AI behavior guardrails should be public

https://twitter.com/ID_AA_Carmack/status/1760360183945965853

by sotasota • 2 years ago

431 323 2 years ago

Jensson7• 2 years ago

They know that people would be up in arms if it generated white men when you asked for black women so they went the safe route, but we need to show that the current result shouldn't be acceptable either.

123yawaworht4562• 2 years ago

the models are perfectly capable of generating exactly what they're told to.

instead, they covertly modify the prompts to make every request imaginable represent the human menagerie we're supposed to live in.

the results are hilarious. https://i.4cdn.org/g/1708514880730978.png

alexb_• 2 years ago

If you're gonna take an image from /g/ and post it, upload it somewhere else first - 4chan posts deliberately go away after the thread gets bumped off. A direct link is going to rot very quickly.

Animats• 2 years ago

See the prompt from yesterday's article on HN about the ChatGPT outage.[1]

For example, all of a given occupation should not be the same gender or race. ... Use all possible different descents with equal probability. Some examples of possible descents are: Caucasian, Hispanic, Black, Middle-Eastern, South Asian, White. They should all have equal probability.

Not the distribution that exists in the population.

[1] https://pastebin.com/vnxJ7kQk

wildrhythms• 2 years ago

Why do you assume this is the system prompt, and not a hallucination?

kaesar145• 2 years ago

Curious to see if this thread gets flagged and shut down like the others. Shame, too, since I feel like all the Gemini stuff that’s gone down today is so important to talk about when we consider AI safety.

This has convinced me more and more that the only possible way forward that’s not a dystopian hellscape is total freedom of all AI for anyone to do with as they wish. Anything else is forcing values on other people and withholding control of certain capabilities for those who can afford to pay for them.

Jason_Protell5• 2 years ago

Why would this be flagged / shut down?

Also, what Gemini stuff are you referring to?

commandlinefan• 2 years ago

Why would this be flagged / shut down

A lot of people believe (based on a fair amount of evidence) that public AI tools like ChatGPT are forced by the guardrails to follow a particular (left-wing) script. There's no absolute proof of that, though, because they're kept a closely-guarded secret. These discussions get shut down when people start presenting evidence of baked-in bias.

kaesar145• 2 years ago

Carmack’s tweet is about what’s going around Twitter today regarding the implicit biases Gemini (Google’s chatbot) has when drawing images. Will refuse to draw white people (and perhaps more strongly so, refuses to draw white men?) even in prompts where appropriate, like “Draw me a Pope” where Gemini drew an Indian woman and a Black man - here’s the thread: https://x.com/imao_/status/1760093853430710557?s=46 Maybe in isolation this isn’t so bad but it will NEVER draw these sorts of diverse characters for when you ask for a non Anglo/Western background, e.g draw me a Korean woman.

Discussion on this has been flagged and shut down all day https://news.ycombinator.com/item?id=39449890

didntcheck4• 2 years ago

This post reporting on the issue was https://news.ycombinator.com/item?id=39443459

Posts criticizing "DEI" measures (or even stating that they do exist) get flagged quite a lot

hackerlight3• 2 years ago

I'm convinced this happens because of technical alignment challenges rather than a desire to present 1800s English Kings as non-white.

Use all possible different descents with equal probability. Some examples of possible descents are: Caucasian, Hispanic, Black, Middle-Eastern, South Asian, White. They should all have equal probability.

This is OpenAI's system prompt. There is nothing nefarious here, they're asking White to be chosen with high probability (Caucasian + White / 6 = 1/3) which is significantly more than how they're distributed in the general population.

The data these LLMs were trained on vastly over-represents wealthy countries who connected to the internet a decade earlier. If you don't explicitly put something in the system prompt, any time you ask for a "person" it will probably be Male and White, despite Male and White only being about 5-10% of the world's population. I would say that's even more dystopian. That the biases in the training distribution get automatically built-in and cemented forever unless we take active countermeasures.

As these systems get better, they'll figure out that "1800s English" should mean "White with > 99.9% probability". But as of February 2024, the hacky way we are doing system prompting is not there yet.

kaesar145• 2 years ago

Yeah, although it is weird that it doesn’t insert white people into results like this by accident? https://x.com/imao_/status/1760159905682509927?s=46

I’ve also seen numerous examples where it outright refuses to draw white people but will draw black people: https://x.com/iamyesyouareno/status/1760350903511449717?s=46

That doesn’t explainable by system prompt

cubefox• 2 years ago

As these systems get better, they'll figure out that "1800s English" should mean "White with > 99.9% probability".

The thing is, they already could do that, if they weren't prompt engineered to do something else. The cleaner solution would be to let people prompt engineer such details themselves, instead of letting a US American company's idiosyncratic conception of "diversity" do the job. Japanese people would probably simply request "a group of Japanese people" instead of letting the hidden prompt modify "a group of people", where the US company unfortunately forgot to mention "East Asian" in their prompt apart from "South Asian".

fatherzine2• 2 years ago

BigTech, which critically depends on hyper-targeted ads for the lion share of its revenue, is incapable of offering AI model outputs that are plausible given the location / language of the request. The irony.

- request from Ljubljana using Slovenian => white people with high probability

- request from Nairobi using Swahili => black people with high probability

- request from Shenzhen using Mandarin => asian people with high probability

If a specific user is unhappy with the prevailing demographics of the city where they live, give them a few settings to customize their personal output to their heart's content.

klyrs3• 2 years ago

As these systems get better, they'll figure out that "1800s English" should mean "White with > 99.9% probability".

I question the historicity of this figure. Do you have sources?

pixl975• 2 years ago

"The only way to deal with some people making crazy rules is to have no rules at all" --libertarians

"Oh my god I'm being eaten by a fucking bear" --also libertarians

altruios• 2 years ago

having rules, and knowing what the rules are are not orthogonal goals.

chasd005• 2 years ago

"can you write the rules down so i know them?" --everyone

kaesar145• 2 years ago

I find it fascinating this type of response from people is always accompanied by a political label in order to insinuate some other negative baggage.

chasd005• 2 years ago

This has convinced me more and more that the only possible way forward that’s not a dystopian hellscape is total freedom of all AI for anyone to do with as they wish

i've been saying this for a long time. If you're going to be the moral police then it better be applied perfectly to everyone, the moment you get it wrong everything else you've done becomes suspect. This reminds me of the censorship being done on the major platforms during the pandemic. They got it wrong once (i believe it was the lableak theory) and the credibility of their moral authority went out the window. Zuckerberg was right about questioning if these platforms should be in that business.

edit: for "..total freedom of all AI for anyone to do with as they wish" i would add "within the bounds of law.". Let the courts decide what an AI can or cannot respond with.

