They should require consent evidence, anonymisation where needed and approval from the relevant business and privacy owners before publication. Synthetic media that uses identifiable people creates both content and identity risk, so the control model must cover lawful basis, provenance and downstream use.
What makes AI-generated media with real people different from ordinary content review?
AI-generated media involving real people is not just a branding or editorial question. It can create a misuse path for likeness, voice, biometric-style attributes and personal data, so the review process needs to decide whether the material is permitted, disclosed, approved and traceable before it reaches an audience.
That means the control point is upstream of publication. Teams should treat the asset as something that can affect privacy, reputation, impersonation and downstream reuse, not simply as a creative output that can be posted once it looks convincing.
The key operational question is whether the media can be confidently tied to a lawful and authorised purpose. If that cannot be shown, the default should be to hold publication until the evidence, approvals and provenance are complete.
What governance controls should sit around consent, provenance and approval?
A workable governance model starts with explicit consent evidence for any identifiable person, plus a documented business justification for why the media should exist at all. Where identification is not necessary to the purpose, anonymisation, masking or other minimisation should be used before approval.
Approval should not sit only with the team creating the asset. The relevant business owner should confirm the intended use, while the privacy owner should validate lawful basis, retention, disclosure and reuse boundaries. That separation matters because one team usually owns the content outcome, but another owns the risk created by publishing it.
For controls that touch provenance and disclosure, use a standard review record that captures source material, transformation steps, approvals and any restrictions on reuse. For media sanitation and disposition decisions, NIST SP 800-88 Media Sanitization is useful where source assets or derivatives need controlled disposal, even though the subject here is publication governance rather than storage cleanup. For the privacy and lawful-use dimension, the NIST Privacy Framework supports a structured view of data use, disclosure and risk management.
When synthetic media is part of a broader AI programme, the governance standard should also connect to AI policy and lifecycle controls. The NIST AI Risk Management Framework and the ISO/IEC 42001:2023 AI Management System Standard both help formalise accountability, approval discipline and traceability for AI outputs that can affect people.
How should teams handle publication, reuse and downstream abuse risk?
Publishing is the point where small mistakes become durable exposure. Once realistic media of a real person is released, it can be copied, repurposed, altered and detached from the original context, so the team should define where it can be used, whether it can be redistributed, and what watermarking, disclosure or platform labeling is required.
Teams should also assume that the same asset may be reused in contexts the original approver never intended, including internal comms, social channels, customer-facing campaigns or hostile impersonation. That makes publication control a governance issue, not a purely creative one.
If the media could reasonably be mistaken for authentic, the organisation should decide whether the goal is likeness, parody, illustration or simulation, because each carries different tolerance for confusion. The more lifelike the output, the stronger the need for identity verification, legal review and provenance tracking before release.
For broader AI deployment governance, the NIST AI 600-1 GenAI Profile is directly relevant because it addresses generative AI governance, provenance and pre-deployment testing. The EU AI Act regulatory framework is also relevant where legal obligations, transparency and deployer duties shape how synthetic media involving real people can be released.
Risk and Threat Considerations
AI-generated media involving real people can create identity abuse, deceptive impersonation and privacy exposure if consent, provenance or approval are weak. The main failure mode is not only a policy breach, it is uncontrolled reuse of a convincing likeness or voice in a context the subject never approved.
Failure mechanism: Teams publish or circulate synthetic media without verified consent, clear lawful basis or change control, which allows realistic content to escape into channels where it can be mistaken for authentic or used outside the intended purpose.
Impact: The organisation can create reputational harm, privacy complaints, legal exposure and downstream impersonation risk, especially when the media is easy to copy, edit or repackage.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
NIST SP 800-53 Rev 5 and NIST AI RMF set the technical controls, while ISO/IEC 42001:2023 and EU AI Act define the regulatory obligations.
| Framework | Control / Reference | Relevance |
|---|---|---|
| NIST SP 800-53 Rev 5 | MP-6 — Media Sanitization | Covers controlled disposal and purging of source media and derivatives. |
| PL-4 — Rules of Behavior | Sets expectations for approved use, sharing, and publication of synthetic media. | |
| AU-2 — Event Logging | Supports traceability for who approved and released synthetic media. | |
| Recommendation — Apply MP-6 to purge source assets and retain only approved derivatives. Define publication and reuse rules for AI-generated media. Log approvals and release actions for synthetic media workflows. | ||
| NIST AI RMF | GOVERN — GOVERN | Directly supports AI accountability, policy, provenance and oversight for generated media. |
| MAP — MAP | Helps map intended use, affected people and disclosure boundaries for the media. | |
| MANAGE — MANAGE | Supports risk treatment and control selection for real-person synthetic media. | |
| Recommendation — Establish accountable approval and provenance controls for generative media. Map stakeholders, intended use and disclosure limits before release. Manage residual risk with documented controls and review gates. | ||
| ISO/IEC 42001:2023 | 4.1 — Understanding the organization and its context | Frames organisational context and responsibilities for AI content governance. |
| 8.2 — AI risk treatment | Supports treating consent, provenance and misuse risks before deployment. | |
| 9.1 — Monitoring, measurement, analysis and evaluation | Enables review evidence and control effectiveness tracking for approvals. | |
| Recommendation — Define the organisational context for synthetic media decisions. Treat synthetic media risks before publication or reuse. Measure approval quality and control effectiveness over time. | ||
| EU AI Act | Transparency and governance obligations for AI systems | Regulates transparency, accountability and deployment of generative AI outputs affecting people. |
| Recommendation — Apply transparency and governance obligations to real-person synthetic media. | ||
Practitioner Guidance
What to verify: Before publication, verify that the subject’s consent is documented, the lawful basis is known, and the media owner can explain why identification was necessary rather than optional. If the answer is unclear, hold the asset rather than trying to compensate with disclaimers after release.
Decision rule: If the content uses a real person’s face, voice or other identifiable attribute, require business-owner approval and privacy-owner approval together, not one or the other. If the asset is not needed to identify the person, anonymise or transform it before it enters the approval queue.
Practitioner takeaway: Treat realistic synthetic media as a governed identity-and-privacy artefact, not a design asset, because the highest-risk mistake is approving distribution before consent, provenance and reuse limits are explicit.