I don't see a description of how Gemini was trying to kill anyone. Rather I see a screenshot of a bunch of text ostensibly generated by Gemini, that uncritically following as if it was expert instructions might cause harm for a person doing such an ill-advised thing.
The model did not "suggest" anything. The model generated text. Google, or more specifically some Google employees (likely following incentives that Google management set up), presented the output of the model as if it might be authoritative answers to users' questions. Some users seemingly fall for the framing/hype and take the generated text as if it is authoritative information to be relied upon. This distinction is critical as we face an onslaught of "AI did <blank>" narratives, when the actual responsibility for such situations lies with specific humans relying on "AI" in such ways.
I think this is typical for software engineers (including myself) to view the world like this: this can't be murder, this is just words, you fell for it yourself. After all words alone are gibberish, the meaning arises in the head of the listener depending on how he interprets them.
However in most countries the justice system works differently. If your neighbor gives you this advice and it kills you, what will matter the most is not the authoritativeness of the advice (he's not a food expert, he was just talking) but the intent. If it also turns out that he had an affair with your wife and mentioned he wants to get rid of you, then he'll very likely go to jail. The "he fell for it and took my advice as if it was authoritative information to be relied upon" defense won't work.
Here (hopefully) the intent is missing. However the whole reasoning that advising partly bears no responsibility for texts it "generates" doesn't work very well.
You're not responding to the argument I'm making, but seemingly to other arguments that have been made over things like 4chan trolling that got people to cook their phones in microwaves.
I'm not saying that users are the only responsible parties here. I'm saying that the humans, including both the users and Google ('s managers and employees), are the only parties that can possibly be responsible. The LLM is not responsible. The LLM is incapable of having intent, as far as we currently know.
I've never used Gemini so I don't know how the user interface comes across, but it might be completely appropriate for Google to be facing liability lawsuits over things like this. What doesn't make sense is narratives that push any of this shared blame onto LLMs or "AI" itself, when the real world effects hinge entirely upon how humans choose to attach these algorithms to the real world.
Thanks for clarification, I agree this is a reasonable view of the situation.
Though I also don't really have problems with headlines like "AI tried to kill a man". True LLM can't be held responsible, but so are e.g. animals. If an unleashed dog kills a man, the owner will be responsible but the headlines will be still about the dog.
Then it can be argued that LLMs don't have intent while the dog had, and many more analogies, arguments and nuances will follow. My point is, this headline is still at an acceptable level of accuracy, unlike many other headlines that distort the reality completely.
The key difference is that everyone mostly understands a dog's capabilities. A reader knows there must be a larger context. Even in your simple example you felt the need to elaborate that it was an "unleashed" dog. If I merely say "a dog killed a man", the reader is left grasping for that larger context and defaulting to thinking a human must ultimately be responsible (most dogs are human owned rather than in wild packs). If I say "a bear killed a man", a similar grasping occurs but defaults to assuming the deceased likely got too close to a wild bear.
Society has no such general understanding for LLMs. If we did, we wouldn't have situations like this in the first place, because people wouldn't be expecting them to generate instructions that can be followed verbatim. Instead we get the marketing buzz of anthropomorphized "AI" as if they're super-human intelligent entities capable of conversing as humans do, but also better. And when they inevitably fail to live up to that, the narrative shifts to the need for "alignment" and "safety", which seems to mostly mean adding some words to the short priming prompt to make it avoid outputting some specific type of thing that was found to be problematic (with the definition of problematic generally being the creation a public relations shitstorm while also not creating profit). But us humans, our only benchmark for comparison, develop our sense of wisdom much slower through repeated learning and development stages. You can't simply read a child an essay of any length about how to drive safely and then hand them the keys to the family car.
https://darwinawards.com/ -- And yes, there is a quite meaningful difference. I know that some people are safety freaks, but I for one, will not grief for a single person that managed to kill themselves by following LLM instructions. There is a point of low IQ where we as a society can't really help anymore.
Moral considerations aside, I'm really skeptical of the assumption that following LLM instructions = low IQ.
As always, HN is a very biased audience. Most of us have probably read about how LLMs are the best bullshitters in the world, and we've seen a lot examples like this that prove how little LLMs can be trusted.
But in the general population? I'd bet there's plenty of intelligent people who haven't had the same exposure to LLM critiques and might trust their output out of ignorance rather than lack of intelligence.
I agree, and I think there might be an argument to be made that people with higher IQ are more prone to believing in (gramatically) well-written texts from authoritative-looking sources.
After all, people with higher IQ are used to reading books and believing them, whereas those with supposedly lower IQ tend to dismiss books and experts, instead believing in their own real-world experiences.
My only issue with concepts such as the Darwin Awards is that they do not reflect upon whether or not the deceased has reproduced. To attribute darwinism to lethal mistakes is meaningless unless we know whether or not the deceased has a child.
Some person with children who dies on an idiotic way has done more for evolution than the still alive and childless me.
Darwin Awards rewards people who are no longer able to reproduce, not taking into account any previous offspring. It'd be pretty hard to discover whether the subject of a random newspaper article had children or not.
So, if you drink and drive and kill a child, is it your responsibility, or is it the fault of alcohol being legal or vehicles not being mandates safe enough? This modern way of dealing with moral hazards is making me sad and afraid of others. After all, who knows whom they are going to blame for their own inability to perform sensible actions?
More like if you drink and then take paracetamol for the headache, the pharma corporation should warn you against performing this seemingly sensible action.
> The bottle being labeled is sufficient, just as the LLM being labeled as sometimes wrong is sufficient.
Labeling LLMs as "sometime wrong" is like labeling drugs as "sometimes have side effects, especially when mixing". It's a truism, such label would be completely useless. You need to take the drug anyway, so knowing that "some side effects exist" doesn't change anything. And often you also do need to take 2-3-4 drugs at same time, so the mixing clause is not helping either.
It took us many decades to build the system around drugs that today you take as granted. Drugs are studied very carefully, with all observed side effects listed explicitly. Drugs compatibility is a big topic and all new drugs are checked in combination with common-use drugs.
At the other end of equation awareness of side effects and mixing was increased dramatically both among doctors and patients, who were previously completely oblivious of them. Mere 100 years ago people were prescribed heroin against cough. Only 60 years ago the thalidomide disaster happened.
If all you can say to people destroyed by Kevadon is "you are at fault, the bottle said CAUTION: new drug", then I'm afraid we see this too differently.
I don't think that's a fair assessment. Garlic and olive oil can both be stored without refrigeration. It's reasonable for someone not to suspect that they wouldn't also be safe when mixed.
The point is there are absolutely no guarantees about the output of an LLM being semantically sensible or logical. It's appropriate for brainstorming and discovery, but any ideas/plans/etc you get from it you need to analyze (either yourself, or relying on an expert). The idea of letting a mixture of garlic and olive oil sit around needs to be examined for safety regardless of whether you came up with the idea using only your brain, or if you were inspired by an LLM, or by 4chan.
I suspect if you haven't experienced how effective LLMs are at making coherent prose that is nevertheless utterly incorrect, you either haven't been asking very technical questions or you've been quite lucky.
On a related note, I have never seen a person trying to kill anyone. Rather I have seen movies about a bunch of things that are ostensibly people holding guns, that uncritically standing in the way of their shot paths as if the guns were toys might cause harm for a person choosing to stand in such a foolish place.
English is such a fun language sometimes! So many idioms.
You jest, but I remember reading a case about a guy who was convicted on 9 counts of attempted murder and managed to get it reduced to one on appeal by pointing out that he only fired one bullet toward a group of 9 people. (Nobody was actually hit.)
