On May 9, 2026, Saudi Arabia's Data and Artificial Intelligence Authority (SDAIA) published Deepfakes Guidelines: Mitigating Risks While Fostering Innovation — a non-binding framework that, instead of reaching for a ban, sorts synthetic media into malicious and beneficial uses across six sectors: marketing, entertainment, retail, education, healthcare, and culture. Developers are asked to embed digital watermarks, keep model documentation, and maintain human-oversight mechanisms; content creators are asked to apply visible, tamper-resistant watermarks, secure explicit consent, and hold auditable records, all aligned with the Kingdom's Personal Data Protection Law (PDPL). It is the newest layer in an AI-governance stack — PDPL, the AI Ethics Principles, the Generative AI Guidelines — that SDAIA is assembling under the Vision 2030 banner of a self-declared "Year of AI 2026."
On the merits, the design is good
The strongest case for this kind of instrument is that it avoids the two failure modes regulators usually fall into with generative media. The first is the prohibition reflex — outlawing "deepfakes" as a category, which sweeps in the ALS patient whose voice is reconstructed, the dubbed film, and the AI tutor alongside the fraudster. The second is paralysis. SDAIA's framework does neither. It treats synthetic media as dual-use technology and regulates the harm — non-consensual likeness, impersonation, fraud — rather than the method.
That is the approach technologists themselves recommend. Watermarking and content provenance — the same lineage as Adobe's Content Authenticity Initiative and the C2PA standard — keep a verifiable signal of origin without dictating what may be made. Tying obligations to consent and to the existing PDPL, rather than inventing a parallel enforcement regime, is exactly the proportionate, evidence-based posture we favor. As the CMS AI-regulation guide notes, Saudi Arabia has deliberately chosen "soft-law mechanisms over statutory intervention," with the PDPL the only binding instrument in the stack. For a frontier-moving technology whose use cases are still being discovered, declining to freeze the rules into statute is a defensible bet.
The soft-law bet has a catch
Non-binding guidance lives or dies on the credibility of the institution issuing it and the legal environment around it. Voluntary watermarking works when a critical mass of actors expects to be caught and sanctioned for the genuinely harmful uses — and when the people the technology is used against can speak freely about it. Both conditions depend on something the guidelines cannot supply: a state that treats online expression as a right rather than a risk to be managed.
This is where the Kingdom's own conduct intrudes. According to a May 20, 2026 statement coordinated by Access Now and Gulf-focused NGOs, Meta began rendering the Facebook accounts of human-rights groups ALQST and Democratic Diwan, and of researchers including Abdullah Alaoudh and defender Yahya Assiri, "unavailable" inside Saudi Arabia from April 30, 2026, at the Saudi government's request — geo-blocking, not removal. Meta's own content-restriction reporting, the statement says, shows more than 100 pages and accounts restricted since March 2026. The same playbook was attempted on X against prominent Saudi activists, though as of May 20 X had not complied.
The point is not that watermarking guidance and account geo-blocking are the same policy. They are not. The point is that they emerge from the same authority structure. A government willing to lean on platforms to make critics invisible is a government for whom provenance and authentication tooling is dual-use in a second, less advertised sense: the infrastructure that proves a video is real can equally be turned to proving a dissident's identity, and the consent-and-records regime that protects a citizen's likeness can be selectively enforced against the people the state already disfavors.
Rhetoric and enforcement are not yet aligned
The gap is visible in individual cases. The Electronic Frontier Foundation relaunched a campaign on May 12, 2026 for Osama Khalid, a Saudi Wikipedia contributor imprisoned over his online writing. A framework that speaks fluently about "safeguarding public trust" in AI sits uneasily beside the continued jailing of a volunteer encyclopedist for ordinary speech. Trust is not a property of a PDF; it is earned by how a state behaves toward its most inconvenient users.
None of this argues for abandoning the model. Sector-differentiated, harm-focused, provenance-based rules are precisely what other jurisdictions — tempted by blunt synthetic-media bans — should study. The financial backdrop is real: Saudi Arabia ranked 14th in the 2025 Global AI Index, government spending on emerging technology rose more than 56% in 2024, and domestic AI firms raised some $9.1 billion, per Arab News. A country investing at that scale has every incentive to get the governance architecture right, and on the technical question of how to regulate synthetic media, SDAIA largely has.
What proportionate governance actually requires
Proportionate regulation is not only about calibrating burdens on developers. It is about the symmetry between what a state asks of technology companies and what it permits of itself. The deepfake guidelines ask creators for consent, transparency, and auditable records. The same standards — disclosure, due process, a right to be heard — should govern the state's own demands to geo-block accounts and prosecute speech. Until that symmetry exists, the Kingdom's AI stack will be read, fairly, as a well-engineered facade over an unreformed approach to expression. The framework deserves credit; the credibility has to be earned elsewhere.