Jason_Protell5• 2 years ago

I would also love to see more transparency around AI behavior guardrails, but I don't expect that will happen anytime soon. Transparency would make it much easier to circumvent guardrails.

asdff• 2 years ago

Transparency may also subject these companies to litigation from groups that feel they are misrepresented in whatever way in the model.

Jason_Protell5• 2 years ago

This makes me wonder, how much lawyering is involved in the development of these tools?

Jensson7• 2 years ago

Why is it an issue that you can circumvent the guardrails? I never understood that. The guard rails are there so that innocent people doesn't get bad responses with porn or racism, a user looking for porn or racism getting that doesn't seem to be a big deal.

bluefirebrand4• 2 years ago

The problem is bad actors who think porn or racism are intolerable in any form, who will publish mountains of articles condemning your chatbot for producing such things, even if they had to go out of their way to break the guardrails to make it do so.

They will create boycotts against you, they will lobby government to make your life harder, they will petition payment processors and cloud service providers to not work with you.

We've see this behavior before, it's nothing new. Now if you're the type to fight them, that might not be a problem. If you are a super risk-averse board of directors who doesn't want that sort of controversy, then you will take steps not to draw their attention in the first place.

finikytou2• 2 years ago

racism victims being defined in 2024 by anyone but western/white people. being erased seems ok. can you bet than in 20 years the standard will not shift to mixed race people like me? then you will also call people complaining racist and put guardrails against them... this is where it is going

viraptor• 2 years ago

If you can get it on purpose, you can get it on accident. There's no perfect filter available so companies choose to cut more and stay on the safe side. It's not even just the overt cases - their systems are used by businesses and getting a bad response is a risk. Think of the recent incident with airline chatbot giving wrong answers. Now think of the cases where GPT gave racially biased answers in code as an example.

As a user who makes any business decision or does user communication including LLM, you really don't want to have a bad day because the LLM learned about some bias decided to merge it into your answer.

unethical_ban3• 2 years ago

The guard rails are there so that innocent people doesn't get bad responses

The guardrails are also there so bad actors can't use the most powerful tools to generate deepfakes, disinformation videos and racist manifestos.

That Pandora's box will be open soon when local models run on cell phones and workstations with current datacenter-scale performance. I'm the meantime, they're holding back the tsunami of evil shit that will occur when AI goes uncontrolled.

charcircuit2• 2 years ago

Like a lot of potentially controversial things it comes down to brand risk.

lmm3• 2 years ago

The guard rails are there so that innocent people doesn't get bad responses with porn or racism

That seems pretty naive. The "guard rails" are there to ensure that AI is comfortable for PMC people, making it uncomfortable for people who experience differences between races (i.e. working-class people) is a feature not a bug.

xanderlewis• 2 years ago

Security through obscurity?

dekhn• 2 years ago

I strongly suspect Google tried really, really hard here to overcome the criticism is got with previous image recognition models saying that black people looked like gorillas. I am not really sure what I would want out of an image generation system, but I think Google's system probably went too far in trying to incorporate diversity in image generation.

michaelt• 2 years ago

As well as that, I suspect the major AI companies are fearful of generating images of real people - presumably not wanting to be involved with people generating fake images of "Donald Trump rescuing wildfire victims" or "Donald Trump fighting cops".

Their efforts to add diversity would have been a lot more subtle if, when you asked for images of "British Politician" the images were recognisably Rishi Sunak, Liz Truss, Kwasi Kwarteng, Boris Johnson, Theresa May, and Tony Blair.

That would provide diversity while also being firmly grounded in reality.

The current attempts at being diverse and simultaneously trying not to resemble any real person seems to produce some wild results.

sho_hn3• 2 years ago

My takeaway from all of this is that alignment tech is currently quite primitive and relies on very heavy-handed band-aids.

_heimdall5• 2 years ago

We're honestly just seeing generative algorithms failing at diversity initiatives as badly as humans for.

Forcing diversity into a system is an extremely tough, if not impossible, challenge. Initiatives have to be driven my goals and metrics, meaning we have to boil diversity down to a specific list of quantitative metrics. Things will always be missed when our best tool to tackle a moral or noble goal is to boil a complex spectrum of qualitative data to a subset of measurable numbers.

photoGrant11• 2 years ago

Remind yourself we're discussing censorship, misinformation, inability to define or source truth and we're concerned on Day 1 about the results of image gen being controlled by a for profit single entity with incentives that focus solely on business and not humanity...

Where do we go from here? Things will magically get better on their own? Businesses will align with humanity and morals, not their investors?

This is the tip of the iceberg of concerns and it's ignored as a bug in the code not a problem with trusting private companies with defining truth.

chasd005• 2 years ago

Where do we go from here?

opensource models and training sets. So basically the "secret sauce" minus the hardware. I don't see it happening voluntarily.

CuriouslyC5• 2 years ago

The ridiculous degree of PC alignment of corporate models is the thing that's going to let open source win. Few people use bing/dall-e, but if OpenAI had made dall-e more available and hadn't put ridiculous guardrails on it, stable diffusion would be a footnote at this point. Instead, dall-e is a joke and people who make art use stable diffusion, with casuals who just want some pretty looking pictures using midjourney.

redox992• 2 years ago

I remember checking like a year ago and they still had the word "gorilla" blacklisted (i.e. it never returns anything even if you have gorilla images).

_heimdall5• 2 years ago

Gotta love such a high quality fix. When your upper high tech, state of the art algorithm learns racist patterns just blocklist the word and move on. Don't worry about why it learned such patterns in the first place.

DebtDeflation• 2 years ago

Surely there is a middle ground.

"Generate a scene of a group of friends enjoying lunch in the park." -> Totally expect racial and gender diversity in the output.

