Can labelling AI-generated content effectively combat the rising tide of misinformation, or does it merely scratch the surface of a deeper trust crisis in our digital age?
An SABC News video that carries a video message from Elon Musk may be easy to believe, which has happened to many recently. What made the Elon Musk deepfake videos so disturbing was not merely their sophistication, but how believable they appeared in an age already drowning in information overload.
In November 2023, a wave of fabricated videos began circulating across social media platforms. Edited to resemble legitimate television news broadcasts, they featured well-known news anchors appearing to endorse what was described as Elon Musk’s “secret investment project.” The production quality was convincing enough that many unsuspecting people reportedly invested money into the fraudulent scheme.
The incident revealed something profound about the AI era now unfolding before us: society is entering a period in which seeing is no longer believing.
That is why, when the South African government later proposed requiring publishers to label AI-generated content, I immediately thought about the Musk deepfake saga. The proposal was clearly driven by legitimate concern. Governments across the world are struggling to understand how societies can protect themselves from synthetic misinformation, manipulated media, and AI-generated deception. But the more I reflected on it, the more I questioned whether labeling alone could meaningfully solve the problem.
At the heart of the proposal lies an assumption: that those producing false or misleading content would voluntarily identify it as AI-generated. History suggests that disinformation campaigns rarely announce themselves. Fraudsters do not typically cooperate with regulatory intentions.
The very actors most likely to misuse AI are also the least likely to comply with labeling requirements. This concern is reinforced by research from the Centre for Data Innovation by researchers Justyna Lisinska and Daniel Castro, who argue that watermarking and labeling systems face fundamental limitations.
Their findings are striking, and they indicate that watermarks can be removed. Open-source models can generate content without labels. Countries with weaker regulations may become safe havens for unlabeled synthetic content.
And perhaps most importantly, the absence of a watermark does not prove content was created by a human being. That insight matters because it dismantles a growing illusion in AI governance: the belief that technical labeling mechanisms alone can distinguish truth from deception.
The researchers go further, arguing that focusing too heavily on labeling risks creating a false divide between AI-generated and human-created content. Both can inform. Both can deceive. Human beings were spreading misinformation long before artificial intelligence arrived. This is why the deeper challenge is not simply identifying AI content.
It is building trusted systems of verification in a world where synthetic media becomes indistinguishable from reality.
That realization leads to what may ultimately become the more effective battleground against deepfakes: devices themselves.
Most people do not interact directly with AI infrastructure. They experience digital reality through devices designed by major technology companies—smartphones, laptops, tablets, televisions, and operating systems. That gives device manufacturers extraordinary strategic importance.
The future fight against deepfakes may depend less on labels attached to content and more on security architectures embedded at the silicon and operating system level.
Imagine smartphones that can authenticate whether audio, video, or images originate from verified capture systems. Imagine operating systems that automatically detect manipulated media patterns before content spreads virally.
Imagine hardware-level provenance systems built directly into cameras and communication platforms.
These interventions would not eliminate misinformation entirely. No technology ever fully eliminates deception. But they could significantly reduce the scale and speed at which synthetic fraud spreads.
This suggests governments may need to rethink where regulation should focus.
Instead of concentrating almost exclusively on publishers and content creators, policymakers may need to engage device manufacturers, operating system developers, and platform infrastructure companies as frontline actors in information integrity. The AI misinformation challenge is ultimately not just a content problem. It is also an infrastructure problem.
And history repeatedly shows that the most important technological battles are often won not at the visible application layer, but deep within the architecture of the systems society depends on every day.
