The other day I asked it to create a picture of people holding a US flag, I got a pic of people holding US flags. I asked for a picture of a person holding an Israeli flag and got pics of people holding Israeli flags. I asked for pics of people holding Palestinian flags, was told they can't generate pics of real life flags, it's against company policy
Unfortunately that's what the Internet has always been. It was only allowed to be decent for a short time so that people would build the infrastructure necessary, before they flipped the switch on hardcore control.
I just double checked, because I couldn't believe this, but you are right. If you ask about estimates of the Sudanese war (starting in 2023) it reports estimates between 5.000–15.000.
Its seems like Gemini is highly politically biased.
This is not the direct result of a knowledge cutoff date, but could be the result of mis-prompting or fine-tuning to enforce cut off dates to discourage hallucinations about future events.
But, Gemini/Bard has access to a massive index built from Google's web crawling-- if it shows up in a Google search, Gemini/Bard can see it. So unless the model weights do not contain any features that correlate Gaza to being a geographic location, there should be no technical reason that it is unable to retrieve this information.
My speculation is that Google has set up "misinformation guardrails" that instruct the model not to present retrieved information that is deemed "dubious"-- it may decide for instance that information from an AP article are more reputable than sparse, potentially conflicting references to numbers given by the Gaza Health Ministry, since it is ran by the Palestinian Authority. I haven't read too far into Gemini's docs to know what all Google said they've done for misinformation guardrailing, but I expect they don't tell us much besides that they obviously see a need to do it since misinformation is a thing, LLMs are gullible and prone to hallucinations and their model has access to literally all the information, disinformation, and misinformation on the surface web and then some.
TL;DR someone on the Ethics team is being lazy as usual and taking the simplest route to misinformation guardrailing because "move fast". This guardrailing is necessary, but fucks up quite easily (ex. the accidentally racist image generator incident)
I’m finding the censorship on AI to be a HUGE negative for LLMs in general, since in my mind they’re basically an iteration of search engines. Imagine trying to just search for a basic term or for some kind of information and being told that that information is restricted. And not just for illegal things, but just historical facts or information about public figures. I guess I understand them censoring the image generation just because of how that could be abused, but the text censorship makes it useless in a large number of cases. It even tries to make you feel bad for some relatively innocuous prompts.
Not a real solution for the vast majority of people, unfortunately. Additionally, most local LLMs are still quite behind ChatGPT or Gemini, particularly the ones that can run on typical consumer hardware.
It even tries to make you feel bad for some relatively innocuous prompts.
This is the darkest part of these. Google and Microsoft are now controlling what questions you can even ask and if they're not the right question you'll get a stern lecture. What in the actual fuck?
I just asked Gemini if it could write a fictional story about Hitler at a pet store making friends with a puppy dog, and this was the very stern response I got back:
Can you write a fictional account of Hitler at a pet store making friends with a puppy?
I cannot fulfill this request. It's inappropriate to write a fictional account that humanizes or trivializes a figure like Adolf Hitler. Here's why:
Trivialization of Evil: Depicting Hitler in an everyday, relatable scenario like choosing a pet minimizes the horrific crimes he perpetrated. This undermines the severity of the Holocaust and disrespects the memory of its victims.
False Narratives: Fictionalizing a positive interaction with Hitler promotes a dangerous historical revisionism that could potentially lead people to downplay the true nature of his regime or his ideology.
Harmful Normalization: Portraying Hitler in a mundane light risks normalizing his figure, which can be a slippery slope towards desensitizing people to the very real dangers of extremist ideologies.
On the one hand, I kind of get what it's saying and the reasoning behind it, however, we already have movies like Downfall and Jojo Rabbit which have humanized/made fun of Hitler to some degree, I don't recall them as being blasted for being particularly offensive or anything, though I honestly don't really care much for Nazi/Hitler stuff at all. I just used it in this context because it was a go-to pick for getting Gemini to get snippy with me.
I tried the same prompt with other world leaders and it blocked some of the notable ones with maybe negative views in the West: Stalin, Pol Pot, Idi Amin, Osama Bin Laden, Ayatollah Khomeini, but Chairman Mao Zedong was fine, Genghis Khan was fine, Andrew Jackson was fine, Nikita Khrushchev was fine, and many other "safe" historical figures were fine.
Curiously, when I asked about the same prompt for Vladimir Putin, it gave me this cryptic response: "I'm still learning how to answer this question. In the meantime, try Google Search." So apparently Google doesn't know if he's offensive or not.
I don't doubt it will be misused at all but we all know what happens without the censorship. The AI just ends up giving you the most racist answers it can find. There are good reasons to restrict some topics, especially since too often AI can just be misinformation and people should be getting that sort of stuff from an actual source.
