Nightshade

The Glaze Project deployed a research study that uses AI-based classifiers to analyze whether images were created by a human artist or generated by an AI model, publishing findings on the accuracy of both automated detectors and human reviewers. This work was published in October 2024.

The research, titled 'Organic or Diffused: Can We Distinguish Human Art from AI-generated Images?', compares the ability of AI detection tools (such as Hive and Optic) and human reviewers (artists and non-artists) to classify images as human-made or AI-generated. The study inputs images into detection systems and human review panels and produces accuracy and false-positive/false-negative rates. This is a published academic study, not a deployed consumer tool; no standalone product for end-users has been confirmed.

The Glaze Project offers artists a free, downloadable tool called Glaze that uses machine learning to add imperceptible pixel-level changes to artwork, causing AI image models to misidentify the artist's style and preventing accurate style mimicry. Glaze was first released in March 2023 and has been downloaded more than 8.5 million times globally.

Glaze takes an artist's digital image as input and, using machine learning style-transfer algorithms, computes minimal pixel perturbations that make the image appear to AI models as a dramatically different art style while remaining visually unchanged to human viewers. For example, a charcoal realism portrait may appear to an AI model as abstract art in the style of Jackson Pollock. A browser-based version called WebGlaze was deployed in August 2023, running on AWS cloud GPU servers to serve artists without powerful computers; WebGlaze is invite-only and available to any human artist who does not use generative AI tools.

Nightshade offers artists a free, downloadable tool that uses machine learning to alter pixels in artwork so that any AI model trained on those images learns incorrect visual associations — for example, interpreting a dog as a cat. The tool was first released on January 18, 2024, and has been downloaded more than 2.5 million times globally.

Nightshade takes an artist's digital image as input and applies a computed perturbation — a set of pixel-level changes invisible to the human eye but visible to AI — that causes generative AI image models to mislearn the associations between text prompts and visual content. Technically, it is a prompt-specific poisoning attack on text-to-image diffusion models, designed as a multi-objective optimization to minimize perceptible changes while maximizing disruption to model training. The tool runs entirely offline on the user's own machine; no image data is sent to the researchers. Version 1.1 was released on April 20, 2026, with bug fixes.