In Transparency We Trust?

Based on desk research reviewing academic literature, industry reports and existing regulation, this study assessed the mechanisms platforms use to disclose AI-generated and synthetic content, weighing both human-facing labels and machine-readable watermarking for effectiveness.

Human-facing methods — visual or audio labels and disclaimers — rated poorly: they are easily stripped with basic editing tools, can be omitted by bad actors, and risk stigmatizing legitimate content since terms like "deepfake" become associated with deception regardless of accuracy. Machine-readable watermarking scored better for tamper-resistance but depends on detection systems that remain unreliable and can be biased, for instance against non-native English speakers. The report cites context such as 2024 elections spanning more than 60 countries and roughly half the world's population, the finding that 96% of deepfakes identified in 2019 were pornographic, and projections that a majority of data used to train future AI systems will itself be synthetic.

It concludes no single disclosure method is sufficient and calls for mandatory, combined technical, regulatory and educational measures that place responsibility at the point of content creation rather than on end users.