing that corpus work-by-work was never realistic.
What SAIP Still Owes the Market
None of this makes the current gap acceptable. A five-factor "necessary for the purpose" test with no published guidance, no opt-out, and a fine for ordinary infringement that just quadrupled to SAR 1 million is a lot of legal exposure to leave undefined (Gowling WLG). SAIP was expected to issue the implementing regulation within 180 days of the law's approval — a window that lines up almost exactly with the August 12 entry-into-force date (Baker McKenzie). It has not appeared. Every week that gap persists, both AI developers relying on the exception and rightsholders trying to understand what they've lost are operating on inference rather than rule text. A broad exception without an opt-out is a legitimate policy choice for a country racing to build an AI sector from a smaller creative-rights base than the EU's. Leaving "necessary for the intended purpose" undefined for months after the law already binds everyone is not — SAIP should close that gap before the ambiguity becomes the norm.