Existing systems solve adjacent parts of the problem
GAPP draws from prior work in adversarial protection, watermarking and provenance, while focusing specifically on machine-readable transformation permission.
01
PhotoGuard by MIT
Purpose
- Make individual images resistant to malicious AI-powered editing.
- Add an imperceptible perturbation before the image is published.
What worked
- Demonstrated that carefully designed perturbations can disrupt diffusion-based image editing.
- Raised the cost of producing convincing edits from a protected image.
Limitations
- Protection is tied closely to the targeted model and editing pipeline.
- Changes in model architecture or preprocessing can weaken effectiveness.
- It disrupts transformation rather than communicating a shared permission.
02
Glaze by UChicago
Purpose
- Protect visual artists from having their individual styles mimicked.
- Apply subtle style cloaks before artwork is shared online.
What worked
- Shifted how protected artwork is represented in a model’s feature space.
- Reduced the accuracy of style mimicry after fine-tuning.
- Was developed and evaluated with substantial artist participation.
Limitations
- Focuses specifically on artistic style mimicry.
- Acts through future model training rather than at transformation time.
- Effectiveness depends on the training pipeline and model behaviour.
03
Nightshade by UChicago
Purpose
- Deter unauthorised model training through prompt-specific poisoned images.
- Make scraped protected data costly or harmful to use.
What worked
- Disrupted targeted concepts using relatively few aligned poison samples.
- Produced effects that could extend to semantically related concepts.
- Changed the economics of using content without permission.
Limitations
- Targets model training rather than image transformation requests.
- Depends on protected images entering future training datasets.
- Poisoning may affect behaviour beyond one protected image.
04
SynthID by Google
Purpose
- Invisibly watermark AI-generated media so its origin can be detected.
- Support provenance and transparency across generative systems.
What worked
- Demonstrated large-scale deployment of invisible watermarking.
- Balanced detectability, fidelity, robustness and efficiency.
- Survived many common image transformations.
Limitations
- Primarily communicates that media was AI-generated.
- Does not communicate whether an image may be transformed.
- Detection requires access to the relevant verification system.
05
Content Credentials
Purpose
- Provide verifiable provenance and content-history information.
- Record claims about origin, creation and editing.
What worked
- Established a shared technical specification.
- Created an interoperable model for structured assertions.
- Connected platforms, publishers, vendors and device makers.
Limitations
- Focuses on provenance and authenticity rather than permission.
- Credentials may become unavailable in unsupported workflows.
- Carries richer information and trust requirements than GAPP needs.