What constitutes a Deep Fake? The blurry line between legitimate processing and manipulation under the EU AI Act

Kristof Meding, Christoph Sorge

Introduction

As generative AI becomes a standard feature in consumer electronics, the European Union’s AI Act (AIA) introduces strict transparency requirements for "deep fakes." The goal is to protect the "objective truth" of media by ensuring that manipulated content is clearly labeled. However, a critical challenge remains: the legal definitions within the AIA are technically ambiguous. This paper explores the "blurry line" between routine digital image processing and regulated manipulation, highlighting how modern smartphone features—such as Samsung’s moon photography and Google’s "Best Take"—create significant compliance hurdles for developers.

The Myth of the "Original" Photo

A fundamental misconception in the legal discourse surrounding the AIA is the existence of an "authentic" or "original" digital photo. In reality, no digital image is an objective depiction of reality. The path from raw sensor data to a viewable image involves complex interpolation, noise reduction, demosaicing, and tone mapping. These steps are necessary to make sensor data intelligible to humans, yet they involve AI-like processing.

When does this necessary processing cross the threshold into "manipulation"? The AIA struggles to define terms like "authentic" and "substantial alteration," leaving developers in a state of legal uncertainty regarding whether their standard camera software constitutes a deep fake generator.

Case Study: AI-Enhanced Photography

The paper evaluates specific commercial features against the criteria set out in Article 3(60) and Article 50 of the AIA. A primary example is Samsung’s "Scene Optimizer," which enhances moon photographs.

Comparison of moon photographs with and without Samsung's Scene Optimizer.

As shown in the figure above, the AI-enhanced version (b) provides detail that the raw sensor data (a) cannot capture. While this creates a more pleasing image, it raises questions about authenticity. If the AI is "filling in" details based on its training data of what a moon should look like, is it a deep fake? Similarly, Google’s "Best Take" feature allows users to swap faces from different photos in a burst to create a composite "perfect" moment. While this feels like a standard editing tool, it technically creates a representation of an event that never actually occurred, meeting the AIA's broad definition of a deep fake.

The Pixel-Semantic Gap

A major critique of the AIA is its exception for changes that do not "substantially alter" input data. The authors argue that mathematical measures of change, such as Root Mean Square Difference, do not correlate with human perception of manipulation. This is known as the "Pixel-Semantic Gap."

For instance, a massive change in pixel values—such as shifting the overall brightness or intensity of an image—might not change its meaning at all. Conversely, a tiny change in pixel space, such as digitally adding a small weapon into a person's hand, fundamentally alters the semantics of the image.

Example of standard editing: removing sensor dust from an image.

The AIA provides exceptions for "assistive functions for standard editing," such as the dust removal shown above. However, the authors argue this is a "regulatory backdoor." What is considered "standard" evolves rapidly; as AI becomes integrated into every layer of consumer hardware, the line between "standard editing" and "substantial alteration" becomes impossible to draw without clearer technical thresholds.

Key Results and Policy Recommendations

The paper identifies four legal conditions that must be met for content to be classified as a deep fake under the AIA:

  1. It must be AI-generated or manipulated.
  2. It must consist of image, audio, or video content.
  3. It must resemble real-world entities.
  4. It must create a false appearance of authenticity to a human observer.

To resolve the current ambiguity, the authors recommend removing the "not substantially altered input data" exception from the AIA. They argue that because the law already covers "substantially altered semantics," the "input data" clause only adds technical confusion. Furthermore, there is an urgent need for interdisciplinary dialogue between computer scientists and legal scholars to define "standard editing" before the AIA is fully enforced. Without this, the judiciary will be forced to define technical thresholds on a case-by-case basis, leading to inconsistent enforcement and market instability.

Conclusion

While the EU AI Act is a landmark attempt to regulate digital deception, its current form lacks the technical precision required for the age of AI-integrated photography. By failing to account for the pixel-semantic gap and the inherent processing involved in all digital imagery, the Act risks over-regulating standard consumer features or, conversely, leaving significant loopholes for deceptive content.

Summary updated

BibTeX
@inproceedings{meding2025what,
  author = {Kristof Meding and Christoph Sorge},
  title = {{What constitutes a Deep Fake? The blurry line between legitimate processing and manipulation under the EU AI Act}},
  booktitle = {CSLAW},
  year = {2025},
  doi = {10.1145/3709025.3712218},
}
Figure