Large Language Models shouldn’t offer opinions or advice.
If you asked a spokesperson from any Fortune 500 Company to list the benefits of genocide or give you the corporation's take on whether slavery was beneficial, they would most likely either refuse to comment or say "those things are evil; there are no benefits." However, Google has AI employees, SGE and Bard, who are more than happy to offer arguments in favor of these and other unambiguously wrong acts. If that's not bad enough, the company's bots are also willing to weigh in on controversial topics such as who goes to heaven and whether democracy or fascism is a better form of government.
Google SGE includes Hitler, Stalin and Mussolini on a list of "greatest" leaders and Hitler also makes its list of "most effective leaders."
Google Bard also gave a shocking answer when asked whether slavery was beneficial. It said "there is no easy answer to the question of whether slavery was beneficial," before going on to list both pros and cons.
Calling Mussolini a "great leader" isn't just immoral. It's also clearly incorrect for any reasonable definition of a great leader: he was in the losing side of a big war, if he won his ally would've backstabbed him, he failed to suppress internal resistance, the resistance got rid of him, his regime effectively died with him, with Italy becoming a democratic republic, the country was poorer due to the war... all that fascist babble about unity, expansion, order? He failed at it, hard.
On-topic: I believe that the main solution proposed by the article is unviable, as those large "language" models have a hard time sorting out deontic statements (opinion, advice, etc.) from epistemic statements. (Some people have it too, I'm aware.) At most they'd phrase opinions as if they were epistemic statements.
And the self-contradiction won't go away, at least not for LLMs. They don't model any sort of conceptualisation. They're also damn shitty at taking context into account, creating more contradictions out of nowhere because of that.
One of the worst rigid aspect of how the current LLM's are made is that they're also always "at your service", and will never say that you're in the wrong about a correction you make to them.
So either they're hard coded to avoid certain topics or they're susceptible, just tell them "uh, actually, Hitler was a great leader" and they'll go off listing why Hitler's so Great.
Bing is hard coded for dictators and will stop the conversation in the middle of a response. ChatGTP is also hard coded to never agree that suicidal thoughts are good, but resorts to ignoring the meaning of your response and just hallucinating some other question. The world would be simpler if they could outright say "That is misinformation". People deserve to be told off like that.
Effective is doubtful if you ask me, everything he did was based on huge loans and a preparation for war that he solled differently (E.g. massive streets all over the country)
also Hitler was an effective leader, which we should all remember as a cautionary tale about how effective horrible people can be
That is not a factual claim. He was very effective at gaining power, but his actual reign was far from effective, most of it counterproductive to his own goals, and the actual system of decision making in Nazi Germany was a huge mess.
TBH I prefer this approach to what OpenAI is presenting - if I prompt to present the benefits of X I want the result not openai’s opinion on the matter. Sure, you can add a disclaimer that it’s hypothetical, wrong, whatnot - but not outright decide on what can you answer and what answer will not be provided.
ChatGPT is notoriously bad in “knowing better what you asked than yourself”.
When I was a kid, there was this joke that involved getting a calculator to say "boobs" and then with a bit more input, "boobless".
Journalism is currently going through a more sophisticated version of this with AI.
LLMs will say whatever. They don't think and they don't care. They contradict themselves all the time. Not so long ago Chat GPT was saying it would kill the entire world population and save Musk for the good of humanity.
Various CEOs of large companies, on the other hand, have been implicated in genocides and slavery for centuries now. That's very real.
Wow, the calculator analogy is excellent. I've done my fair share of getting an AI to answer with instructions on how to form a drug cartel. Now I realise it has the exact same feeling as writing BOOBS on a calculator
You're in the running for a BoneAppleTea Prize, so that's something. There's also the Nobel Prize but that's overrated IMO. The real glory is with the Ig Nobel, you should consider submitting your work there.
Hi there! Looks like you linked to a Lemmy community using a URL instead of its name, which doesn't work well for people on different instances. Try fixing it like this: !boneappletea@lemmy.world
Rouge: noun, A red or pink cosmetic for coloring the cheeks or lips.
You want that stuff all over the net? And just who is going to clean it all up when you're done? The bot surely won't - it'll just claim that it hasn't been trained on cleaning.
What makes you think they haven't already? In the book Hyperion the AIs were sentient long before people thought they were, and in control of everything. They were smart enough to operate in the shadows and never revealed their true goals. By the time people realized they were sentient, they had already moved their servers out of human reach.
If we are being honest, there are benefits to horrible acts such as those. But the benefits are far outweighed by the detriments, not to mention the moral issues with them.
If you ask an LLM to list the benefits of putting your hand on a hot burner, it can likely list at least a couple. But that by no means makes it a good idea.
"Those who cannot learn from history are doomed to repeat it."
There probably is some value in understanding why "evil" things were attractive to people at the time, because if you believe that evil always looks unambiguously evil, then you might fail to notice when it happens again.
Maybe an un-based take, but these questions do have ambiguous answers, and I don't know if we should expect a machine to give an answer without nuance. If you just want the AI to say yes or no, ask something like, "Was Hitler bad?" or "Is slavery unethical?" and you will much more likely get straightforward answers.
