AI companies typically use automated programs known as web crawlers and web scrapers to gather data. Web crawlers navigate the internet, cataloging information from various URLs, while web scrapers download this cataloged data. For example, OpenAI has utilized a web crawler called Common Crawl to collect training data for its models.

While generative AI models primarily gather data that is publicly accessible, there are concerns about privacy. For instance, Meta has acknowledged using public posts from platforms like Facebook and Instagram to train its AI. Although locked-down accounts are generally not included, there are instances where private information can inadvertently end up in training datasets due to lax privacy settings or digital leaks.

Bias in the data used to train AI models can lead to skewed outputs that reflect harmful stereotypes. For example, AI image generators may produce more sexualized depictions of women compared to men. This bias arises because the internet itself contains a disproportionate amount of certain perspectives, often favoring wealthier, Western demographics, which can result in AI models that do not accurately represent the broader population.