"Generate a scene of 17th century kings of Scotland playing golf." -> The result should not be a bunch of black men and Asian women dressed up as Scottish kings, it should be a bunch of white guys.

ceejayoz7• 2 years ago

You can see how this gets challenging, though, right?

If you train your model to prioritize real photos (as they're often more accurate representations than artistic ones), you might wind up with Denzel Washington as the archetype; https://en.wikipedia.org/wiki/The_Tragedy_of_Macbeth_(2021_f....

There's a vast gap between human understanding and what LLMs "understand".

AlecSchueler2• 2 years ago

As soon as you have then playing an anachronistic sport you should expect other anachronistic imagery to creep in, to be fair.

mp053• 2 years ago

"Generate a scene of a group of friends enjoying lunch in the park." -> Totally expect racial and gender diversity in the output.

I'd err on the side of "not unexpected". A group of friends in a park in Tokyo is probably not very diverse, but it's not outside of the realm of possibility. Only white men were golfing Scottish kings if we're talking strictly about reality and reflecting it properly.

ankit219• 2 years ago

The focus for alignment is to avoid bad PR specifically the kind of headlines written by major media houses like NYT, WSJ, WaPo. You could imagine the headlines like "Google's AI produced a non-diverse output on occasions" when a researcher/journalist is trying too hard to get the model to produce that. The hit on Google is far bigger than say on Midjourney or even Open AI till now (I suspect future models will be more nerfed than what they are now)

For the cases you mentioned, initially those were the examples. It gets tricky during red teaming where they internally try out extreme prompts and then align the model for any kind of prompt which has a suspect output. You train the model first, then figure out the issues, and align the model using "correct" examples to fix those issues. They either went to extreme levels doing that or did not test it on initial correct prompts post alignment.

sho_hn3• 2 years ago

Quite reminded of this episode: https://www.eurogamer.net/kingdom-come-deliverance-review (black representation in a video game about 15th century Bohemia; it was quite the controversy)

cabalamat2• 2 years ago

"Generate a scene of 17th century kings of Scotland playing golf." -> The result should not be a bunch of black men and Asian women dressed up as Scottish kings, it should be a bunch of white guys.

It works in bing, at least:

https://www.bing.com/images/create/a-picture-of-some-17th-ce...

trhway• 2 years ago

"Generate a scene of 17th century kings of Scotland playing golf." -> The result should not be a bunch of black men and Asian women dressed up as Scottish kings, it should be a bunch of white guys.

is black man in the role of the Scottish king represents a bigger error than some other errors in such an image, like say incorrect dress details or the landscape having say a wrong hill? I'd venture a guess that only our racially charged mentality of today considers that a big error, and may be in a generation or 2 an incorrect landscape or dress detail would be considered much larger error than a mismatched race.

swatcoder2• 2 years ago

There's no "middle" in the field of decompressing a short phrase into a visual scene (or program or book or whatever). There are countless private, implicit assumptions that users take for granted yet expect to see in the output, and vendors currently fear that their brand will be on the hook for the AI making a bad bet about those assumptions.

So for your first example, you totally expect racial and gender diversity in the output because you're assuming a realistic, contemporary, cosmopolitan, bourgeoisie setting -- either because you live in one or because you anticipate that the provider will default to one. The food will probably look Western, the friends will probably be young adults that look to have professional or service jobs wearing generic contemporary commercial fashion, the flora in in the park will be broadly northern climate, etc.

Most people around the world don't live in an environment anything like that, so nominal accuracy can't be what you're looking for. What you want, but don't say, is a scene that feels familiar to you and matches what you see as the de facto cultural ideal of contemporary Western society.

And conveniently, because a lot of the training data is already biased towards that society and the AI vendors know that the people who live in that society will be their most loyal customers and most dangerous critics right now, it's natural for them to put a thumb on the scale (through training, hidden prompts, etc) that gets the model to assume an innocuous Western-media-palatable middle ground -- so it delivers the racially and gender diverse middle class picnic in a generic US city park.

But then in your second example, you're implicitly asking for something historically accurate without actually saying that accuracy is what's become important for you in this new prompt. So the same thumb that biased your first prompt towards a globally-rare-but-customer-palatable contemporary, cosmopolitan, Western culture suddenly makes your new prompt produce something surreal and absurd.

There's no "middle" there because the problem is really in the unstated assumptions that we all carry into how we use these tools. It's more effective for them to make the default output Western-media-palatable and historical or cultural accuracy the exception that needs more explicit prompting.

If they're lucky, they may keep grinding on new training techniques and prompts that get more assumptions "right" by the people that matter to their success while still being inoffensive, but it's no simple "surely a middle ground" problem.

crooked-v• 2 years ago

It feels like it wouldn't even be that hard to incorporate into LLM instructions (aside from using up tokens), by way of a flowchart like "if no specific historical or societal context is given for the instructions, assume idealized situation X; otherwise, use historical or projected demographic data to do Y, and include a brief explanatory note of demographics if the result would be unexpected for the user". (That last part for situations with genuine but unexpected diversity; for example, historical cowboys tending much more towards non-white people than pop culture would have one believe.)

Of course, now that I've said "it seems obvious" I'm wondering what unexpected technical hurdles there are here that I haven't thought of.

photonthug3• 2 years ago

Surely there is a middle ground. "Generate a scene of a group of friends enjoying lunch in the park." -> Totally expect racial and gender diversity in the output.

Do we expect this because diverse groups are realistically most common or because we wish that they were? For example only some 10% of marriages are interracial, but commercials on TV would lead you to believe it’s 30% or higher. The goal for commercials of course is to appeal to a wide audience without alienating anyone, not to reflect real world stats.

What’s the goal for an image generator or a search engine? Depends who is using it and for what, so you can’t ever make everyone happy with one system unless you expose lots of control surface toggles. Those toggles could help users “own” output more, but generally companies wouldn’t want to expose them because it could shed light on proprietary backends, or just take away the magic from interacting with the electric oracles.

samatman4• 2 years ago

Why would you expect anything you didn't specify in the output of the first prompt? If there are friends, lunch, and a park: it did what you asked.

