Information su sentenza Google AI

We got big news in the past couple of days concerning a European court ruling against Google. But bear with me, this is not one of those knee-jerk anti-U.S. tech decisions we often see coming out of Europe. This ruling makes some sense—and could have broader implications for the AI sector. A court in Germany found that Google is responsible for what its AI models say in AI-written answers to search results. While the ruling is preliminary, according to Google, the finding spotlights an unresolved question about who’s responsible for costly errors made by AI models.

The court ruled in favor of two businesses that sued Google over AI Overviews results describing those businesses as scams when in fact they weren’t. (German publication The Decoderreported the ruling.) The errors were made by Google’s AI—the judgment noted that the AI Overviews answer said things not mentioned by the websites that were the supposed source. “These are unique assertions invented by Google’s AI tool,” the ruling said, which Google “must accept responsibility for.” 

The ruling highlights how AI-written responses could undercut the legal protections Google and other tech firms enjoy in the U.S. under section 230 of the Communications Decency Act. That law says tech firms aren’t liable for content that appears on their services. Some of those same protections apply in Germany, the ruling indicated: “For standard search engines, courts usually rule that operators do not have to check every link in advance.” 

But as the ruling points out, an AI Overview answer, unlike a traditional search response, “weaves information together to create brand-new text statements.” (In a statement, Google pointed out that “the overwhelming majority of responses” in AI Overviews are accurate, although it acknowledges Overviews can get things wrong at times.)

Google had argued, according to the ruling, that users could check AI answers by looking at the websites the answers link to. But the court noted, “If an AI overview were legally treated as completely unreliable text that required users to check every single link anyway, the entire feature would lose its stated purpose.” Good point!