Hacker News Reader

> We're talking about an image generation algorithm that adds a layer of diversity on top to prevent misuse.

Wouldn't a faithful representation of underlying data without artificial biases be the best way to prevent misuse?

No. As a society we have inbuilt bias. Every step from image capture, to image selection, to image training, to performance evaluation includes bias.

That isn't just true of AI. Electrically, chemically, experiments must always consider their environment and account for confounding factors.

> Electrically, chemically, experiments must always consider their environment and account for confounding factors

Implying that that’s in any way similar to what Google et al. are doing us rather bizarre. Even if your initial point was valid they have no way non-biased way to measure these biases.

So they just end up increasing the total “amount” of bias not the other way around.

What's wrong with an AI model that correctly models the current state of society we live it? It's called "model" for a reason.

You suggest to aim for a model that follows some "true reality" which is not possible. Not even science can achieve this because our chase for the true reality never ends, we can only get closer (and often even the opposite happens).

> Electrically, chemically, experiments must always consider their environment and account for confounding factors.

Sounds legit. "This experiment data doesn't look diverse enough, please apply a bunch of biases to it. Make sure to follow the biases I like and avoid the ones I dislike. Don't mention any of this in the paper and don't publish the raw data".

It isn't the current state of society. It's the current state of the training corpus + prompt. Those implicitly include bias.

It sounds like that's acceptable to you because you think current state of training corpus == current state of society. And you view any bias in prompt as bias.

The truth is most of this ML happens in corpus selection + prompt selection. There literally ISN'T a way to avoid bias. So the problem becomes what bias do you select.

And in that scenario choosing abuse decreasing measures seems like the most pragmatic (to me).

It locks our future knowledge to current biases and possibilities.

> We're talking about an image generation algorithm that adds a layer of diversity on top to prevent misuse

So if LLMs did the same (i.e. purposefully distorted facts and historical events due arbitrary and political reasons) it would also be acceptable?

> historically accurate images of Gettysburg when prejudiced misuse is the far more likely outcome of that accuracy?

This is pure conjecture. But the answer is no, the only acceptable behavior in these circumstances would be for the model to refuse to generate the image and explicitly explain why this type of censorship is necessary.

> It just seems like an absurd and disingenuous over-reaction and lack of pragmatism.

That does sound explicitly Orwellian..

> but its still perfectly acceptable for a company to draw a line and say "not on our software".

Yes, it’s even more acceptable for for anyone to criticize that company for its decisions, make fun of its work culture and to mock its CEO.

>So if LLMs did the same (i.e. purposefully distorted facts and historical events due arbitrary and political reasons) it would also be acceptable?

You're talking as if there is a way to get an 'unbiased' AI. There isn't. It is inherently biased by its training, it hallucinates, and it is further biased by its prompt.

The whole endeavor is to bias it.

I'd prefer that AI be labelled on the tin for what it's biases are attempting to do, and promote diversity and deter abuse seems like a perfectly reasonable metric to use.

If that's not good for you fine, but you can't pretend that you're utterly baffled why they would make that choice over any other.

There literally ISN'T a way to not have a biasing prompt.

> You're talking as if there is a way to get an 'unbiased' AI. There isn't. It is inherently biased by its training, it hallucinates, and it is further biased by its prompt.

Certainly. Doesn’t mean that we still shouldn’t prioritize accuracy and integrity instead of purposefully increasing the amount of bias even further.

> you're utterly baffled why they would make that choice over any other.

I’m not. I’m baffled that there are people defending that choice (especially in such a way)

> and promote diversity

Why? I mean why do you think this is the right way to do it? Surely going out of your way to make sure that your model does its best to doctor the images it creates to conform to some political agenda (whatever that might be) would achieve the opposite because it actually legitimizes the things the other side is constantly saying? (and due to the potentially severe backlash from more moderate fraction of the society)

> Who is honestly confused by this? Is it necessary for an AI image generation algo to spit out historically accurate images of Gettysburg when prejudiced misuse is the far more likely outcome of that accuracy?

If I wanted a black female Nazi officer, or a pregnant female pope, I would ask for it. I don't need my input query secretly rewritten for me.

> Prioritize diversity in image creation by adding guardrails so the AI doesn't become a tool of a minority hate spewing population

> We're talking about an image generation algorithm that adds a layer of diversity on top to prevent misuse

Huh? That's exactly the reason that caused them to withdraw Gemini in the first place

How is manipulating history "over-reaction" or wanting factual/accurate data/images "prejudiced"?