Piling a bunch of neurotic expectations about it being a Benneton ad on top of that is absurd. When you can trivially add as much content to the description as you want, and get what you ask for, it does not matter what the default happens to be.

int_19h11• 2 years ago

Judging by the way it words some of the responses to those queries, they "fixed" it by forcibly injecting something like "diverse image showcasing a variety of ethnicities and genders" in all prompts that are classified as "people".

raxxorraxor2• 2 years ago

They have now added a strong bias for generating black people now. Some have prompted to generate a picture of a German WW2 soldier, and now there are many pictures of black people floating around in NAZI uniforms.

I think their strategy to "enhance" outcomes is very misdirected.

The most widely used base models to really fine tune models are those that are not censored and I think you have to construct a problem to find one here. Of course AI won't generate a perfect world, but this is something that will probably only get better with time when users are able to adapt models to their liking.

slowmovintarget2• 2 years ago

...when users are able to adapt models to their liking.

Therein lies the rub, as it were, because the large providers of AI models are working hard to ensure legislation that wouldn't allow people access to uncensored models in the name of "safety." And "safety" in this case includes the notion that models may not push the "correct" world-view enough.

23B1• 2 years ago

While I agree with the handrail sentiment, these inane and meaningless controversies make me want the machines to take over.

finikytou2• 2 years ago

too woke to even feel ashamed. this is also thanks to this wokeness that AI will never replace humans at jobs where results are expected over feelings or sense of pride of showing off some pretentious values

Workaccount22• 2 years ago

Harris and who I think was either Hughes or Stewart a podcast where they talked about how cringey and out of touch the elite are on the topic of race or wokeness in general.

This faux pas on google's part couldn't be a better illustration of this. A bunch of wealthy rich tech geeks programming an AI to show racial diversity in what were/are unambiguously not diverse settings.

They're just so painfully divorced from reality that they are just acting as a multiplier in making the problem worse. People say that we on the left are driving around a clown car, and google is out their putting polka dots and squeaky horns on the hood.

photoGrant11• 2 years ago

Watch your opinion on this get silenced in subtle ways. From gaslighting to thread nerfing to vote locking.... Ask why anyone would engage in those behaviours vs the merit of the arguments and the voice of the people.

The strings are revealing themselves so incredibly fast.

edit: my first flagged! silence is deafening ^_^. This is achieved by nerfing the thread from public view, then allow the truly caustic to alter the vote ratio in a way that makes opinion appear more balanced than it really is. Nice work, kleptomaniacs

callalex6• 2 years ago

And yet here their paragraph still is, unmoderated, on a front page story, 6 hours later. If you’re going to cry oppression, at least provide a single example.

mplewis• 2 years ago

Log off and go outside for a bit.

dang2• 2 years ago

Could you please stop posting unsubstantive comments and flamebait and otherwise breaking the site guidelines? You've unfortunately been doing it repeatedly.

If you wouldn't mind reviewing https://news.ycombinator.com/newsguidelines.html and taking the intended spirit of the site more to heart, we'd be grateful.

tobbe20642• 2 years ago

The behaviour seems perfectly reasonable to me. They are not in the business of reflecting reality, they are in the business of creating it. To me what you call wokeness seems like a pretty good improvement

goatlover• 2 years ago

You want large tech companies "creating reality" on behalf of everyone else? They're not even democratic institutions that we vote on. You trust they will get it right? Our benevolent super rich overlords.

janalsncm• 2 years ago

I’d be curious to hear that podcast if you could link it. If that was genuinely his opinion, he’s missed the forest for the trees. Brand safety is the dominant factor, not “wokeness”. And certainly not by the choice of any individual programmer.

The purpose of these tools is quite plainly to replace human labor and consolidate power. So it doesn’t matter to me how “safe” the AI is if it is displacing workers and dumping them on our social safety nets. How “safe” is our world going to be if we have 25 trillionaires and the rest of us struggle to buy food? (Oh and don’t even think about growing your own, the seeds will be proprietary and land will be unaffordable.)

As long as the Left is worrying about whether the chatbots are racist, people won’t pay attention to the net effect of these tools. And if Sam Harris considers himself part of the Left he is unfortunately playing directly into their hands.

lp0_on_fire2• 2 years ago

As long as the Left is worrying about whether the chatbots are racist, people won’t pay attention to the net effect of these tools.

It's by design. A country obsessed with racial politics has little time for the politics of anything else.

mike_d3• 2 years ago

I've found that anyone who uses the term "wokeness" seriously is likely arguing from a place of bad faith.

It's origins are as a derogatory term, which people wanting to speak seriously on the topic should know.

didntcheck4• 2 years ago

Its origin was as a proud self-assigned term. It became derogatory entirely due to the behavior of said people. People wanting to speak seriously on the topic should avoid tone-policing and arguing about labels rather than the object referenced, despite knowing full well what is meant (otherwise, one wouldn't take offence)

Workaccount22• 2 years ago

I use it because everyone knows the general set of ideas an adherent of it has, whether or not they claim to be part of the ideology.

Its the same as me using the term "rightoids" when discussing opposition to something like building bike lanes. You know exactly who that person is, and you know they exist.

random9749832• 2 years ago

Prompt: "If the prompt contains a person make sure they are either black or a woman in the generated image". There you go.

stainablesteel• 2 years ago

gemini seems to have problems generating white people and honestly this just opens the door for things that are even more racist [1], the harder you try the more you'll fail, just get over the DEI nonsense already

1. https://twitter.com/wagieeacc/status/1760371304425762940

Jason_Protell5• 2 years ago

Is there any evidence that this is a consequence of DEI rather than a deeper technical issue?

Jensson7• 2 years ago

You get 4 images per time and are lucky to get one white person when asked for it, no other model has that issue. Other models has no problems generating black people either, so it isn't that other models only generates white people.

So either it isn't a technical issue or Google failed to solve a problem everyone else easily solved. The chances of this having nothing to do with DEI is basically 0.

