Sora (OpenAI) and GDPR / AI Act: generative video and deepfakes
In brief
- Sora (OpenAI) produces realistic videos from text, with 'cameo' (integration of people). It is the prototype deepfake risk: never depict a real person without documented consent. OpenAI applies provenance metadata (C2PA) and a watermark, but the deployer remains responsible for visible disclosure (article 50). Frame transfers (CLOUD Act) and set out prohibited uses (deepfakes of executives/clients) in the AI charter.
Overview and data processed
Sora is OpenAI's video generation model: from a text description (or images), it produces realistic video sequences. Certain features allow the integration of people's likenesses.
Sora processes the prompts and images/videos provided, and can generate identifiable people. Integrating the likeness of a real person involves their image and voice. US-based provider: exposure to the CLOUD Act. OpenAI applies provenance metadata (C2PA) and a watermark.
Risks for your organisation
The major risk is deepfake: a video depicting a real person without consent, or deceptive content appearing authentic. In addition there are transfers outside the EU and copyright on generated content.
Main GDPR risks:
- Representation of real people without consent (right to image, voice as personal data).
- Transfer outside the EU (CLOUD Act); prompts containing personal data.
- Origin of training data (scraping prohibited); copyright.
- Dissemination of deceptive content (disinformation).
GDPR obligations
- Never depict a real person without consent; document authorisations.
- Sign the DPA (business plans); frame transfers (SCC); do not enter unnecessary personal data in prompts.
- Enter in the record of processing activities; DPIA if high-risk use.
- Inform the persons depicted.
- Involve the DPO; train teams. No DPO? See our outsourced DPO.
AI Act obligations
Article 50 (applicable from 2 August 2026, technical deadline 2 December 2026): the provider marks technically; the deployer who produces a deepfake must disclose clearly. Lighter regime for artistic/satirical content. GPAI; AI literacy. Prohibition (article 5): untargeted facial scraping prohibited. See the AI Act guide.
AI Act risks: undisclosed deepfakes (article 50); removal/alteration of provenance markings; deceptive content on matters of public interest.
Cybersecurity and good practices
Cyber risks: identity theft via generated video, disinformation and brand damage, transfer outside the EU; NIS 2 depending on size.
Good practices: prohibit the representation of real people without documented consent; preserve and display provenance markings (C2PA/watermark), visible disclosure; frame transfers; set out prohibited uses in the AI charter; editorial control before dissemination; record, training.
Checklist:
- No representation of real people without consent, provenance marking preserved and displayed, visible disclosure of deepfakes
- DPA + transfers framed (SCC), no unnecessary personal data in prompts, DPIA if high-risk use
- AI charter (prohibited uses), editorial control before dissemination, teams trained
Sanctions, conclusion and recommendations
Sanctions: GDPR (up to €20M or 4%); AI Act (€15M or 3% for article 50, €35M or 7% for prohibited practices); civil and criminal — right to image, article 226-8 and 226-8-1 of the criminal code.
Recommendation: Sora puts deepfake in the spotlight: its power demands strict discipline. Never the likeness of a real person without consent, preserve and display provenance markings, disclose clearly and frame transfers. Set out prohibited uses in the AI charter. Compare with Veo, Runway and Pika. See the AI tools comparison.
Take action
Assess the GDPR and AI Act risk level of your Sora use cases with the Data Comply One self-assessment.
Discover the AI Act software →Test my compliance for free