Excavating AI / ImageNet Roulette
Click to visit website
About
Excavating AI is an investigative project and digital tool designed to uncover the social and political implications of image-based machine learning training sets. Developed by Kate Crawford and Trevor Paglen, the platform centers on the "archeology of datasets," specifically scrutinizing ImageNet, one of the most influential benchmarks in artificial intelligence. The project includes an experimental software tool called ImageNet Roulette, which allows users to see how they would be classified by a deep-learning model trained exclusively on ImageNet's "Person" categories. By making these hidden classifications visible, the tool illustrates the shaky and often offensive assumptions that underpin how AI systems recognize and interpret human identity. Technically, the ImageNet Roulette tool utilizes an open-source Caffe deep-learning framework. When a user uploads a photograph, the application runs a face detector to locate subjects and then assigns them a label based on the thousands of categories found in the original ImageNet "Person" hierarchy. These categories include labels ranging from professions like "welders" to derogatory terms and nonsensical classifications such as "snob," "failure," or "closet queen." This process highlights the problematic nature of "supervised learning," where human laborers—often Amazon Mechanical Turk workers—are tasked with flattening complex human experiences into static, often biased, noun-based categories. This classification process is revealed to be a subjective and often ideological act rather than a purely scientific one. The project is primarily intended for researchers, AI ethicists, artists, and students of media theory who seek to understand the non-neutral nature of data. It serves as a powerful case study for those investigating the resurgence of "physiognomic" AI—the pseudoscientific attempt to derive character or morality from facial features. Unlike standard AI diagnostic tools that aim to "fix" bias through more data, Excavating AI argues that the very act of classification is a political intervention. It provides a rare forensic look at the historical and sociological layers of datasets that have already been integrated into critical social infrastructures like hiring, education, and law enforcement protocols. What makes Excavating AI unique is its critical, investigative approach to the "black box" of training data. Rather than offering a commercial service, it acts as a provocateur, revealing the "treachery of images" where labels and referents are arbitrarily linked. By highlighting datasets like ImageNet, JAFFE, and IBM’s Diversity in Faces, the project demonstrates how technical systems can replicate 19th-century phrenology and race science. It stands out as an essential resource for anyone questioning the power dynamics of who gets to decide what an image means and how those representations affect real-world rights, liberties, and forms of self-determination.
Pros & Cons
Provides a rare forensic look into the hidden labels used to train major AI systems.
Uses open-source frameworks for transparency in its experimental results.
Highlights critical ethical issues like phrenology and race science in modern AI.
Does not store user-uploaded images, respecting immediate privacy during the experiment.
Documented by world-class researchers from the AI Now Institute at NYU.
The 'Person' categories in the original dataset are currently down for maintenance or removed from public servers.
Produces offensive and derogatory labels by design to illustrate dataset bias, which may be distressing.
Limited to noun-based classifications derived from the historical WordNet hierarchy.
Functionality may be limited as several major source datasets have been taken offline by their respective institutions.
Use Cases
Ethical AI researchers can use the tool to demonstrate the real-world impact of biased training data to policymakers.
Media theory students can perform an 'archeology of datasets' to trace how historical prejudices are replicated in machine vision.
Artists and activists can visualize the 'physiognomic' assumptions of AI to create critical commentary on surveillance.
Computer science educators can use the project as a case study to teach students about the non-neutrality of supervised learning.
Platform
Features
• ethical ai research insights
• open-source software framework
• classification visualization
• imagenet 'person' category analysis
• historical dataset archeology
• deep-learning caffe framework
• automated image labeling
• face detection and bounding boxes
FAQs
What is the primary purpose of ImageNet Roulette?
It is an experimental tool created to show how technical systems are trained using problematic data. It classifies users based on the 'Person' category in ImageNet to reveal the biases inherent in such datasets.
Does the tool store the photos I upload?
No, the creators state that ImageNet Roulette does not store the photos people upload to the application. The tool is designed for research and demonstration purposes rather than data collection.
Why are some of the labels returned by the tool offensive or nonsensical?
The labels are drawn directly from the ImageNet 'Person' category, which contains many problematic, misogynistic, and racist terms assigned by crowdsourced workers. The tool's purpose is to make these hidden classifications visible.
What happened to the 'Person' category images on the ImageNet website?
In early 2019, approximately 1.2 million photos in the 'Person' category were taken down or moved 'under maintenance' following research into their ethical and privacy violations. The tool uses a backup of these categories.
How does the tool determine a label if no face is detected?
If the face detector fails to locate a face, the application sends the entire uploaded scene to the Caffe model. It then returns a label in the upper left corner of the processed image based on the model's classification of the background.
Pricing Plans
Open Access
Free Plan• Image classification visualization
• Face detection output
• ImageNet 'Person' labels
• Archeological analysis findings
• Open-source Caffe framework usage
Job Opportunities
There are currently no job postings for this AI tool.
Ratings & Reviews
No ratings available yet. Be the first to rate this tool!
Featured Tools
adly.news
Connect with engaged niche audiences or monetize your subscriber base through an automated marketplace featuring verified metrics and secure Stripe payments.
View Details
Veo 4
Create cinematic 4K videos up to 30 seconds with synchronized audio and realistic motion using advanced AI models designed for professional content creators.
View Details
Nano Banana
Create and edit professional-grade visuals for designers using natural language commands powered by Google Gemini for character consistency and 4K realism.
View Details
GPT Image 2
Generate photorealistic AI images with 95%+ text accuracy and 4K resolution. Create professional-grade posters, logos, and marketing assets with perfect text.
View DetailsVeo 4
Produce cinematic AI videos using text, image, and audio references with native lip-syncing and consistent character identity for high-quality storytelling.
View Details
ToolCenter
Find the best AI solutions for your workflow with a curated directory of over 1,700 tools across categories like design, development, and content creation.
View Details
Sceneform
Design hyper-realistic AI influencers and viral social media content with an all-in-one studio for persona building, motion syncing, and batch video rendering.
View Details
Grok Imagine
Transform creative ideas into cinematic 2K videos and photorealistic images with xAI’s Aurora engine, featuring precise motion control and multi-modal inputs.
View Details
Salespeak
Provide founder-level sales expertise across web, email, and LLM search with AI agents that learn your product in minutes to capture intent and convert buyers.
View Details