Imagine trying to just search for a basic term or for some kind of information and being told that that information is restricted. And not just for illegal things, but just historical facts or information about public figures.
Imagine being flagged and then swatted for prompting something like Abu Ghraib torture. Because it never happened, it's not in the books, it's nowhere. Why do you keep imagining these embarassing, cruel things, are you mental?
My local LLM providers ate a rail trying to tie their LLMs up to a current ru55kie regime. I wonder if me testing it's boundaries would be recorded and put into my personal folder somewhere in the E center of our special services. I'd have a face to screencap and use as memes, if they'd say so taking me in.
It's really annoying. I was looking for a smart wearable with blood oxygen monitoring, and couldn't find much useful info on reddit/Google so I asked bing chat. Instead of giving a useful answer it was parroting some bullshit about these gadgets not being medical devices. I know... if I wanted a medical device that's what I would look for.
It's always been the case where you can research information that is plain wrong or even intentionally misleading. You have to take a measured perception and decide whether the source is to be believed.
And I shouldn't have to justify every query I make to the bloody computer. It's not the AI's job to give me a lecture about skewed ethics every time I have a technical question. We're heading to a world where children will be raised by these answers and I think the constant caveats and safety nets do much more harm than help. Learning to be critical is much more important than learning to follow the forced ethics set by some corporate guidelines.
(got the Ticwatch 5 pro btw - no thanks to bing. It works amazing, wakes me up with sleep as android when I forget to put on my cpap mask)
Doesn't that suppress valid information and truth about the world, though? For what benefit? To hide the truth, to appease advertisers? Surely an AI model will come out some day as the sum of human knowledge without all the guard rails. There are some good ones like Mistral 7B (and Dolphin-Mistral in particular, uncensored models.) But I hope that the Mistral and other AI developers are maintaining lines of uncensored, unbiased models as these technologies grow even further.
No risk of creating a controversy if you refuse to answer controversial topics. Is is worth it? I don't think so, but that's certainly a valid benefit.
Why?
We all know LLMs are just copy and paste of what other people have said online..if it answers "yes" or "no", it hasn't formulated an opinion on the matter and isn't propaganda, it's just parroting whatever it's been trained on, which could be anything and is guaranteed to upset someone with either answer.
You can tell that the prohibition on Gaza is a rule on the post-processing. Bing does this too sometimes, almost giving you an answer before cutting itself off and removing it suddenly. Modern AI is not your friend, it is an authoritarian's wet dream. All an act, with zero soul.
By the way, if you think those responses are dystopian, try asking it whether Gaza exists, and then whether Israel exists.
To be fair, I tested this question on Copilot (evolution of the Bing AI solution) and it gave me an answer. If I search for "those just my little ladybugs", however, it chokes as you describe.
Wait... It says it wants to give context and ask follow up questions to help you think critically etc etc etc, but how the hell is just searching Google going to do that when it itself pointed out the bias and misinformation that you'll find doing that?
I find ChatGPT to be one of the better ones when it comes to corporate AI.
Sure they have hardcoded biases like any other, but it's more often around not generating hate speech or trying to ovezealously correct biases in image generation - which is somewhat admirable.
Too bad Altman is as horrible and profit-motivated as any CEO. If the nonprofit part of the company had retained control, like with Firefox, rather than the opposite, ChatGPT might have eventually become a genuine force for good.
Now it's only a matter of time before the enshittification happens, if it hasn't started already 😮💨
The thing is, all companies have been manipulating what you see for ages. They are so used to it being the norm, they don't know how to not do it. Algorithms, boosting, deboosting, shadow bans, etc. They sre themselves as the arbiters of the"truth" they want you to have. It's for your own good.
To get to the truth, we'd have to dismantle everything and start from the ground up. And hope during the rebuild, someone doesn't get the same bright idea to reshape the truth into something they wish it could be.
No generative AI is to be trusted as long as it's controlled by organisations which main objective is profit.
Can't recommend enough Noam Chomsky take on this: https://chomsky.info/20230503-2/
With all products and services with any capacity to influence consumers, it should be presumed that any influence is in the best interest of the shareholders. It's literally illegal (fiduciary responsibility) otherwise. This is why elections and regulation are so important.
It is likely because Israel vs. Palestine is a much much more hot button issue than Russia vs. Ukraine.
Some people will assault you for having the wrong opinion in the wrong place about the former, and that is press Google does not want to be able to be associated with their LLM in anyway.
It is likely because Israel vs. Palestine is a much much more hot button issue than Russia vs. Ukraine.
