Pense
As we fight to determine where AI will sit in society, courts around the world are taking tech companies to court. And this week, I find myself reporting on Germany. And unfortunately 'Guten Tag' is about the only German I know. So as I continue on my quest to inform you about AI, I won't be able to break down this particular filing line by line. I can only refer you to English language sources on the ruling.
A case was under litigation about whether or not tech companies should be held liable for AI search overviews. The case came up because false information was published about two German publishers. At the moment LLMs are not accurate and make up information. I have no idea how Google search summaries work, but I'd imagine they would consider the first page of results, review the on-page content, and summarize using some agent protocol. And while their method should eliminate hallucinations it doesn't. Some researchers found that it is only correct 72% of the time (New York Times). When I was in school that was a 'D' so the 'D' student is trying to tell the other students what is the correct answer. That is laughable.
If these numbers are to be believed, then valid sources aren't required for AI search overviews. Within 5-ish years, I expect factual hallucinations will be negligible if practical engineering can win over hype. And I say this as a person who is developing source of truth mechanisms for my own application. Accuracy will be better when AI delusion dies down. So, this isn't even the real issue.
The real issue is in how Google defended its claim against liability.
Google sought free speech protection for a series of math equations. And the court ruled only flesh and blood humans can have that.
According to The Decoder, the German court ruled around these things:
Are AI search overviews required for search results? → No
Should AI overviews be covered by search engine liability laws (the ones where search engines aren't held accountable for the search results)? → Only in their dreams.
Should users be solely responsible for verifying search results? → Nope, personal responsibility doesn't absolve publishers from liability.
Is AI overviews covered by free speech laws? → Nice try, but no. Overviews are math and math can only express the opinions of the corporate entity responsible for its creation. People have personal rights against slander that supersede a collection of equations.
For the record, Google's own documentation claims AI Overviews are "built to only surface information that is backed up by top web results" and "generally don't hallucinate in the ways that other LLM experiences might." You can read that for yourself in their 2024 AI Overviews paper and their 2025 AI Overviews and AI Mode paper. The German court disagreed.
You can read the full ruling here (PDF). For additional context, ArsTechnica covers the ruling's broader implications.
So by summarizing searches, Google became a publisher and received the related liabilities. Three cheers for the German judges who exercised their common sense and good judgment. Someone buy them a beer.
This is great and all, but I know all the dirty little secrets of the past twenty years of search. So, it got me thinking, "How does AI know if information is false?" It doesn't. Consider this: AI can be used to spin up sites with hundreds of pages in less than an hour. If I wanted to falsify search results I can spin up enough for AI to summarize false information as fact. You may be tempted to think — wouldn't Google know the site is new and discount the information? That doesn't matter. This tactic has been used for decades to drown out bad PR or to promote false narratives and credentials of fake gurus. An old site, a new site, or a big site has already been used for just this tactic before AI could do it in less than an hour. That's why startups like GroundNews (not sponsored) spun up in the first place.
So, if Google shifts from simply being an avenue of finding information to vetting and summarizing the information, their liability shifts because they have now published information and are responsible for the correctness of their representation. They can state that information may be false at the bottom. But that is no shield from liability.
And do many consumers scroll that far down? I would argue probably not since we know that the average search doesn't leave the first page (Search Engine Land). According to Pew Research, users who encounter an AI summary are less likely to click through to linked sources at all. Thus, the disclaimer is useless unless it is at the top of the summary: "This information may be false and the URLs could be fake — please don't take our experiment seriously." But that would hurt their bottom line. And even if disclosed, I'm not sure I want them shielded from liability. Publishers have to publish retractions. Why shouldn't they?
If you want to know more about the different agent protocols being developed to improve accuracy, subscribe to the Tuesday edition. The link is below. I'll be breaking down the Model Protocols (MCP, SCP, G2CP, and more this coming Tuesday).
https://newsletter.geekyinsights.com/upgrade
This has a been A Geeky Production
