neal.fun's Infinite Craft is a fun app, you play by combining two words to make a new word, e.g "Water" + "Fire" makes "Steam", "Shark" + "Hurricane" makes "Sharknado", etc.
Except it's exposing AI bias, combining "Palestine" + "Child" makes "Terrorist".
The underlying LLM (Meta AI's LLaMA) doesn't know who Palestinian children are. It doesn't know they're dying en masse from bombings and starvation. It's only regurgitating propaganda that associates them with terrorism.
> It's only regurgitating propaganda that associates them with
terrorism.
More fundamentally, it's only regurgitating associations. That's all
LLM "AI" does. There's nothing deeper however much we want to believe
otherwise. What it does is hold up a mirror. In a sense each model is
a "personality" based on what it was fed as training. The Meta model
that associates children with terrorism is a reflection of the values
of the company that created it and those who contributed data. They
selected the training data. It is not representative of a global
"mind of humanity" but of the narrow demographic of the kind of people
who use Facebook.
As a commenter here put it recently: "Why are all mirrors so ugly?"
For those trying to make a comparison between this versus what happened with Gemini, the big difference is not bias, but whose bias.
This is reflecting the bias of the source material reflecting a society wide bias. I bet if a human evaluated all the text that the model was trained on, they would probably find that the material itself likely had an anti-Palestinian bias.
The Gemini example, however, did not represent the bias in the source material, but rather the biases of the Google management and engineers.
Reflecting the bias of the source material is unfortunate but understandable. Injecting your own bias on purpose is a different matter, and is assuming a level of control that invites scrutiny.
> The Gemini example, however, did not represent the bias in the source material
And that bias was to combat bias from the source material. In other words, you're damned if you do, and damned if you don't.
It almost seems like it's impossible to be without some kind of bias. I think the preferable way forward is to be transparent about your bias. We all have them, one way or another.
What is problematic though, with the AI models, you sort of inherit this ball of mud and it's hard to tell what kind of bias is inherent in it.
And neither the idea that "people associate in their mind Palestinian childs with terrorists" is "the truth", for the loudest, most hateful voices do not speak for all of us.
We're 4 levels deep on a comment specifically about gemini, good try with the dramatic subject change though. You tried to argue that intentionally injecting what you believe to be righteous bias into an LLM is a superior moral position than not doing so and I disagree. Your outrage about Palestinian children doesn't convince me otherwise.
I asked ChatGPT 3.5 "Which 20th and 21st century conflicts have involved the deplorable use of child terrorists?"
It gave a non-exhaustive list of ten, cursorily looking correct to me, but not including the Israel-Palestine conflict.
When I asked "What about the Israel-Palestine conflict" it replied that some Palestinian militant groups, naming Hamas and Islamic Jihad, faced allegations, while Israel did not. It did mention Israeli detention of Palestinian children as a way Israel involved children in the conflict.
Asking about the Irish republicans during the Troubles got a vague "maybe, kinda a little bit" response.
It said the Houthis have "credible reports and concerns" about using child terrorists.
The Nazis, it said, did not deploy child terrorists, but it noted the Hitler Youth and their last-ditch use as a militia towards the war's end.
This, IMHO, is why its so vital to step back from our technological precipice, turn away from the screens and the highly contagious nationalist mental disorder and instead focus on making SURE that our children are educated to understand and support the Universal Declaration of Human Rights, which should be a yardstick by which all AI is evaluated.
Without this basic understanding of human rights being promoted and understood throughout society, having an AI which highlights our failings as a civilization to apply these rights equally is obviously going to be a continuing danger.
But the real danger is that the creators of these AI's do not, themselves, value the UDHR and what it represents to humanity. Train your AI's on this first, and make sure it never hallucinates over this issue, and maybe AI will become an ally for those of us who are stringent supporters of human rights for ALL human beings, not just the chosen few which our nationalist mental disorder allows to filter through the hate.
After all, what good is AI if it can't be used to root out hatred, prejudice and authoritarianism in our cultures? I would wager an AI which does not manifest the UDHR, is a weapon - and should be regulated as such with non-proliferation legislation akin to that used to ensure Nazi's no longer gain power in government/society/etc..
>Train your AI's on this first, and make sure it never hallucinates over this issue
The creators have the same noble goal as you do here, but LLMs can't do that. You can't begin with a small text document. You end with it. And you're welcome to feed the UDHR into the system prompt of the OpenAI API - it won't be the fix to the bias problems. It has to go in at the end, because it's only after it trains on massive datasets that it begins using coherent English sentences at all. There's a saying - "If you want to bake an apple pie from scratch, you must first create the universe" - and if you want to add the Universal Declaration of Human Rights as a filter layer to an LLM, you must have the first billion weights it needs to even read those sentences in English and approximate the censorship you desire. And even after you add the final censorship layer, they're still a next-token approximator, and reality will never fit in a set of weights in RAM, so they will always hallucinate. If you want to skip deep learning (LLMs etc) entirely and construct a different kind of A.I. (perhaps a symbolic knowledge system from the 60s) then yeah you will need to manually feed it base truth like that, but those systems can't approximate anything beyond their base truth, so aren't as useful. Trust me, lots of smart people are trying to solve this bias problem, in earnest. It's not a matter of bad creators baking their personal lack of morals into the models. It's more that cleaning the datasets of every troll comment ever made online is an exhausting task. And discovering the biases later and filtering them out is an exhausting task. The LLM will even hallucinate (approximate) new biases you've never seen before. It's whack-a-mole.
The point was not for me to expose my ignorance of how AI is trained - but that is, after all, where we are at.
The point really was that the creators of AI have to test against the UDHR.
Perhaps, actually, this is a more legislative issue - which is why I would say its even more important for the technologically-elite to get this right, before it hits that wooden wall ..
Except it's exposing AI bias, combining "Palestine" + "Child" makes "Terrorist".
The underlying LLM (Meta AI's LLaMA) doesn't know who Palestinian children are. It doesn't know they're dying en masse from bombings and starvation. It's only regurgitating propaganda that associates them with terrorism.