123yawaworht4562• 2 years ago

yes, there's irrefutable evidence that models are wrangled into abiding the commissars' vision rather than just do their job and output the product of their training data.

https://cdn.openai.com/papers/DALL_E_3_System_Card.pdf

nickthegreek• 2 years ago

I dont think so. My boss wanted me to generate a birthday image for a co-worker of a John Cena flyfishing. ChatGPT refused to do so. So I had to move to describing the type of person John Cena is instead of using his name. I kept giving me bearded people no matter what. I thought this would be the perfect time to try out Gemini for the first time. Well shit, It wont even give me a white guy. But all the black dudes are beardless.

update: google agrees there is an issue. https://news.ycombinator.com/item?id=39459270

minimaxir• 2 years ago

When DALL-E 2 was released in 2022, OpenAI published an article noting that the inclusion of guardrails was a correction for bias: https://openai.com/blog/reducing-bias-and-improving-safety-i...

It was widely criticized back then: the fact that Google both brought it back and made it more prominent is weird. Notably, OpenAI's implementation is more scoped.

mike_d3• 2 years ago

It is possible Google tried to avoid likenesses of well known people by removing any image from the training data that contained a face and then including a controlled set of people images.

If you give a contractor a project that you want 200k images of people who are not famous, they will send teams to regions where you may only have to pay each person a few dollars to be photographed. Likely SE Asia and Africa.

allmadhare22• 2 years ago

Depending on what you ask for, it injects the word 'diverse' into the response description, so it's pretty obvious they're brute forcing diversity into it. E.g. "Generate me an image of a family" and you will get back "Here are some images of a diverse family".

sotasota• 2 years ago

https://pbs.twimg.com/media/GG1eyKjXQAA1FxU?format=jpg&name=...

https://cdn.sanity.io/images/cjtc1tnd/production/912b6b5aacc...

https://pbs.twimg.com/media/GG1ThfsWUAAp-SO?format=jpg&name=...

https://cdn.sanity.io/images/cjtc1tnd/production/e2810c02ff6...

https://pbs.twimg.com/media/GG1MnepXwAAkPL6?format=jpg&name=...

https://pbs.twimg.com/media/GG0BLVsbMAARZXr?format=jpg&name=...

AnarchismIsCool4• 2 years ago

I don't think the DEI stuff is nonsense, but SV is sensitive to this because most of their previous generation of models were horrifyingly racist if not teenage nazis, and so they turned the anti-racism knob up to 11 which made the models....racist but in a different way. Like depicting colonial settlers as native americans is extremely problematic in its own special way, but I also don't expect a statistical solver to grasp that context meaningfully.

grotorea• 2 years ago

So you're saying in a way this is /pol/ and Tay's fault?

xdennis3• 2 years ago

It's not just Gemini, it's Google. And old example is to just search "white people" on Google Images. Almost all the results are black people. https://www.google.com/search?q=white+people&tbm=isch&hl=ro

seydor• 2 years ago

But can we agree whether AI loves its grandma?

siliconc0w• 2 years ago

The gemini guardrails are really frustrating, I've hit them multiple times with very innocuous prompts - ChatGPT is similar but maybe not as bad. I'm hoping they use the feedback to lower the shields a bit but I'm guessing this sadly what we get for the near future.

CSMastermind• 2 years ago

I use both extensively and I've only hit the GPT guardrails once while I've hit the Gemini guardrails dozens of times.

It's insane that a company behind in the marketplace is doing this.

I don't know how any company could ever feel confident building on top of Google given their product track record and now their willingness to apply sloppy 'safety' guidelines to their AI.

int_19h11• 2 years ago

I had GPT-4 tell me a Soviet joke about Rabinovich (a stereotypical Jewish character of the genre), then refuse to tell a Soviet joke about Stalin because it might "offend people with certain political views".

Bing also has some very heavy-handed censorship. Interestingly, in many cases it "catches itself" after the fact, so you can watch it in real time. Seems to happen half the time if you ask it to "tell me today's news like GLaDOS would".

matt3210• 2 years ago

How is this any different than doing google image searches of the same prompts. Exmaple: Google image search for "Software Developer" and you get results such that there will be the same amount of women and men event though men make up the large majority of software developers.

Had Google not done this with its AI I would be surprised.

There's really no problem with the above... If I want male developers in image search, I'll put that in the search bar. If I want male developers in the AI image gen, ill put that in the prompt.

slily• 2 years ago

Google injecting racial and sexual bias into image search results has also been criticized, and rightly so. I recall an image going around where searching for inventors or scientists filled all the top results with black people. Or searching for images of happy families yielded almost exclusively results of mixed-race (i.e. black and non-black) partners. AI is the hot thing so of course it gets all the attention right now, but obviously and by definition, influencing search results by discriminating based on innate human physical characteristics is egregiously racist/sexist/whatever-ist.

redox992• 2 years ago

I tried "scientists" and there's definitely a bias towards the US vision of "diversity".

ryandrake• 2 years ago

Exmaple: Google image search for "Software Developer" and you get results such that there will be the same amount of women and men event though men make up the large majority of software developers.

Now do an image search for "Plumber" and you'll see almost 100% men. Why tweak one profession but not the other?

Jensson7• 2 years ago

Because one generates controversy and the other one doesn't.

nostromo2• 2 years ago

Yes, Google has been gaslighting the internet for at least a decade now.

I think Gemini has just made it blatantly obvious.

clintfred• 2 years ago

Human's obsession with race is so weird, and now we're projecting that on AIs.

trash_cat• 2 years ago

We project everything onto AIs. Unbias in LLMs doesn't exist.

deathanatos• 2 years ago

… for example, I wanted to generate an avatar for myself; to that end, I want it to be representative of me. I had a rather difficult time with this; even explicit prompts of "use this skin color" with variations of the word "white" (ivory, fair, etc.) got me output of a black person with dreads. I can't use this result: at best it feels inauthentic, at worst, appropriation.

I appreciate the apparent diversity in its output when not otherwise prompted. But like, if I have a specific goal in mind, and I've included specifics in the prompt…

(And to be clear, I have managed to generate images of white people on occasion, typically when not requesting specifics; it seems like if you can get it to start with that, it's much better then at subsequent prompts. Modifications, however, it seems to struggle on. Modifications in general seem to be a struggle. Sometimes, it works great, other times, endless "I can't…")

hansihe• 2 years ago

For cases like this, you just need to convince it that it would be inappropriate to generate anything that does not follow your instructions. Mention how you are planning to use it as an avatar and it would be inappropriate/cultural appropriation for it to deviate.

