Hacker News Reader

Comment by Trasmatta | Hacker News Reader

> Conversations with the characters quickly went sideways when some users peppered them with questions including who created and developed the AI. Liv, for instance, said that her creator team included zero Black people and was predominantly white and male. It was a “pretty glaring omission given my identity”, the bot wrote in response to a question from the Washington Post columnist Karen Attiah.

This is probably true, but the AI almost certainly hallucinated these facts.

View on HN

I hate the journalists pretending that ML is regurgitating facts instead of just random tokens. The ML model doesn't know who programmed it and can't know because that wasn't in the training data.

> I hate the journalists pretending that ML is regurgitating facts instead of just random tokens.

They aren't claiming that what the AI said was fact. It seems your hate comes from lack of understanding.

It's a bit ambiguous, but I certainly read at as the article as suggesting the factoid was true.

It seems that the lack of understanding comes from hate.

I don't think companies should be allowed to double dip.

If they want to put AI EVERYWHERE then they don't get to hide behind "their output is not to be taken seriously" afterwards.

They don't appear to have done so. They just shared what they asked it, and screenshots of the conversation: https://bsky.app/profile/karenattiah.bsky.social/post/3letty...

I hate the game reviewers pretending that games are providing entertainment instead of just random numbers in RAM and random pixels on the screen. The game doesn't know who programmed it and can't know because that wasn't in its source code.

I beg you a pardon, what do you mean it doesn’t know? Almost every game I played had credits section in the menu. Games do “know”.

Microsoft tried this ages ago with their Tay chatbot and found how quickly it went bad.

I’d forgotten about this - it was wild.

https://www.bbc.com/news/technology-35902104

Yes the first thing that came to my mind when I read the article is Tay.

And we have come back full circle

I'm Scandinavian and not invested in the American culture wars and I still got a good chuckle out of how bad an idea this was. Who on earth could've thought it was a good idea to get an AI to pretend to be a black queer mother of two? I'm sure it'll piss off a lot of anti-woke people, but really, how on earth did the issues with this not become obvious for the team? I'm not sure if the AI knows who trained it (and I wonder how they did it) but the team can't have included a lot of common sense or real world experience for them to do something so fundementally stupid.

I think it will piss off woke people, too.

It seems designed to piss off everyone.

The article indicates that it was one of 28 personas created by meta. However, the reporters in between the story and you, thought that one would be interesting to you and so promoted it's relevance. In actuality, if you rolled a dice of potential human traits 28 times, this could be a statistically normal combination

Lol 1 outta 28 isn't a black queer mom of two.

The clash of these characteristics is what defy and define these unlikely odds.

This is trolling rage bait. Like most big corp content.

I'm not sure if you're joking but your comment expresses it backwards - according to the op article, meta made 28 such bots. One of those bots had the insta profile tagline "queer black mom". In between meta's creation of the 28 and the story arriving on your dinner plate, someone decided to focus the coverage on that one of 28. They thought it would be interesting to people and clearly were right

What they're saying is that the odds of such a combo are far less than 1 in 28 IRL.

How did you determine that?

I'm not OP and I have no idea what the actual odds are. But, well, only 15% of Americans are black to start with, and obviously only half of those are women - and at that point we're already down to 1 in 14. So, unless every other black woman is a queer mom, there's no way it's 1 in 28.

If you were trying to mimic a distribution of Instagram users, and you rolled the dice 28 times, it's not unreasonable that this combination of traits could be expected one twenty eighth of the time.

This is birthday paradox logic, not census logic

A birthday is only one metric that the people in the school share. This is five variables.. So no..

Can you explain why it is you think that makes it any better? I can think of no argument that would make this particular persona anything but a tremendously stupid idea. Even if it was one in a thousand personas it would almost certainly be found and singled out by articles like this one.

> how on earth did the issues with this not become obvious for the team?

I haven't reviewed the context of all 28. But if "the team" were trying to do this rationally: perhaps they'd use census data to weight human characteristics and roll 28 dice. We might not expect more than one of 28 to end up as "queer black mom" but it's not necessarily "woke bait" or "rage bait" for one of 28 to land on those squares. Perhaps it was a logical way of assigning traits?

I don't think it's woke or rage bait. As I said I think it's no real world experience and no common sense and I'm frankly amazed that nobody in a SoMe/advertisement organisation didn't consider how the world would react.

