Deepfakes, Disinformation and AI — Where Do We Draw the Line?

In 2023, creating a convincing deepfake required specialist software and hours of compute. In 2026, it takes a phone, an internet connection, and thirty seconds. The capability gap between what a nation state could produce and what a teenager can produce has collapsed. This is not a problem technology can fully solve — it requires social, legal, and technical responses working together.

What detection can and cannot do

Watermarking and provenance systems — like the C2PA standard adopted by Adobe, Google, and others — attach cryptographic metadata to AI-generated content, allowing platforms to verify its origin. When content passes through these systems cleanly, verification works well. When it does not — when someone screenshots a video or uses a model without watermarking support — detection rates fall significantly.

AI-based detection classifiers have a worse problem: they can be fooled. A deepfake generator and a deepfake detector are engaged in an arms race, and the generator is currently winning. Detection tools are useful as one signal among many, but treating them as reliable gatekeepers creates false confidence.

Where the real leverage is

• Platform responsibility: requiring provenance labels on synthetic media before algorithmic amplification
• Legal frameworks: several jurisdictions now criminalise non-consensual deepfake content and political deepfakes within election periods
• Media literacy: the most durable protection is scepticism — checking sources and treating viral content with appropriate scrutiny
• Industry standards: the C2PA credential ecosystem is growing, but adoption needs to be near-universal to be effective

The honest answer is that no single intervention is sufficient. The goal is not to make deepfakes impossible — it is to raise the cost of deception high enough that most bad actors choose easier targets, while giving individuals the tools to protect themselves.