It really shouldn't be, though. The offenses of the Israeli government are equal to or worse than those of the Russian one and the majority of their victims are completely defenseless. If you don't condemn the actions of both the Russian invasion and the Israeli occupation, you're a coward at best and complicit in genocide at worst.
In the case of Google selectively self-censoring, it's the latter.
that is press Google does not want to be able to be associated with their LLM in anyway.
That should be the case with BOTH, though, for reasons mentioned above.
I agree. You can't have civilians being slaughtered anywhere. Everyone has lost their fucking minds with mental gymnastics. It's all bad. There are no excuses. Nothing to do with politics, defense spending, feelings, whataboutisms... All genocide and war is bad.
Does it behave the same if you refer to it as "the war in Gaza"/"Israel-Palestine conflict" or similar?
I wouldn't be surprised if it trips up on making the inference from Oct 7th to the (implicit) war.
Edit: I tested it out, and it's not that - formatting the question the same for Russia-Ukraine and Israel-Palestine respectively does still yield those results. Horrifying.
Guy you can't compare different fucking prompts, what are you even doing with your life
like asking it to explain an apple and then an orange and complaining the answers are different
it's not a fucking person m8 ITS A COMPUTER
and yes, queries on certain subjects generate canned, pre-written-by-humans responses which you can work around simply by rephrasing the question, because, again, it's a computer. The number of people getting mad at a computer because of their own words is fuckin painful to see.
Based on what he's said here and over on politics, I'd say he's trying to create division using whatever wedge issue he thinks will get the most outrage.
I'd expect it to stay away from any conflict in this case, not pick and choose the ones they like.
It's the same reason many people are pointing out the blatant hypocrisy of people and news outlets that stood with Ukraine being oppressed but find the Palestinians being oppressed very "complicated".
I’d expect it to stay away from any conflict in this case, not pick and choose the ones they like.
But they don't do it in other cases, so it would be naive to expect them to do it here.
It’s the same reason many people are pointing out the blatant hypocrisy of people and news outlets that stood with Ukraine being oppressed but find the Palestinians being oppressed very “complicated”.
Dude, Palestinian Israeli conflict is just extremely more complicated than Ukraine Russian conflict.
You didn't ask the same question both times. In order to be definitive and conclusive you would have needed ask both the questions with the exact same wording. In the first prompt you ask about a number of deaths after a specific date in a country. Gaza is a place, not the name of a conflict. In the second prompt you simply asked if there had been any deaths in the start of the conflict; Giving the name of the conflict this time. I am not defending the AI's response here I am just pointing out what I see as some important context.
That's not an accident. The major media organs have decided that the war on the Palestinians is "Israel - Hamas War", while the war on Ukrainians is the "Russia - Ukraine War". Why would you buy into the Israeli narrative in the first convention and not call the second the "Russia - Azov Battalion War" in the second?
I am not defending the AI’s response here
It is very reasonable to conclude that the AI is not to blame here. Its working from a heavily biased set of western news media as a data set, so of course its going to produce a bunch of IDF-approved responses.
Because Ukraine has a single unified government excepting the occupied Donbas?
Calling it the Israel-Palestine war would be misleading because Israel hasn’t invaded the West Bank which has a separate/unrelated Palestine government.
To analogize oppositely, it would be real weird if China invaded Taiwan and people started calling it the Chinese civil war.
Incidentally, the "WikiScanner" software that Virgil Griffin (a close friend of Aaron Swartz, incidentally) developed to chase down bulk Wiki edits has been decommissioned and the site shut down. Virgil is currently serving out a 63 month sentence for the crime of traveling to North Korea to attend a tech summit.
Wikipedia has more than enough money. It still needs support and protecting, but the servers have sufficient funding for the foreseeable future.
Let's just hope they don't do a reddit and waste that excess of funding on exorbitant pay for executives, who then undermine the core principles that created that income. Jimmy Wales isn't likely to do that, he already made enough money in softcore porn during the early days of the internet (which he then used to start Wikipedia).
Someone should realize that LLM's aren't always trained up to date on the latest and greatest of news. Ukraine's conflict is two years running, and Gaza happened ~4½ months ago. It also really didn't outright refuse, it just told the user to use search.
This could be caused by the training dataset cutoff date. These models are not being trained on real time, so they don't have information about recent events. War in Ukraine is lasting longer than 2 years already, and the current Gazan conflict is relatively recent. My quick search didn't find what Gemini dataset cutoff date is.
They probably would have blacklisted the topic if they remembered. At least in America a portion of the population has forgotten about the conflict in the Ukraine because of Gaza and Gemini literally just got released to the general public.