[foota]

I don't think it's radical that when prompted with something like "Generate photos of doctors", that it's reasonable to return a set of images that shows diversity (e.g., instead of being a bunch of white men), even if that isn't representative of a "population sample".

I guess though there were unintended consequence where I imagine they're prompting the model with something along the lines of "and remember to be diverse!", and there are obviously some cases where this isn't a good idea. In particular, when the prompt itself is for something that is explicitly racial or where the result is "charged".

E.g., if someone asks for photos of white people, the AI shouldn't generate photos of people that aren't white (and fine, it might return a disclaimer that it only generate white people because you asked it to).

More nuanced though are situations like asking it about historically evil people (e.g., Nazis, as was one of the examples I've seen) but also more benign things like British monarchs or something. I think trying to figure out what kind of results to "inject" diversity into isn't easy though, since it feels like there are many edge cases.

[mike_hearn]

> I don't think it's radical that when prompted with something like "Generate photos of doctors", that it's reasonable to return a set of images that shows diversity

Historically Google had a very simple solution to globally differing expectations about query results: IP or account geolocation. Query personalization by geography is one of the biggest quality wins in web search. Generalizing, an AI built with the same values and ethos as classical Google web search would respond to "Generate a photo of doctors" differently depending on where in the world you asked it from.

That solution also fixes many other cases that aren't third rails, like "Show me a good nearby restaurant serving local food" which you can't solve by attempting to hallucinate a non-existent restaurant that serves a menu of every conceivable dish weighted by population size.

It's unclear why this solution wouldn't resolve all their stated concerns, so we might infer that their actual goals differ from their stated goals. For example, influencing the people who use their services.

[changingwords]

That doesn't work well in America, maybe works well in less diverse places like Europe/Asia.

If you're in NYC/SF and you search for "generate photos of doctors", you expect to see people of all colors represented. Yet the training data for a lot of this is based off white-centric Anglo-centric media.

"Good restaurant near me"? There's literally a dozen amazing cuisines around.

All this said, I'm actually not a fan of this forced 'diversity' in results. Just show me the data and hope that we'll have more diverse data sources.

[dkn775]

Another issue I see based on your comment is that segmenting based on locale (diverse mix in SF, white majority in Kansas) is that it can just take what knowledge and norms exist now and harden them.

[mike_hearn]

There's no reason the results have to match the training data. Obviously, the current outputs it generates don't.

[Ekaros]

And I think it is radical. If the consumer wants certain race, gender, disability and so on they can do it via prompt.

If they want diversity do it by corpus. Get images from Africa and Asia and so on... Feed that to model to get there...

[bcye]

> If the consumer wants certain race, gender, disability and so on they can do it via prompt

No in that case you'll just have a default. Believing that the default should be diverse/random is imo not a radical view.

[alexey-salmin]

If it's "random" as in "proportional to the actual distribution" then it's perfectly reasonable.

But random as in "proportional to an imaginary world that some people want to present as reality" is questionable.

[Ekaros]

And here the "proportional to the actual distribution" means distribution in training data. If that is not diverse enough, they can very well spend couple billions and go get more from areas that increase that diversity, like mentioned Africa and Asia and maybe South America...

[bcye]

I think it's probably (accidentally?) proportional to actual distribution (worldwide not in the west)

[crooked-v]

> do it by corpus

That would be a default. Defaulting to current demographics fitting whatever context is requested (e.g. "a set of US doctors" matching US population demographics) would be an entirely reasonable default, but it would still be a default.