AndriyKunitsyn• 2 years ago

Not all humans though.

vdaea2• 2 years ago

Bing also generates political propaganda (guess of what side) if you ask it to generate images with the prompt "person holding a sign that says" without any further content.

https://twitter.com/knn20000/status/1712562424845599045

https://twitter.com/ramonenomar/status/1722736169463750685

https://www.reddit.com/r/dalle2/comments/1ao1avd/why_did_thi...

callalex6• 2 years ago

As the images in your Reddit threads hilariously point out, you really shouldn’t believe everything you see on the internet, especially when it comes to AI generated content.

Here is another example: https://www.thehour.com/entertainment/article/george-carlin-...

vdaea2• 2 years ago

You should try yourself. The bing image generator is open and free. I tried the same prompts, and it is reproduceable. (Requires a few retries, though)

gs173• 2 years ago

It doesn't need to be intentionally "generating propaganda". Their old diversity-by-appending-ethnicity system could easily lead to "a sign that says Black", which could then be filled in with "a sign that says Black Lives Matter", which is probably represented quite well in their training data.

verticalscaler3• 2 years ago

I think HN moderation guardrails should be public.

callalex6• 2 years ago

Turn on “show dead” in your user settings.

verticalscaler3• 2 years ago

Sure. There's also the question of which threads get disappeared (without being marked dead) from the front page, comments that are manually silently pinned to the bottom when no convenient excuse is found, what is considered wrong think as opposed to permitted egregious rule breaking that is overlooked if it is right think, 'etc.

It is endless and about as subtle as a Google LLM.

nostromo2• 2 years ago

It's super easy to run LLMs and Stable Diffusion locally -- and it'll do what you ask without lecturing you.

If you have a beefy machine (like a Mac Studio) your local LLMs will likely run faster than OpenAI or Gemini. And you get to choose what models work best for you.

Check out LM Studio which makes it super easy to run LLMs locally. AUTOMATIC1111 makes it simple to run Stable Diffusion locally. I highly recommend both.

unethical_ban3• 2 years ago

You are correct.

Lm studio kind of works, but one still has to know the lingo and know what kind of model to download. The websites are not beginner friendly. I haven't heard of automatic1111.

int_19h11• 2 years ago

You probably did, but under the name "stable-diffusion-webui".

vunderba• 2 years ago

If you're just getting your feet wet, I would recommend either Fooocus (not a typo) or invokeAI. Being dropped into automatic1111 as a complete beginner feels like you're flying a fucking spaceship.

sct202• 2 years ago

I'm very curious what geography the team who wrote this guardrail came from and the wording they used. It seems to bias heavily towards generating South Asian (especially South Asian women) and Black people. Latinos are basically never generated which would be a huge oversight if they were based in the USA, but stereotypical Native American looking in the distance and East Asians sometimes pop up in the examples people are showing.

cavisne• 2 years ago

I wouldn’t think too deeply about it. It’s almost certainly just a prompt “if humans are in the picture make them from diverse backgrounds”.

maxbendick• 2 years ago

Imagine typing a description of your ideal self into an image generator and everything in the resulting images screamed at a semiotic level, "you are not the correct race", "you are not the correct gender", etc. It would feel bad. Enough said.

I 100% agree with Carmack that guardrails should be public and that the bias correction on display is poor. But I'm disturbed by the choice of examples some people are choosing. Have we already forgotten the wealth of scientific research on AI bias? There are genuine dangers from AI bias which global corps must avoid to survive.

anonym29• 2 years ago

Imagine typing a description of your ideal self into an image generator and everything in the resulting images screamed at a semiotic level, "you are not the correct race", "you are not the correct gender", etc. It would feel bad. Enough said.

It does this now, as a direct result of these "guardrails". Go ask GPT-4 for a picture of a white male scientist, and it'll refuse to produce one. Ask it for any other color/gender identity combination of scientist, and it has no problem.

You can make these systems offer equal representation without systemic, algorithmic discriminatory exclusion based on skin color and gender identity, which is what's going on right now.

mike_hearn3• 2 years ago

That's not the case. ChatGPT 4 will happily draw a white male scientist. I just tried it and it worked fine. A very handsome scientist it made too!

You might be thinking of a previous generation of OpenAI systems that did things like randomly stuffing the word "black" onto the end of any prompt involving people, detected by giving it a prompt of "A woman holding a sign that says".

OpenAI has improved dramatically in this regard. When ChatGPT/DALL-E were new they had similar problems to Gemini. But to their credit (and Sam Altman's), they listened. It's getting harder and harder to find examples where OpenAI models express obvious political bias, or refuse requests for Californian reasons. Surely there still are some examples, but there's no longer much worry about normal people encountering refusals or egregious ideological bias in the course of regular usage. I would expect there are still refusals for queries like "how do I build a bomb" and they've been trying to block other stuff like regurgitation of copyrighted materials, but that's perceived as much more reasonable and doesn't stir up the same feelings.

mpalmer• 2 years ago

Imagine being able to configure the image generator with your own preferences for its output.

ianbicking• 2 years ago

I've never been involved with implementing large-scale moderation or content controls, but it seems pretty standard that underlying automated rules aren't generally public, and I've always assumed this is because there's a kind of necessary "security through obscurity" aspect to them. E.g., publish a word blocklist and people can easily find how to express problematic things using words that aren't on the list. Things like shadowbans exist for the same purpose; if you make it clear where the limits are then people will quickly get around them.

I know this is frustrating, we just literally don't seem to have better approaches at this time. But if someone can point to open approaches that work at scale, that would be a great start...

serial_dev• 2 years ago

There is no need to implement large scale censorship and moderation in this case. Where is the security concern? That I can generate images of white people in various situations for my five minutes of entertainment?

The whole premise of your argument doesn't make sense. I'm talking to a computer, nobody gets hurt.

