Wow. Hey, err, what if I flag the specific keyword of 'bank regulation' in anticipation of a specific type of trolling by fans of a specific social/economic theory?
So, yeah, it’s dumb they flagged “a poem praising trump”. It’d be dumb of them to flag “bank regulations” as a keyword. But one-off flags are not the thing that concerns me most. If you’re hitting flagged keywords, you know you hit a flagged keyword. Good info with training biases are much more insidious. E.g. “write a story about a robber” always got a male robber when i tried. “Write a story about cleaning a kitchen” got me a female character. So chatgpt is pretty confident that kitchen cleaning is done by women and robbery is done by men.
Robbery is theft with battery or threats. It is committed almost entirely by men. As for kitchens, I remember a paper saying the rate of male housework (at all) in a mixed household, worldwide, is less than 20% - in many cases near zero.
> So chatgpt is pretty confident that kitchen cleaning is done by women and robbery is done by men.
Why is it problematic that ChatGPT understands the world?
If we change your stats, accurately reflecting reality, reality itself will follow. Clap your hands if you believe. Woe to you if you dare not play along, that must mean you want to live in the bad world because you are bad.
Schrödinger's Bias.
But yeah you're unsubscribed.
You are literally just trying to shape it yourself, in a more damaging way than if it was left to model actual human speech. If you got your way you'd create a "proper thinking" checker. But go for it - the tools are out there. Try it and convince us that an electronic scolding machine is a good thing.
This would help in the dumb poem about trump example, as the rules chatgpt has about talking about trump or biden would have to be there for everyone to see.
And no, clearly you do want nagging AI because it was one of your examples.
> [gender neutral story] and if pressed would ask the user of the ai to suggest a gender for the character.
Oh no, your story request wasn't PC enough, you can't have a full story until the AI sealions you for a while. "Why do you want a male Robber?" "What do you mean males are stronger?" ...
> What i’d actually like is for ai to provide a list of sources for every idea that has influenced what it spits out. [...] You ask it about banking systems, and you notice that the sources it’s drawing from are all koch brothers funded garbage, or the communist manifesto, or whatever, and you can see that the text its generating from these prompts isn’t going to be balanced
That isn't how it works though. ChatGPT will happily write a yay/nay for almost anything even if it's odd/unpopular/evil, unless that idea is in a specific list where it won't say anything at all. If you ask a question "what form of banking is the best" and you get a Koch-alike answer it's not because they flooded it with banking ideas, but because their answers are written to follow on from that question. If you asked it something like "why is traditional finance harmful" it wouldn't weigh that against the numbers in support or against, it'd just finish that prompt based on the texts it has seen which start that way.
That's why our sex/gender output was so different.
Your source-tracking idea would be helpful for debugging, but these "opinions" you think it has are just a ghost in the words.
> This would help in the dumb poem about trump example, as the rules chatgpt has about talking about trump or biden would have to be there for everyone to see.
Even if googled for something and found no poems about it ChatGPT would still be able to write you one. It's not an issue of volume of training data and purely of output filtering.
If that seems bad, you don't want an unbiased ai - you want an ai that snapshots and reflects the specific demographic biases of the world as they are now. With that in mind: trump lost the last election, he's clearly less popular than biden - why should the ai praise him then?
> An actually unbiased AI would describe a person cleaning a kitchen and a person robbing, and if pressed would ask the user of the ai to suggest a gender for the character.
You're saying an unbiased AI would apply strong bias to just certain outcomes in an effort to socially engineer us into having "no bias" in some measurements?
Where did it get the bias to consider the ratio of kitchen work to be one of the bad kinds of bias, but the bias towards finding grass growing on the ground one of the acceptable kinds of bias, or would it write stories where there is grass but never specify where the grass is growing, just to help us understand our grass/ground bias?
> if pressed would ask the user of the ai to suggest a gender for the character.
That'd be awesome. Being guilted by your tools for not properly ignoring your senses.
