EU AI Act deepfake labeling loosens in final guidance; implications for AI builders
Published by AINave Editorial • Reviewed by Ramit
The European Commission released final draft guidance on the EU AI Act's transparency rules, softening certain deepfake labeling requirements and creating a more nuanced compliance path for AI-generated content. The guidance, published on July 20, gives companies just two weeks to prepare before the rules take effect on August 2.
What happened
The final guidance clarifies which AI-generated content must be labeled under the EU AI Act. Under the Act, content that "would falsely appear to a person to be authentic or truthful" must carry a mark identifying it as AI-generated or manipulated. A separate code of practice sets out how companies can follow these rules.
The key change: replacing backgrounds "for clearly aesthetic purposes" would likely only have a "minor impact" on whether viewers perceive an image as authentic, potentially exempting such images from mandatory labeling. An earlier draft only exempted extending or editing background details.
The guidance also softened broadcast labeling rules. The earlier draft required deepfake labels "also at later stages or persistently" and stated that an end-credits warning was insufficient. The final version says labeling is needed only "where possible" and "as appropriate" and is silent on end credits.
Why AI builders should care
For teams building AI-powered advertising, content creation, or media tools, the exemptions reduce labeling friction for common use cases. Aesthetic background swaps in product photos or social media images may not require a label, which matters for advertisers who argued that excessive labeling would cause "labeling fatigue" and reduce its effectiveness.
However, the Commission also noted that advertisements should still not mislead audiences about a product's actual characteristics. Builders should ensure their AI editing tools don't produce deceptive content even if labeling isn't required.
Practical implications
| Scenario | Labeling requirement |
|---|---|
| Aesthetic background replacement in an image | Likely exempt under the new guidance |
| Broadcast deepfake warning at start of program | Required |
| Broadcast deepfake warning at later stages or end credits | Required only "where possible" and "as appropriate" |
| AI-generated content that could mislead about product characteristics | Must be labeled |
Compliance begins August 2, but enforcement will vary significantly by member state. Several capitals have not yet decided which national bodies will oversee implementation, nearly a year after their own deadline for doing so passed. This means the practical risk of enforcement may differ across EU markets.
Caveats
The guidance is final draft guidance, not the full legal text. National authorities may interpret the exemptions differently. The "where possible" and "as appropriate" language for broadcast labeling leaves room for interpretation, and the exact thresholds for what counts as "aesthetic" background replacement are not specified in the cited guidance. Builders should monitor how their specific member state implements enforcement before relying on these exemptions.
This analysis is based on a single source article covering the guidance release. The full guidance document may contain additional details not captured here.
