Efforts to combat deepfakes have primarily focused on detecting manipulated content after it has been created. ETSI's new specification takes a proactive approach, treating deepfake prevention as a security and governance challenge that must be addressed throughout the AI system's lifecycle. Building on the baseline cybersecurity requirements set out in ETSI’s EN 304 223 standard for AI models and systems, the new specification helps organisations to reduce the risk of AI systems creating harmful deepfakes. This also supports objectives laid out in the EU AI Act, extending security beyond systems classified as high-risk.

ETSI’s TC SAI outlines the following measures:

• Risk assessments before deployment: Identifying potential misuse of generative AI systems and implementing safeguards before they are released.

• Human oversight: Establishing processes for human review and intervention, ensuring that organisations retain control over system behaviour.

• Safeguards against misuse: Implementing measures such as content filtering, AI-generated-content labelling and user warnings where appropriate.

• Ongoing monitoring: Reviewing emerging threats, investigating incidents and updating safeguards as new risks are identified.

• Evidence-based assessment: Maintaining documentation and records that demonstrate whether appropriate controls have been implemented and are operating as intended.

Alongside these safeguards, the specification also outlines test scenarios and pass/fail criteria. These practical measures allow organisations to demonstrate whether reducing deepfake-related risks have actually been met, rather than relying on commitments to responsible AI development.

"Harmful, misleading and illegal deepfake material such as abusive images or fraudulent product sites can be created from the misuse of legal and commercially available software,” said Alex Cadzow, Rapporteur and Secretary of ETSI TC SAI. “As deepfakes become more pervasive, AI organisations need support in addressing the risks much earlier, starting with how generative AI systems are designed. This specification provides organisations with a clear set of requirements to help reduce those risks, alongside ways to assess whether the necessary safeguards are in place.”