A court in Munich declared that Google is liable for their "AI summaries" and all its hallucinations.
-
@lymphomation @david_chisnall @tante
As for judging the veracity of LLM output based on its citations and what other sources say, by the time you've done that, you would've done better not to use an LLM at all - it would've been both less work and higher quality outputIt's also not good. It turns out that existing ML models trained on x-ray data overfit on specific measurement errors for individual x-ray machines and produce surprisingly poor results when you try to use them on a different x-ray machine, of the same model in the same hospital, let alone a different model.
There was a paper published near the start of the year debunking a load of the claims about ML in radiology.
But that doesn't stop it being the go-to example for boosters.
-
A court in Munich declared that Google is liable for their "AI summaries" and all its hallucinations. This is an important step to bring "AI" slop in line with all other products on the market: "AI" products are basically the only ones where a provider can just deliver unchecked garbage and put all the liability on the consumer. I hope to see aggressive change here.
-
As I've said elsewhere today:
There are strict legal limits on where you can limit liability. Your calculator can't have that disclaimer at all because (in both the EU and USA) there are very strict limits on disclaimers of liability for physical machines (which is an issue that comes up in open-source hardware quite often).
Even in software, claiming in your marketing that your product does one thing and then having a disclaimer in the license that says that it does not, in fact, do that thing is generally a problem: you may not be liable for the damages from failing to do the thing, but there's a good chance that you're liable for fraud. A disclaimer of liability isn't a get out of jail free card, it's a statement of intent.
@david_chisnall @tante once more, the law makes arbitrary distinctions. What if I grow a biological calculator? (which one can...)
-
@david_chisnall @tante once more, the law makes arbitrary distinctions. What if I grow a biological calculator? (which one can...)
It's not really an arbitrary distinction. The relevant law treats software as a component. The liabilities apply to a final product. The product liability laws cover products delivered to customers. It's the responsibility of the product builder to ensure that components meet the requirements and to use contract law to enforce any liability that's necessary to propagate along the supply chain.
The EU's CRA takes a similar view: an open-source project does not have any liability but a product that incorporates that project must do its own due diligence to ensure compliance.
-
A court in Munich declared that Google is liable for their "AI summaries" and all its hallucinations. This is an important step to bring "AI" slop in line with all other products on the market: "AI" products are basically the only ones where a provider can just deliver unchecked garbage and put all the liability on the consumer. I hope to see aggressive change here.
@tante I'll eat my hat if something actually comes of it.
So to speak -
@tante This is the kind of thing I couldn't have come up with, because I would never have considered what the LLM is spitting out to be Google's words.
Guess there is no real way out. They get what they wanted and I isn't legally theft (however incompetent)... but now it's their words so guess they are responsible.
Look forward to how they try to fight this one. Sorry it's not our words, we actually stole the entirety of human creation.
@theeclecticdyslexic @tante they made/own/run the thing, so why not? They've put the summaries front and centre. They integrated it when there was absolutely no need. The LLM didn't have to be doing any of those things.
-
A court in Munich declared that Google is liable for their "AI summaries" and all its hallucinations. This is an important step to bring "AI" slop in line with all other products on the market: "AI" products are basically the only ones where a provider can just deliver unchecked garbage and put all the liability on the consumer. I hope to see aggressive change here.
@tante as far as I know, this is sort of incorrect, cause this ruling is just a local court
-
Google's defence needs to be amplified by anyone talking to politicians about 'AI' regulation:
Google is explicitly saying in their legal filing that the outputs from their LLM should not be trusted and that users should know that.
That's one hell of an admission. Imagine saying that about any other category of product.
Q-tips says it. "don't clean your ears". What do 90% of Q-tips buyers do? Clean their ears.
-
J jwcph@helvede.net shared this topic
-
Google's defence needs to be amplified by anyone talking to politicians about 'AI' regulation:
Google is explicitly saying in their legal filing that the outputs from their LLM should not be trusted and that users should know that.
That's one hell of an admission. Imagine saying that about any other category of product.
@david_chisnall @tante I remember years ago being on the Tube in London, and the it stopped, and the recommendation was "Please use other means to get to your destination". Which I felt was rather like saying "We cannot do our job, please find someone who can do our job of getting you there"
It feels a little like this.
-
It's also not good. It turns out that existing ML models trained on x-ray data overfit on specific measurement errors for individual x-ray machines and produce surprisingly poor results when you try to use them on a different x-ray machine, of the same model in the same hospital, let alone a different model.
