Home › News

Navigating Global AI Rules for Online Deepfakes

Published: 12.09.2026

When a developer deploys a model that allows a user to create a deepfake online with a few text prompts, a precise structural question arises: which jurisdiction’s rules govern that output? The underlying technology operates with borderless indifference, yet the legal frameworks constraining it are rigidly territorial. This disconnect between synthetic media generation and regulatory perimeters forms the central tension in governing artificial intelligence today. Understanding global standards is less about finding a single universal statute and more about mapping the overlapping, sometimes contradictory, obligations that apply when synthetic media crosses borders.

Disentangling general AI governance from synthetic media rules

Regulatory debates frequently conflate two distinct categories: general AI safety—concerning systemic risk, bias, and model autonomy—and the specific harms of synthetic media. Deepfakes represent a subset of AI output where the primary risk is not emergent machine behaviour, but human deception. The regulatory response therefore splits. General AI frameworks, such as the risk-based tiering in the European Union, treat foundational models as infrastructure. Synthetic media rules, by contrast, target the point of deployment and the specific intent of the end user.

This distinction matters for compliance. A platform hosting a model must satisfy structural requirements for model transparency and data provenance, while simultaneously navigating output-level constraints that govern how the generated media may be used. Failing to separate these layers leads to misapplied compliance strategies, where platforms over-index on model documentation while neglecting the distinct legal liabilities of the generated deepfake itself.

The fragmented legal perimeter

There is no unified global standard. Instead, operators face a patchwork of regional frameworks that categorise the harm—and the corresponding duty—differently.

The European Union: transparency as a baseline

The EU AI Act remains the most comprehensive regional framework. For synthetic media, it mandates transparency obligations. Providers of systems that generate synthetic audio, image , video, or text must ensure the output is machine-readable as artificially generated. Crucially, the obligation extends to deployers: anyone using a deepfake must disclose that the content has been artificially created or manipulated. The exception is content generated for obviously artistic, creative, or satirical purposes, provided the right of freedom of expression is respected. This carve-out introduces significant interpretive ambiguity, shifting the burden onto platforms to define the boundary between satire and deception.

The United States: a state-level patchwork

In the absence of a federal statute specifically targeting synthetic media, US regulation operates through state-level interventions, primarily targeting the harm rather than the technology. Several states have criminalised non-consensual intimate deepfakes, while others have enacted election-specific disclosure requirements for synthetic media. The regulatory focus is ex-post (addressing the harm after distribution) rather than ex-ante (preventing the generation at the model level). For a service that enables users to create deepfakes online, this means the legal exposure is determined by the jurisdiction of the victim or the electoral district, not the jurisdiction of the server.

Asia-Pacific: strict provenance and targeted restriction

China’s deep synthesis regulations establish a rigid ex-ante framework. Providers must label all synthetic content visibly and maintain algorithmic registries with the state. Furthermore, services supporting deep synthesis must verify the real identity of users. South Korea has amended electoral laws to criminalise the distribution of deepfakes related to candidates within a strict timeframe before elections. These frameworks treat synthetic media as an inherent threat to social stability, minimising artistic carve-outs and prioritising state-level provenance tracking.

Technical dependencies and enforcement exceptions

Legal mandates for disclosure and labelling depend entirely on the viability of technical standards. A regulatory requirement that a deepfake be "machine-readable" as artificial presupposes an agreed-upon standard for watermarking or metadata embedding. Currently, the Coalition for Content Provenance and Authenticity (C2PA) provides the most robust architecture for content credentials, binding cryptographic provenance data to the media file.

However, a critical dependency exists: metadata fragility. C2PA credentials survive platform re-uploads and basic compression, but they are stripped by simple screenshots or transcoding workflows not integrated into the C2PA ecosystem. Regulations that mandate labelling without accounting for the technical ease of stripping metadata create an enforcement gap. A user can generate a compliant deepfake, take a screenshot of the output, and distribute an identical, non-compliant version. The regulatory exception for bad-faith actors is that they simply ignore the technical compliance layer, leaving the framework to penalise only those who comply imperfectly, rather than those who circumvent entirely.

Practical checks for platform operators

For any service providing the capability to create deepfake online, navigating this fragmented landscape requires operationalising the legal constraints. The following checks provide a practical baseline for compliance:

• Geofencing capability over content: Because the legality of specific deepfake categories (such as political satire or non-consensual imagery) varies drastically by jurisdiction, platforms must implement geo-specific access controls. It is insufficient to apply a global moderation policy; the generation capability itself must be restricted based on the user's jurisdiction and the subject classification of the target media.
• Provenance embedding at generation: Do not rely on post-hoc labelling. Integrate C2PA-compliant provenance signing directly into the model's output pipeline. Even though metadata can be stripped, demonstrating that the platform embedded robust credentials at the point of creation satisfies the ex-ante transparency obligations under the EU AI Act and Asian regulations.
• Intent-classification prompts: Before generation, implement a classification step that queries the use case. While users may lie, forcing a declaration of intent (e.g ., parody, educational, personal) creates a necessary audit trail. If a user declares satire but the output targets an election in a restricted jurisdiction, the system can flag the discrepancy for human review before rendering the output.
• Identity verification for high-risk synthesis: Align with jurisdictions that require deployer accountability. Implementing tiered access—where generating a face-swap of a public figure requires authenticated identity, while generating a stylised avatar does not—balances friction with regulatory compliance.

The limitation of voluntary commitments

Major model providers frequently announce voluntary commitments to develop responsible synthetic media tools, often focusing on red-teaming and safety filters. While valuable, these commitments are structurally insufficient. They lack the permanence of statute, can be revised unilaterally, and fail to address the open-source proliferation of fine-tuned models. Once a base model is released, community fine-tuning can strip safety filters, enabling users to create deepfakes online outside the perimeter of the original provider's voluntary framework. Regulation must therefore target the distribution infrastructure—hosting platforms and app stores—rather than relying solely on the self-governance of foundation model developers.

Designing for the strictest common denominator

The absence of a singular global standard does not permit inaction. Because synthetic media is inherently cross-border, the most restrictive jurisdiction a platform's output touches will often dictate the enforcement posture. Operators must design systems that satisfy the strictest transparency and provenance requirements—currently those mandated by the EU and China—and treat the artistic and satirical carve-outs as jurisdiction-specific relaxations of that baseline, rather than default permissions. The practical reality is that compliance is an engineering problem of dynamic policy application, requiring systems that evaluate the user, the subject, and the jurisdiction before a single pixel is generated.