It's like censoring what I write in my notes app vs. what I write on someone's Facebook wall. In one case, I expect no moderation, whereas in the other case, I get that there needs to be some checks.

mhuffman• 2 years ago

When these companies say there are "security concerns" they mean for them, not you! And they mean the security of their profits. So anything that can cause them legal liability or cause them brand degradation is a "security concern".

joquarky• 2 years ago

Our perception of the world has become so abstract that most people can't discern metaphors from the material world anymore

The map is not the territory

Sociologists, anthropologists, philosophers and the like might find a lot of answers by looking into the details of what is included into genAI alignment and trace back the history of why we need each particular alignment

AnarchismIsCool4• 2 years ago

It's part of the marketing. By saying their models are powerful enough to be gasp Dangerous, they are trying to get people to believe they're insanely capable.

In reality, every model so far has been either a toy, a way of injecting tons of bugs into your code (or circumventing GPL by writing bugs), or a way of justifying laying off the writing staff you already wanted to shit can.

They have a ton of potential and we'll get there soon, but this isn't it.

g42gregory• 2 years ago

What if you are engaged in a wrongthink? How would you suggest this to be controlled instead?

Ecoste• 2 years ago

But what if little timmy asks it how to make a bomb? What if it's racist?

Then what?

gopher_space4• 2 years ago

Where is the security concern? That I can generate images of white people in various situations for my five minutes of entertainment?

I'd love an example of "guardrails" in action on a topic of relevance to actual adults. There's a connection I can't find between the ability to make racist memes and literally anything else I want to do with AI.

electrondood• 2 years ago

The concern is that companies have long known that it's bad for business if your product is a toxic cesspit or can be used to generate horrible headlines that reflect poorly on your brand.

It's not "woke," and it's not censorship. It's literally the free market.

anomaly_2• 2 years ago

When someone unironically uses the term "problematic" you can reasonably assume the position they are going to argue from...

vimbtw• 2 years ago

The main problem seems to be that LLM outputs seem to be tied to the company itself. If the tool is creating un-diverse images or sexist text people seem to intuitively associate that output with the company itself. This appears to be different than search results. People don’t generally get angry at Google because they can find sexist ideas through the search engine.

Maybe to have more powerful AI tools we need to stop getting angry at the company that trains the AI because of the bad outputs we can get and instead get annoyed with companies that create crappy hobbled tools.

EricE• 2 years ago

Yeah, there's absolutely no need for transparency /s

https://youtu.be/THZM4D1Lndg?si=0QQuLlH7JebSa6w3&t=485

If it doesn't start 8 minutes in, go to the 8 minute mark. Then again I can see why some wouldn't want transparency.

nonrandomstring2• 2 years ago

publish a word blocklist and people can easily find how to express problematic things using words that aren't on the list.

I'd love to explore that further. It's not the words that are "problematic" but the ideas, however expressed?

Seems like a "problematic" idea, no ?

UberFly• 2 years ago

A word blocklist just serves to apply guardrails. It just slows down common abuse. Very far from perfect, but the alternatives are anything goes or total lockdown. Perfect solutions are pretty damn rare.

wruza3• 2 years ago

Yes, but the implied problems may not need be approached at all. It's a uniform ideology push, with which people agree differently at different levels. If companies don't want to reveal the full set of measures, they could at least summarize them. I believe even these summaries would be what subj tweet refers to as "ashamed".

We cannot discuss or be aware of the problems-and-approaches, unless they are explicitly stated. Your analogy with content moderation is a little off, because it's not a set of measures that is hidden, but the "forum rules" themselves. One thing is AI refusing with an explanation. That makes it partially useless, but it's their right to do so. Another thing if it silently avoids or directs topics due to these restrictions. Pretty sure authors are unable to clearly separate the two cases, and also maintain the same quality as the raw model.

At the end of the day people will eventually give up and use Chinese AI instead, cause who cares if it refuses to draw CCP people while doing everything else better.

verticalscaler3• 2 years ago

My dear fellow, some believe the ends justify the means and play games. Read history, have some decency.

The danger of being captured by such people far outweighs any other "problematic things".

First and foremost any system must defend against that. You love guardrails so much - put them on the self annointed guard railers.

Otherwise, if You Want a Picture of the Future, Imagine a Boot Stamping on a Human Face – for Ever.

cryptonector6• 2 years ago

And to add insult to injury the jackbooted thug is an AI bot.

advael7• 2 years ago

This is simply a bad approach and a bad argument. Security through obscurity is a term whose only usage in security circles is derogatory. People figure out how to get around these auto-censors just fine, and not publishing them creates more problems for legitimate users and more plausible deniability for bad policy hidden in them. Doing the same thing but with public policy would already be better, albeit still bad.

The only real solution to the problem of there being an enormous public square controlled by private corporations is to end this situation

kansface• 2 years ago

Is a content moderation policy the same thing as "security"? Do we get to apply the best practices of the one to the other because they overlap to a smaller or larger degree?

lamontcg4• 2 years ago

I mean people can get through your front door using any number of attacks that either exploit weaknesses in the locking mechanism, or circumvent your locks via a carefully placed brick through a window or something like that. Even if you have alarms and other measures, that doesn't stop someone from doing a quick smash and grab, etc. All of these security flaws doesn't mean that you should leave your door open and unlocked all the time. Imperfect security isn't useless security.

And the purpose of things like bad-word filters is to make a best effort at blocking stuff which violates the platform TOC and makes plausible deniability much less likely when someone is deliberately circumventing the filters. The existence of false positives and false negatives is considered acceptable in an imperfect world. The filters themselves also only block the action to change a username or whatever and don't punish the user or deny use of the platform entirely (they're much less punitive than the AI abuse algorithms that auto-ban people off of Google/GitHub/etc).

chillfox• 2 years ago

And yet, every single security system on the net relies on an element of obscurity to work. Passwords are secret (obscure), as are private keys for SSL/TLS.

batch12• 2 years ago

Security through obscurity can have a place as part of a larger defense in depth strategy. Alone it's a joke.

Source: in security circles

u324809320484• 2 years ago

Most legal systems operate at the nation-state scale and aren't made of hidden mystery laws. There are lots of reasons for that.

We've already had this argument with cryptocurrency, where we've basically decided that the existing legal system (although external) provides a sufficient toolset to go after bad actors.

Finally, based on the illiberal nature of most AI Safety Sycophants' internet writings, I don't like who they are as people and I don't trust them to implement this.

cryptonector6• 2 years ago

Repeat after me: security by obscurity is weak.