There was a paper published near the start of the year debunking a load of the claims about ML in radiology.
But that doesn't stop it being the go-to example for boosters.
@david_chisnall Could you, please, link the paper?
-
It's also not good. It turns out that existing ML models trained on x-ray data overfit on specific measurement errors for individual x-ray machines and produce surprisingly poor results when you try to use them on a different x-ray machine, of the same model in the same hospital, let alone a different model.
There was a paper published near the start of the year debunking a load of the claims about ML in radiology.
But that doesn't stop it being the go-to example for boosters.
I base my perspective on review articles published to journals. This is one of them with an image capture of the review articles conclusion. This one of many such published papers.
Review
Redefining Radiology: A Review of Artificial Intelligence
Integration in Medical Imaging -
@tante Good article. Too bad that the fall down at the end: '... the fallout could hit not just Google but every AI provider whose systems paraphrase content from the web.'
The whole point is that ai(llm) is not paraphrasing content from the web. It is making shit up.
@MeneerDeBruin @tante yes, and i think this is the point missed by many, that an LLM is paraphrazing or summarizing or inferring... including making stuff up, because it is designed to do so: it has a "sycophancy bias" (yes, a real thing) and "unanchors" itself from references so it can string together snippets in order to please users so they continue to engage with it. It does not speak truth, or curate, when it summarizes.
-
Indeed no other industry or product could get away with that ... - let us just hope that this court's view is upheld through future instances
@sebastian @david_chisnall @tante
I think the odds of that are zero, I'm assuming that Google will find someone to bribe before long.
-
@lymphomation @david_chisnall @tante
AI abilities are "jagged" because AI is an umbrella term for a bunch of disparate technologies and the AI used for radiology is almost entirely unrelated to the AI at issue here@sabik @lymphomation @david_chisnall @tante
"But Science!" is just the new talking point for AI astroturf.
Notice how all the boosters started saying that and "jagged" at the same time?
-
@tante ”A German court has ruled that Google is directly liable for what its AI search overviews say. Previous case law shielding search engine operators from liability doesn't apply to AI overviews.”
This is the correct action to take. You are operating a technology that produces this information. You are not connecting people to somebody else’s voice.
@gimulnautti @tante This should be cemented on the EU level. It might get Google thinking about spreading lies.
-
Q-tips says it. "don't clean your ears". What do 90% of Q-tips buyers do? Clean their ears.
@nosrednayduj @david_chisnall @tante I can tell I haven't had enough caffeine yet, because now my brain is going down the rabbit hole of "we have to clean our ears with Q-tips because getting it done safely at the doctor is not financially accessible to most Americans."
But Europeans probably do it too, and I know Asian countries have those little ear spoons which presumably have similar risks of eardrum puncture.
-
Hallucinations aren't Google's fault, and they can't just fix them. It's a property of AI. So just don't use it for topics you don't understand. And frankly, I don't know anyone who still uses Google. Even my mom uses DuckDuckGo.
@PerryNoID
If Google knowingly uses substandard or faulty tools that produce bad outcomes, those bad outcomes are Google‘s responsibility.If I build and operate a soda vending machine that sometimes dispenses antifreeze in Fanta bottles, that’s on me too after all.
-
@sabik @lymphomation @david_chisnall @tante
"But Science!" is just the new talking point for AI astroturf.
Notice how all the boosters started saying that and "jagged" at the same time?
-
@theeclecticdyslexic @tante they made/own/run the thing, so why not? They've put the summaries front and centre. They integrated it when there was absolutely no need. The LLM didn't have to be doing any of those things.
@noodlemaz @tante Oh, for sure. I just never would have thought of it, because I was so preoccupied with the legal questions of theft, and how anything it says is almost always in the training data somewhere (or multiple somewheres).
I do think this ruling is a consequence of their efforts to convince people it is "intelligent". People wouldn't consider it something to listen to otherwise. It would just be an inconvenience, as it is for people who know better.
-
Google's defence needs to be amplified by anyone talking to politicians about 'AI' regulation:
Google is explicitly saying in their legal filing that the outputs from their LLM should not be trusted and that users should know that.
That's one hell of an admission. Imagine saying that about any other category of product.
@david_chisnall @tante I would love for someone to hold them to actually proving that statement.
Do people?
Does the general public, in fact, treat LLM output as just statistically likely text, not real information?
All of their advertisement sure seems to be designed to ensure people *don’t*, and instead put all their trust in this machine garbage.