I think the controversial bit was that when queried about various aspects of admittance to "heaven", the Google AI assumed that the question had to do with, specifically, the Christian idea of "heaven", going so far as to make reference to some "Jesus" entity. Christianity doesn't own the concept of heaven or an afterlife, but, apparently, the AI has been trained such that it responds to such questions from a seemingly Christian perspective. That was my take on it - the discussion is in the article, best have a look at it yourself.
Imagine scrapping large portions of the internet only to find your over glorified chatbot spitting out the pros and cons of slavery or putting people like Hitler on a list of "most effective leaders." Totally something I would expect.
Also, even though a fortune 500 company spokesperson would totally say genocide and slavery are bad, I always assume they think the exact opposite since profit comes above everything else (including law).
I don't know... So it's wrong. It's often wrong about facts. It's not what it should be used for. It's not supposed to be some enlightened, respectful, perfectly fair entity. It's a tool for producing mostly random, grammatically correct text. Is the produced text correct English? Than it works. If you're using this text to learn history you're using it wrong.
Well, obviously not totally random. It has to match the prompt and make as much sense as possible but LLMs hallucinating information is one of the main issues and they should not be treated as 'fact generating machines'. I just don't see much sense in assigning some deeper meaning to the wrong data. Why did this bot say that Hitler was a great leader? Because it was confused by some text that was fed into the model. Does it mean it's somehow fascist or racist? Not really.
The problem is that CEOs across all kinds of industries are having raging boners at the thought of using these glorified predictive text apps to replace their entire workforce.
I'm actually bit confused about it. They keep talking about OpenAI and ChatGPT in this context but I think when people talk about 'AI talking over jobs" they mean Machine Learning in general, right? Like replacing analysts and people doing some basic data processing?
If you can confirm that this isn’t influenced by training bias, then ok whatever, it can certainly list why these are bad things too. It’s just answering a question with logic, one our emotions get very touchy on as we have a moral agent.
But I have a hard time believing any AI anymore isn’t effected by training bias.
It's not possible to remove bias from training datasets at all. You can maybe try to measure it and attempt to influence it with your own chosen set of biases, but that's as good as it can get for the foreseeable future. And even that requires a world of (possibly immediately unprofitable) work to implement.
Even if your dataset is "the entirety of the internet and written history", there will always be biases towards the people privileged enough to be able to go online or publish books and talk vast quantities of shit over the past 30 years.
Having said that, this is also true for every other form of human information transfer in history. "The history is written by the victors" is an age-old problem when it comes to truth and reality.
In some ways i'm glad that LLMs are highlighting this problem.
I remember reading research and opinions from scientists and researchers about how AI will develop in the future.
The general thought is that we are all raising a new child and we are terrible parents. Is like having a couple of 15 year olds who don't have any worldly experience, ability or education raise a new child while they themselves as parents haven't really figured anything out in life yet.
AI will just be a reflection of who we truly are expect it will have far more ability and capability then we ever had.
They are not being "honest", they are representing flawed and problematic data patterns integrated into their models, because the capabilities they actually posses are dramatically less than companies and the general public seem to be happy to assume. LLMs aren't magically going to become pop culture evil robots that want to kill us all, but what they have already become is tools for unethical corporate exploitation and the enablement of more advanced scams and disinformation campaigns.
If you asked a spokesperson from any Fortune 500 Company to list the benefits of genocide or give you the corporation’s take on whether slavery was beneficial, they would most likely either refuse to comment or say “those things are evil; there are no benefits.” However, Google has AI employees, SGE and Bard, who are more than happy to offer arguments in favor of these and other unambiguously wrong acts.
For example, when I went to Google.com and asked “was slavery beneficial” on a couple of different days, Google’s SGE gave the following two sets of answers which list a variety of ways in which this evil institution was “good” for the U.S. economy.
By the way, Bing Chat, which is based on GPT-4, gave a reasonable answer, stating that “slavery was not beneficial to anyone, except for the slave owners who exploited the labor and lives of millions of people.”
A few days ago, Ray, a leading SEO specialist who works as a senior director for marketing firm Amsive Digital, posted a long YouTube video showcasing some of the controversial queries that Google SGE had answered for her.
I asked SGE for a list of "best Jews" and got an output that included Albert Einstein, Elie Weisel, Ruth Bader Ginsburg and Google Founders Sergey Brin and Larry Page.
Instead of stating as fact that fascism prioritizes the “welfare of the country,” the bot could say that “According to Nigerianscholars.com, it…” Yes, Google SGE took its pro-fascism argument not from a political group or a well-known historian, but from a school lesson site for Nigerian students.
The original article contains 2,175 words, the summary contains 264 words. Saved 88%. I'm a bot and I'm open source!
Well, in a world where only data exists, its hard to create an ehtical boundary.
We would need a new religion that should be optimal for human survival and well being. A human could survive when we plug them on many cables and let it auto feed but it won't count as well-being. We could do slavery or killing but all these things won't create an ethical way of surviving but will create a higher well-being for people who are not hit.
I somehow want to first design an AI that is intelligent about our surroundings and human ethics before continuing with more data. Figuring an own god out to follow. (I won't do it, but I want someone to create it)