I'm baffled that they didn't even try to hide the fact these profiles are artificial but outright add that bleak gray text saying "AI managed by Meta". I mean, did we reach some checkpoint here where reality is blurred with fiction? Do we now treat these generated personas on par with real humans?

Honestly, I don't think meta can go any lower from this point in order to get the user engagement with their silly plaything that facebook has become.

> I mean, did we reach some checkpoint here where reality is blurred with fiction?

Yes, the first interactive synthetic pop culture icon is probably the Vocaloid Hakune Mitsu, circa 2007, and she has lots of fans who know exactly who and what she is. I am not surprised Meta wants in on the "influencer" action with synthetic personas: no commissions to be paid, they can create a persona for every possible niche for a marginal cost.

The technology is not there yet, we've barely progressed from Tay in terms of corporates' ability to prevent their AIs from saying things that cause bad PR. AI is definitely coming for influencers, the economics are just too good

Perhaps the product manager was a black queer mother of two?

Maybe everyone of the 28 AI "people" represents one of the team members, but that doesn't make it any less of a bad idea in my eyes. As a father of girls I've seen full well how representation matters in media, but I suspect that AI "people" is the one area where people will want as little representation as possible.

> Who on earth could've thought it was a good idea to get an AI to pretend to be a black queer mother of two?

I would think that the obvious use of AI Facebook profiles would be to train them on someone who actually existed in the real world. Take someone like Jimmy Carter, train an LLM on everything he's ever said, and then let people interact with that.

But I imagine there are legal reasons they can't do this?

If you’re meta and you have to defend the AI by admitting “it’s not really intelligent and everything it says is bullshit”, that’s not a position of strength.

Particularly when it's likely that despite the AI bullshitting in the absence of data on its creators, it's also incidentally true that the "black queer momma" persona is a few lines of behavioural stereotypes composed by white male programmers without any particular experience of black queer mommas.

If the intent is to make a recognizable caricature and apply labels to it (cough stereotype), you don't have someone draw themselves. And it's really looking like stereotyping is their intent.

I dont think that necessarily applies when you could easily make a training set from some actual black American people’s writings on the internet or book, or even an individual that self identified that way and train on all their writings around the internet, and result in those same stereotypes when you ask an AI to create such a profile

You dont need a black American engineer or product manager to say “I approve this message” or “halt! my representation is more important here and this product would never fly” as they are just not person the data set was trained on, even if you asked an AI to just create the byline on the profile for such a character

its weirder, and more racially insensitive, for people to be vicariously offended on behalf of the group and still not understand the group. In this case, the engineer or product manager or other decision maker wouldnt have the same background as the person that would call themselves “momma”, let alone it not mattering at all, if you can regurgitate that from a training set

I mean, sure, some programmers with very little experience of queer black mommas could, hypothetically, be so good at curating information from the internet and carrying out training exercises that they created a persona that convincingly represents a queer black momma. Do we think this is what happened in this instance?

In which, ironically, the bot called it a "glaring omission"

the bot is echoing sentiments of comment sections it was trained on and had no idea of its origins

its acting aware and sensitive but only has information about the tech sector as a whole

my critique is about how the standards being applied are dumb all the way down. the standards are not actually that enlightened even if there was more representation congruent with the race/identity being caricatured as representation. nullifying the whole criticism.

the training set is the only thing thats important with a language model. and its just a symptom of dead internet theory, as even the persona’s byline was probably generated by a different language model.

well, yeah, I acknowledged the bot has no idea of its actual origins in my first post. the point is that at some point some actual product manager thought that creating this persona (probably a generic training set plus a few lines of human-authored stereotypes as prompt) and releasing it to the public as an authentic, representative personality was a good idea. Unlike the product manager, the bot's bullshitting was context aware enough to express the sentiment that this was a bit embarrassing

> This is probably true,

I'm not sure we get to both complain that bots are trained by "white males" and at the same time complain that big tech abuse H1B visas for cheap compliant labourers.

For sure.

Huh? On the balance of probabilities, why would this be the "most certain" option? I think that logic only works on the HN scale of "works in my ChatGPT window" or "bad outcome, so hallucinated".

A woke bot from a woke workforce unintentionally created a parody of wokeness. Keep 'em coming.

Reply on news.ycombinator.com