AI & Society
Only 38% of AI image generators use adequate watermarking, and 18% use deep fake labelling

In a 2026 paper in Policy & Internet, researchers at Maastricht University report that current adoption is low. They found that 38% of AI image generators currently use adequate watermarking, and 18% currently use adequate deep fake labelling.
Watermarking embeds information in an image to show it is AI-generated. Deep fake labelling flags an image as AI-generated. The authors combine a legal analysis of the 2024 EU AI Act with an empirical check of AI image generators.
What did the empirical check find?
The authors found minority adoption for both practices. Only 38% of AI image generators currently use adequate watermarking. Only 18% currently use adequate deep fake labelling. The authors note this gap even as AI-generated images have become hard for people to tell apart from real images.
What does the legal analysis say?
The paper identifies four categories of generative AI deployment scenarios. The authors outline how the legal obligations could apply in each category. Watermarking and AI labelling rules are becoming legal requirements in many jurisdictions, the authors write, including under the 2024 EU AI Act.
What do the authors propose?
They suggest a range of ways to improve how the legally required techniques are put into practice. They also say they publicly share their tooling for detecting watermarks.
FAQ
How many AI image generators were assessed?
The abstract does not say how many were assessed or which ones.
What does adequate mean in this study?
The abstract uses the word but does not state the technical cutoff the authors applied.
What is the main takeaway?
Among the AI image generators studied, adoption of adequate watermarking and deep fake labelling remains low.
Limits and what's next
The abstract describes an empirical and legal analysis of AI image generators, but it does not say how many generators were studied or which ones. It also does not say what technical standard the authors used for adequate watermarking or deep fake labelling. The setting is the legal context of the 2024 EU AI Act. This is one study, and the abstract reports no independent replication. The authors suggest avenues for improving the mandated techniques and say they share their watermark detection tooling publicly.
Source: Bram Rijsbosch et al., “Missing the Mark: Adoption of Watermarking for Generative AI Systems in Practice and Implications Under the New EU AI Act,” Policy & Internet, 2026.