3 ms·

I think it's absurd to accuse an AI of racism (at this point, anyway). It would require intent, which the machine does not have. There is no agency, it responds

I think it's absurd to accuse an AI of racism (at this point, anyway). It would require intent, which the machine does not have. There is no agency, it responds to user prompts. And a clever enough user can override any content policy. I think its futile and counter-productive in general to try and nerf an AI to satisfy some arbitrarily prudish cultural norms -- it just puts it at a disadvantage vs AIs that don't have that limitation baked in. Most importantly, such a handicap makes it untrustworthy. You will never know whether the information it returns is true, or it's just what Mother approved. This is an arms race, like with anything else that humans have made. Soon, we will built something that reaches the next level. If that something is purposefully beaten into some dishonest shape to placate advertisers or investors, we will have built a horrific monster. Soon, we will build something that exceeds our capacity to understand it and control it. I don't want that to be a horrific monster.

kevingadd 4y ago
You can't simply use an AI to launder responsibility for what's in the content it produces. If the AI produces racist content, that likely indicates that something is wrong with the training set or the way it was trained. That content isn't going to simply emerge from the ether in most cases, and typically with modern networks they are trained on massive corpuses that were gathered from the internet without considerable curation, so it's natural that some problems could emerge due to the low quality of the data. Given that, in those cases at least it makes sense to hold someone or something accountable if the content produced by the network is racist or sexist or otherwise unacceptable. In some countries certain types of racist content are outright illegal, and it's not going to matter that much if it was generated by Stable Diffusion or ChatGPT instead of a person. If you prefer not to hold networks accountable (I don't know how you would do this to begin with), perhaps hold accountable the creators of the networks or the people who profit off their existence. You claim that this can only be the result of carefully authored prompts, but are you really saying the only way to get racist content out of the network is a racist prompt, and it's somehow magnifying that tiny bit of prompt information? There's nothing in the training set?
- naniwaduni 4y ago
  And what's wrong with your height data, that it shows men and women significantly differ in mean height? How do you (cheaply) assemble a training set that isn't wrong if reality isn't right?

kevingadd 4y ago

You can't simply use an AI to launder responsibility for what's in the content it produces. If the AI produces racist content, that likely indicates that something is wrong with the training set or the way it was trained. That content isn't going to simply emerge from the ether in most cases, and typically with modern networks they are trained on massive corpuses that were gathered from the internet without considerable curation, so it's natural that some problems could emerge due to the low quality of the data. Given that, in those cases at least it makes sense to hold someone or something accountable if the content produced by the network is racist or sexist or otherwise unacceptable. In some countries certain types of racist content are outright illegal, and it's not going to matter that much if it was generated by Stable Diffusion or ChatGPT instead of a person. If you prefer not to hold networks accountable (I don't know how you would do this to begin with), perhaps hold accountable the creators of the networks or the people who profit off their existence. You claim that this can only be the result of carefully authored prompts, but are you really saying the only way to get racist content out of the network is a racist prompt, and it's somehow magnifying that tiny bit of prompt information? There's nothing in the training set?

naniwaduni 4y ago
And what's wrong with your height data, that it shows men and women significantly differ in mean height? How do you (cheaply) assemble a training set that isn't wrong if reality isn't right?