German court ruling declares Google's AI Overviews are Google's own words and makes it liable for false answers
159 points by gspr
159 points by gspr
Title trimmed slightly without altering meaning.
This seems eminently reasonable to me. If you're a search engine and just point to websites that might match the user's query, that's one thing. But if you decide to go read those sites and instead provide your user with a summary, you're their source of information. You shouldn't be able to ignore the responsibilities that come with that (then of course different jurisdictions will disagree about what those responsibilities are, but that's a different matter).
My issue with this comment is that Google has provided users with summaries for years. (Snippets, which you can’t even tell Google to use a particular thing for IIRC)
Then I read the article and it sounds like it’s saying only the parts not taken from search results are liable for damages? Which is kind of weird. Maybe I misunderstood.
The court explicitly cites existing judgements about how important search engines are for a functioning internet, and how presenting snippets of other websites doesn't incur legal liabilities as such, even if the cited texts are slanderous. (The worst that can happen is that Google is served a takedown notice.)
In contrast, the AI's utterances were found to originate from Google itself (some of the statements of fact uttered by the AI weren't even supported by the sources that it cited), hence Google is legally on the hook for what its own systems produce and present to the public.
The german legal concept/standard is "making another's opinion/statement your own".
Merely providing a snippet of another website, derived in a way were the sentences are lifted verbatim and without changing their meaning, you're not making that website's statement your own. That means if someone goes and sues you for defamation, you might still be required to take it down, but all liability lies at the original source of the statement. There is some leeway to allow shortening or summarizing a sentence, but generally the more you move away from the original the higher your risk becomes of crossing the threshold.
But once you move away from quoting and instead take something someone else said and say it in your own words, you've made their opinion/statement your own. It's not longer just what they said, it's what you say (there is some exceptions around this for when you're just presenting two sides of a defammation case, for example, so you can still summarize such disputes, but it's not applicable to this case). You are now liable if something is wrong. That this applies to summaries of untrue or defamatory statements is well established in Germany.
The court further argued:
Then I read the article and it sounds like it’s saying only the parts not taken from search results are liable for damages? Which is kind of weird. Maybe I misunderstood.
I don't understand what you mean here. Do you mean the stuff written under the headline "Search engine liability rules don't apply to AI 'search'"? I read that as referring to the rules that exempt search engines from liability when providing search results. That those rules don't apply to AI summaries, meaning that Google is liable.
Or do you mean something else?
I think I got the wrong impression from:
The court also pointed to a protection gap. If Google were only liable for obvious violations, victims would have no real legal recourse when the AI makes false claims. The third parties whose websites served as sources hadn't even made the statements in question. So victims couldn't sue the sources, and under existing rules they couldn't effectively sue Google either.
And
But AI overviews generate "independent, new, and substantive statements" by evaluating and combining content from various third-party sites.
(Particularly it doesn’t seem AI results are independent if they solely regurgitate other websites)
But I can’t find a paragraph I swear I read, so I probably misread the article to begin with ^^;
I wonder if this is the way to legislate LLMs out of search. If LLMs are non-deterministic and companies are responsible for what they say, then there is arguably no possibility of decisively stopping these algorithms from hallucinating slander.
Google covers 80 percent of the legal costs; the plaintiffs pay 10 percent each.
How come the plaintiffs had to pay if the court ruled almost entirely in their favor? Is it because of the previously mentioned two dismissed "minor requests"?
I read the judgement, and to my non-lawyer understanding some utterances of Google's AI that the plaintiffs complained about fall under freedom of speech (being utterances of opinion instead of utterances of facts), hence the splitting of costs since the defendant wasn't found guilty of those.
I don't think "determinism" is what matters here. An LM-based program that is not deterministic can generally be made deterministic for a given input. Even if decoding is stochastic, it can be made deterministic by setting a random seed.
The problem with LM-based programs is that they cannot be proven to give the correct output for a given input.
Note an important thing here: Germany has no case law, so this has signal effect, but other courts might decide differently.
I think the ruling shows well: current law is very capable of handling that aspect of AI - especially when it comes to liability - you are responsible for what you produce. If read it and it's very reasonable.
I didn't know about the handling of precedent in continental law.
Here's a good explanation for anyone who is interested: https://max-eup2012.mpipriv.de/index.php/Precedent,_Rule_of
I was in a very funny conversation with a very well paid US lawyer in which he stated that he recommends against a certain thing to be based in Germany because there's little case law there and I was like "yep, that's because case law isn't a thing - the law applies".