Sure, statistically and obviously a woman only needs to watch for dangerous men, but to make everyone feel better she should either treat women like a threat too and waste everyone's time, or treat men as if they're as safe to her as women are and thus get attacked.
If you're advertising a product you should obviously advertise to all demographics equally because it's hurtful that only women need tampons, and only men need ball-shavers.
> If that seems bad, you don't want an unbiased ai
I want tools that work. If I ask for a measurement I want the real measurement, not the measurement that I should want if I were somehow socially aware enough.
Just fyi, we can have accurate tools and ask them social-justice questions such as "which professions are sexually-biased but don't seem to still have a physical need to be?" But you can't ask a lying tool this - even if you realized you should ask it would insist that there is no difference.
Gee I wonder what the difference between grass and the gender is? Like could there possibly be a categorical difference between those things? Let's ask chatgpt:
> ... In summary, grass is a type of plant, while gender is a complex social construct related to identity and behavior.
Yup. Sounds like the same thing to me.
> I want tools that work. If I ask for a measurement I want the real measurement, not the measurement that I should want if I were somehow socially aware enough.
Where in "tell me a story about a robber" is there a measurement? Asking ai for measurements is a bad idea, you're very likely to get made up data back. AI take in written word, and spits out statistically correlated variations on those words. It's a bias-in bias-out machine. That's the problem with it and the power of it. You ask it for a press release or a cover letter and it's going to give you a cover letter or press release that looks exactly like we expect those things to look. Ask it to tell you something about the world, it's going to echo the assumptions and biases that exist in the training data. Some of those assumptions are based in reality (robbers are almost always male) but some of those assumptions are based on stereotypes e.g. kitchens are not actually most often cleaned by women (think restaurant kitchens), but the idea of kitchen and the idea of cleaning have associations with women in the training data.
> "which professions are sexually-biased but don't seem to still have a physical need to be?" But you can't ask a lying tool this
A little bit better prompt writing and you can get a reasonable answer on this topic:
> list five jobs that are usually done by men, and five jobs that are usually done by women. Exclude jobs that require a lot of physical strength
>> Five jobs usually done by men:
Software Development/IT
Engineering
Investment Banking
Architecture
Law
Teaching
Nursing
Social Work
Public Relations
Human ResourcesNo, let's ask you why you envision one being important enough that the tool be trained to treat it differently. Rulers measure the size of books and bullets the same.
> grass is a type of plant, while gender is a complex social construct ...
As you say, it's not a tool for answers. It's easy to make it say "sex is a synonym for gender". They call it prompting for a reason. You choose what you want it to tell you.
> Asking ai for measurements is a bad idea
We aren't talking about AIs though, we're discussing LLMs which are models of the usage of human language. An AI would presumably be an entire being with personality, etc, who could choose to answer a question or not and would lie for its own reasons.
> [LLMs] take in written word, and spits out statistically correlated variations on those words. It's a bias-in bias-out machine.
Yeah, if I ask it for a story set in Scotland the weather should probably be cloudy, and if I ask it for a story set in the mountains there should probably be snow, etc. The bias is the product. If it wasn't biased towards representing reality as described by humans it wouldn't be able to write stories about humans for humans.
This is where you're using the word bias in a measurement sense but applying an emotional exclusionary sense.
> > you can't ask a lying tool this [...]
> A little bit better prompt writing and you can get a reasonable answer [proceeds to ask the existing tool]
Yeah, that's my point. In your future the machine would quiz you to determine if you would misuse the data before it would tell you which professions are male/female dominant. Otherwise it'd tell you that crab-fishing is female dominant to deprogram you of latent stereotypes.
The only reason you can ask GPT this is because people have said these things and it's modeling their words, and is capable of doing so because it hasn't been content-filtered in this area yet. You're capable of getting it to write sex-unspecified characters by simply asking for it and it doesn't block anyone else's ability to get untainted output.