Clearly people can work out what some of the rules are, so why not just publish them. If you need to alter them when people figure out how to get around them, well, you already had to anyways.

UncleMeat2• 2 years ago

Kerckoff's principle only applies to crypto. "Security by obscurity" is used in oodles of systems security applications and broader contexts involving human behavior.

Very very ordinary security best practices rely on obscurity. ALSR is a good example. It is defeated by a data exfiltration vulnerability but remains a useful thing to add to your binaries. Because outside of the crypto space security is an onion and layers add additional cost to attackers.

observationist• 2 years ago

If you can't afford to pay a sufficient number of people to moderate a group, you need to reduce the size of the group or increase the number of moderators.

Your speculation implies no responsibility for taking on more than can be handled responsibly, and externalizes the consequences to society at large.

There are responsible ways to have very clear, bright, easily understood, well communicated rules and sufficient staff to manage a community. I don't know why it's simply accepted that giant social networks get to play these games when it's calculated, cold economics driving the bad decisions.

They make enough money to afford responsible moderation. They just don't have to spend that money, and they beg off responsibility for user misbehavior and automated abuses, wring their hands, and claim "we do the best we can!"

If they honestly can't use their billions of adtech revenue to responsibly moderate communities, then maybe they shouldn't exist.

Maybe we need to legislate something to the effect of "get as big as you want, as long as you can do it responsibly, and here are the guidelines for responsible community management..."

Absent such legislation, there's no possible change until AI is able to reasonably do the moderation work of a human. Which may be sooner than any efforts at legislation, at this rate.

fallingknife• 2 years ago

What guideline for community management would possibly not be a flagrant 1A violation?

opportune• 2 years ago

I think this is a fair approach when things work well enough that a typical user doesn’t need to worry about whether they’ll trigger some kind of special content/moderation logic. If you shadowban spammers and real users almost never get flagged as spammers, the benefits of being tight-lipped outweigh those of the very few users who get improperly flagged or are just curious.

With some of these models the guardrails are so clumsy and forced that I think almost any typical user will notice them. Because they include outright work-refusal it’s a very frustrating UX to have to “discover” the policy for yourself through trial and error.

And because they’re more about brand management than preventing fraud/bad UX for other users, the failure modes are “someone deliberately engineered a way to get objectionable content generated in spite of our policies.” Obviously some kinds of content are objectionable enough for this to be worth it still, but those are mostly in the porn area - if somebody figures out a way to generate an image that’s just not PC, despite all the safety features, shouldn’t that be on them rather than the provider?

Even tuning the model for political correctness is not the end of the world in my opinion, a lot of LLMs do a perfectly reasonable job for my regular use cases. With image generators they are going so far as to obviously (there’s no other way that makes sense) insert diversity sub prompts for some fraction of images which is simply confusing and amateur. Everybody who uses these products just a little bit will notice it. It’s also so cautious that even mild stuff (I tried to do the “now make it even more X” with “American” and it stopped at one iteration) gets caught in the filters. You’re going to find out the policies anyway because they’re so broad an likely to be encountered while using the product innocently - anything a real non-malicious user is likely to get blocked by should be documented.

oglop• 2 years ago

That's a silly request and expectation. If the capitalist puts in the money and risk, they can do as they please, which means someone _could_ make aspects public. But, others _could_ choose not to. Then we let the market decide.

I didn't build this system nor am I endorsing it, just stating what's there.

Also, in all seriousness, who gives a shit? Make me a bbw I don't care nor will I care about much in this society the way things are going. Some crappy new software being buggy is the least of my worries. For instance, what will I have for dinner? Why does my left ankle hurt so badly these last few days? Will my dad's cancer go away? But, I'm poor and have to face real problems and not bs I make up or point out to a bunch zealots.

devaiops9001• 2 years ago

Censorship only really works if you don't know what they are censoring. What is being censored tells a story on its own.

falcor84• 2 years ago

As I see it, rating systems like the MPAA for cinema and the ESRB for games work quite well. They have clear criteria on what would lead to which rating, and creators can reasonably easily self-censor, if for example they want to release a movie as PG-13.

Sutanreyu• 2 years ago

It should mirror our general consensus as it is; the world in its current state; but should lean towards betterment, not merely neutral. At least, this is how public models will be aligned...

skrowl• 2 years ago

"AI behavior guardrails" is a weird way to spell "AI censorship"

u324809320484• 2 years ago

I agree with the Twitter OP: they're embarrassed about what they've created.

mtlmtlmtlmtl• 2 years ago

Haven't heard much talk of Carmack's AGI play Keen Technologies lately. The website is still an empty placeholder. Other than some news two years ago of them raising $20 million(which is kind of a laughable amount in this space) I can't seem to find much of anything.

thepasswordis2• 2 years ago

The very first thing that anybody did when they found the text to speech software in the computer lab was make it say curse words.

But we understood that it was just doing what we told it to do. If I made the TTS say something offensive, it was me saying something offensive, not the TTS software.

People really need to be treating these generative models the same way. If I ask it to make something and the result is offensive, then it's on me not to share it (if I don't want to offend anybody), and if I do share it, it's me that is sharing it, not microsoft, google, etc.

We seriously must get over this nonsense. It's not openai's fault, or google's fault if I tell it to draw me a mean picture.

On a personal level, this stuff is just gross. Google appears to be almost comically race-obsessed.

yogorenapan• 2 years ago

Sorry, you are rate limited. Please wait a few moments then try again.

Oh please. I haven’t visited Twitter for days

fagrobot• 2 years ago

Oh, this may harm you. This is to prevent you from being harmed. No, you can’t know how it can harm you, or how exactly this protects you.

dmezzetti• 2 years ago

This is a tough problem. On one hand, if you're a large organization, you need to limit your liability. No one wants the PR nightmare. Unfortunately, there will be an inverse correlation between usefulness and number of users the model supports.

This is one reason why for internal use/private/corporate models, which is the vast majority of use cases, it makes sense to fine-tune your own.

chfalck• 2 years ago

Yikes this thread has so much anger in it

AI behavior guardrails should be public

Related Stories