(I also want to state that I really loved working with that lawyer, so no shade here if he reads and remembers that :))
Sounds fair. Either it's Google's, in which case it's their fault or it's someone else's then I'd argue it's copyright infringement.
By that logic the whole concept of a search engine is copyright infringement.
No, the case very clearly draws a line for answering a query and pointing you to sources, maybe showing an excerpt where the search engine is an intermediate and protected by specific laws and high court decisions, where an AI summary does not fall under that protection.
Yes, but OP was trying to extend that into a copyright infringement argument
No.
"Either it's Google's, in which case it's their fault"
or
"it's someone else's then I'd argue it's copyright infringement."
Is totally okay. Because in the second case, Google would try to draw a line to the "vague" original, which totally gives rise to an argument that they presented a new derivative without properly getting permission from a source. That would open the door for an argument that claims copyright infringement.
As stated in my response there are specific laws that give search engines (and indices in general) certain permissions as long as various conditions are met.
These make (try to) make it clear that something is search and just "previewed".
Since LLM producers claim the output is essentially original it's hard to argue it's a preview.
Also just to be clear I am not talking about whether these laws are good or bad but it seems legally consistent with these things.
A search engine refers you to the original page. An AI summarization engine authors a new page for you with the data taken from another page.
At least here there is explicit law that says indexes are fair use and it actually was a topic with Google Image Search with the ruling stating that thumbnails are allowed. With the giant backdoor of thumbnails not being defined.
So yes they would be if there was no explicit exclusion.
(I've been working at a search engine before so I got to talk to a lawyer about these things)
A musician in my country had a concert cancelled by the venue because the Google AI Overview falsely claimed he was a child molester.
IMO he should have had legal cause to sue Google for defamation and loss of income (although he did not attempt a lawsuit).
An analysis by AI startup Oumi for the New York Times found that Google's AI Overviews with the current Gemini 3 model answered correctly 91 percent of the time.
That's solid enough for everyday use by most people.
In what crazy universe is a search engine that lies in one out of ten searches "solid enough" for anyone, let alone "most people"?
Whether 91% is good or not obviously depends on the baseline. Would be interested in how often the pre-AI-Overviews Google surfaced correct vs incorrect answers in its top few results. I'd guess that was closer to 80%, but really depends on the mix of questions and how you define correctness.
At last some sanity -- if you believe an AI is a valid way to answer questions, then you need to be willing to stand by them.
If you know those answers are false, and yet still vend them, you need to be liable for those results. People know not to trust random webpages, but google is presenting its answers both as itself, and over the top of websites (especially the web pages whose correct answers it stole the content from, even if it then randomizes the correct information to make it false).
Claiming that users should just accept that anything presented is wrong is unethical: present accurate information, or don't present information, don't present garbage and false information as if accurate and then include a disclaimer that it is nonsense and can't be trusted.
Any company that uses a slop engine to provide information has to accept that a user has a perfectly reasonable expectation of accuracy in automated responses. If you can't ensure that your slop generator can achieve the basic requirement of repeating correct information, then don't pretend it can.
Every company or person who uses these content and open source stealing programs should be liable for everything they produce - be that the consequences of incorrect information, or copyright violations from theft of OSS code.
Wouldn't a disclaimer under every summary absolve them from this law? People are gonna fall for it just as much but you can't be liable if you never claimed to give a true answer.
That's what Google argued they do, but the court ruled you can't publish untrue statements of fact followed by "lol just kidding, check sources and don't believe me".
More formally, the court found that a lot of the AI summaries in evidence were "fresh" statements of fact originating with Google itself (as opposed to statements of opinion, which enjoy broader legal protection) and furthermore that some of the false statements of fact weren't even supported by the sources linked..
I don't know if you can disclaim your way out of slander and defamation, else those wouldn't be able to be prosecuted under free speech either? It's all a mess, but this position seems positive to me...
Is this consistent? Are they going to hold me accountable for incorrectness on, say, my blog? If so, who decides I am incorrect? Its one thing to hold them accountable for generated content without a source. Its another to hold them accountable because they used an available source.
Regulating speech, even if generated, isn't a good idea, in my opinion.
FWIW, the German constitution agrees with you, and protects free speech: https://www.gesetze-im-internet.de/gg/art_5.html
And now you get to read the judgement, which explains why and how untrue, slanderous statements of fact are less protected than statements of opinion.
Good news: Google didn't defend their speech or the truth of their statements at all in that court case. They only defended that it's not their speech and the court disagreed.
So you believe you should have no recourse if someone bought a billboard and put your face and full name on, asserting as fact that you are guilty of (insert heinous act here)?