This article is general information, not legal advice. Duties depend on jurisdiction, the creator’s role, the type of content and publication context.
European Union: Article 50
The European Commission says the transparency obligations begin applying on August 2, 2026. Providers of certain generative systems must support machine-readable marking of generated or manipulated content. Separate duties apply to deployers — people or organizations using a system professionally under their authority.
People must be informed when exposed to deepfakes, and a separate rule covers AI-generated public-interest text published without human editorial review. This is not phrased as “every blemish removal needs a prominent badge.” Definitions, materiality, role and applicable exceptions still matter.
United States: multiple layers, not one AI label law
The federal TAKE IT DOWN Act became law in May 2025. It addresses nonconsensual publication of intimate images, including digital forgeries, and requires covered platforms to operate a removal process. It targets a specific harm and consent problem rather than imposing a universal label on every AI-assisted file.
California’s AI Transparency Act became operative January 1, 2026 and principally regulates large generative-system providers, including detection tools and required disclosures. California AB 2655 separately addresses materially deceptive election-related content on large platforms during specified periods.
Generation, material alteration and basic retouching
Legal significance usually turns on the result and context, not merely whether machine learning ran. Voice denoising is not the same as synthetic audio making a person say invented words. Removing a temporary blemish is not the same as replacing an identity or creating an event that never occurred.
- Realistic deepfakes of people or events carry higher disclosure and liability risk.
- Political ads, elections, public-interest claims and intimate content receive stricter treatment.
- Commercial claims may still violate ordinary deception rules regardless of AI use.
- Instagram, TikTok or ad-network rules may be stricter than applicable law.
What can remain private
A duty to tell an audience that content is synthetic does not necessarily require publication of a full prompt, seed, filename, GPS position or internal workflow. Data minimization remains sensible: provide the required disclosure without exposing unrelated personal or commercial information.
Metadata cleaning cannot be used to erase a disclosure duty. Visible notices, platform settings and machine-readable markers solve different problems. For a high-risk campaign, document jurisdiction, organizational role, material edits and legal review before publishing.
A practical pre-publication matrix
Ask four questions: does the content depict a real person; does it make them do or say something that did not happen; does it concern elections, public interest, advertising or intimate material; and are you acting professionally for an organization? The more yes answers, the stronger the case for explicit disclosure and legal review.
Then inspect the file. C2PA and EXIF help document technical history, but the ultimate legal duty comes from law and context, not from a single metadata field.