Google confirms it’s training AI on scraped web data

Everything you have ever published openly on the web might already be feeding an AI. Now, Google has officially acknowledged it in black and white.

On July 1, Google quietly updated its privacy policy, a change first spotted by Gizmodo. By July 5, The Verge confirmed the shift: the search giant explicitly stated that it trains AI models like Bard and Cloud AI on publicly scraped web data, rather than just using it for legacy tools like Google Translate.

Google spokesperson Christa Muldoon downplayed the revision, telling The Verge that the company has long been transparent about using open web info and simply clarified that newer services like Bard are included under the same umbrella.

From language tools to all AI

Take a closer look at the revision history, and you will see this was far more than minor housekeeping.

Google replaced the specific term "language models" with "AI models." That subtle linguistic swap opens the door wide: Google can now feed your publicly accessible web footprint into virtually any machine learning system it builds, well beyond text generation.

Interestingly, that critical note was not front and center. It was tucked away under an embedded link for public sources inside the policy's "Your Local Information" section.

While Google claims it only harvests publicly available data, the policy fails to mention how, or even if, it will prevent copyrighted material from slipping into its training pipelines. Many web publishers specifically prohibit web scraping for AI training in their terms of service, and privacy regulations like Europe's GDPR strictly protect unauthorized data use.

The scramble for data moats

Google's bold policy update exposes the massive legal gray zone AI labs currently operate in.

Competitors like OpenAI have grown notoriously secretive about what went into training models like GPT-4, especially regarding social media posts and copyrighted human art. Whether training on public internet text qualifies as fair use remains untested at scale, but lawsuits are mounting. Gannett, the largest newspaper publisher in the U.S., is already suing Google over digital ad monopolies, while critics slam Google's AI search previews as "plagiarism engines" that cannibalize web traffic.

Meanwhile, the platforms holding that precious conversational data are panicking. Reddit and Twitter recently pushed draconian API limits and anti-scraping walls to stop AI companies from free-riding on their databases, triggering massive user backlashes in the process.

Behind the scenes, the human workers sorting through these scraped terabytes to keep systems safe continue to face grueling hours and extreme conditions.

Why it matters

For regular internet users and creators, this marks the end of passive publishing. Anything you leave on the open web can now be ingested, processed, and redistributed through an AI interface without credit, compensation, or consent.

For the AI landscape, Google is essentially normalizing the idea that the entire open internet is its training ground. It is leaning on its decades of crawling dominance to build a competitive shield against OpenAI, betting that legal definitions of fair use will bend in its favor.

We think this move will only accelerate the closure of the open web. As platforms realize tech giants will vacuum up every unlocked byte, more paywalls and walled gardens will go up. The tragedy is that the very internet that made generative AI possible is now being locked down to protect itself from it.