Common Crawl - dataset

Transparency, accountability 🙈

Common Crawl is seen to suffer from important transparency and accountability limitations, including:

• Data collection and processing. There is insufficient transparency regarding the methods used for data collection, filtering processes, and the criteria for determining "high-quality" content. This obscurity can lead to misunderstandings about the dataset's reliability.

Risks, harms 🛑

Common Crawl is seen to presents multiple risks and harms associated with its use, particularly in the context of training generative AI models, including: 

• Bias/discrimination. The crawling methodology prioritises frequently linked pages, often leading to underrepresentation of digitally marginalised communities. Most content is in English, which limits the dataset's global applicability and inclusivity.

• Simplistic filtering. Filtered versions of Common Crawl used for training AI models often rely on basic algorithms that fail to adequately remove harmful content. This can result in biased outputs that reinforce stereotypes and misrepresent diverse perspectives.

• Copyright. Common Crawl gathers data from various websites, including copyrighted material, without explicit permission. This raises significant legal concerns regarding intellectual property rights.

• Safety. The dataset may inadvertently include illegal or socially unacceptable material, such as hate speech or explicit content. This poses risks for researchers and organisations that may unintentionally propagate such content.

• Research validity. The uncurated nature of Common Crawl data may compromise the validity of research findings derived from it. Researchers may draw conclusions based on datasets that do not accurately represent the broader web or societal contexts.

• Potential for harmful outputs. When used in AI training, the lack of careful curation can lead to models that generate harmful or misleading outputs, impacting users and society at large.

Research, advocacy 🧮

• Mozilla Foundation. How Common Crawl’s Data Infrastructure Shaped the Battle Royale over Generative AI

• Mozilla Foundation: Training data for the price of a sandwich

• Baack S. A Critical Analysis of the Largest Source for Generative AI Training Data: Common Crawl

• Luccioni S., Viviano J. What’s in the Box? An Analysis of Undesirable Content in the Common Crawl Corpus

News, commentary, analysis 🗞️

• https://www.technologyreview.com/news/509931/a-free-database-of-the-entire-web-may-spawn-the-next-google/

• https://www.wired.com/story/the-fight-against-ai-comes-to-a-foundational-data-set/

• https://www.theguardian.com/technology/2023/apr/20/fresh-concerns-training-material-ai-systems-facist-pirated-malicious

• https://www.washingtonpost.com/technology/interactive/2023/ai-chatbot-learning/