What is GAPP

The Generative AI Permissions Protocol (GAPP) is a proposed technical protocol that lets image owners say whether their image can be edited by generative AI.

This is done by adding a small, invisible signal directly to the image. People still see the image normally, while participating AI models can detect the signal when the image is uploaded for editing.

If the signal is detected, the AI model does not edit the image.

Taking a cooperative approach to image protection

Many existing image-protection methods try to disrupt how AI models work, interfere with training, or make certain AI edits harder. However, these methods can be difficult to maintain as models change and may be less likely to be accepted by AI providers, because they are designed to break the very systems those providers operate.

GAPP tries a different approach. Instead of attacking the AI system, it gives image owners a way to clearly say what they allow, and gives participating AI platforms a way to recognise and respect that choice.

The goal is to create a practical responsible-AI solution that platforms can realistically adopt.

The signal must survive the real world

A permission signal only works if it stays with the image. Images rarely stay untouched after they are published online. They get resized, compressed, cropped, converted, reposted, and re-captured through screenshots. This creates one of the biggest technical challenges: Robustness

An embedded signal needs to stay detectable, even after common image edits done by people. If the signal disappears too easily, the permission disappears with it. So robustness is essential for GAPP to work in reality.

What GAPP is not trying to do

GAPP does not stop people from viewing, downloading, sharing, screenshotting, or manually editing an image. It does not prevent people from using generative AI. It does not force AI platforms to adopt or support the protocol. It is not a watermark, C2PA/Content credentials, or a metadata-provenance signal.

GAPP simply gives image owners a technical way to express their choice, and gives responsible AI platforms a practical way to respect it.

The research question

Can a visually unobtrusive signal embedded in an image provide a practical and interoperable way for image owners to communicate generative-AI transformation permissions to AI models?

Answering that requires investigating detectability, visual quality, robustness, efficiency, security, interoperability, and ultimately adoption by the AI platforms themselves.