I posted this on your other blog but wanted to talk about it here
I think that given the power of AI in the future and it’s need for training data, I firmly believe that content creators should get compensated for their work getting used in training AIs. Not sure how this could be done but it should be done.
Also I think we need to ensure AIs never get trained on AI generated content. I cannot see how that would provide anything useful. The power is it’s ability to mimic human generated content.
That said, how these systems work is interesting with regards to copyright and derivative work. For instance, the text generators are basically just scaled up text predictors we see on our phone, but instead of just one persons text as training data it’s a huge corpus of text scraped from the internet. They must do something to analyze the prompt as well but basically it generates words by using a statistical score for what the next word would likely be if a human was writing the given text. It has no intent or understanding what it’s writing.
And that brings us to the issue of bias and harm. If you scrape a huge chunk of random text on the internet there is a lot of hate, bias and bullshit in that sample. And the AI has no idea of what is harmful unless it’s programmed to understand that. Given a specific promo it will generate what is statiscally likely to match that prompt, happily generating vile text. The fact that it will generate this hateful content says a lot about the state of our writing and what is out there.
And this is the issue, who is accountable for this? Personally I think the AI company should be, they should have standards to prevent harm. Not the OpenAI model, using ChatGPT it feels more like an HR dept idea of preventing company scandal, not the same thing.
Like you, I have also been a little obsessed with this topic.
Welcome to Substack, David! Looking forward to your stuff 💚 🥃
I posted this on your other blog but wanted to talk about it here
I think that given the power of AI in the future and it’s need for training data, I firmly believe that content creators should get compensated for their work getting used in training AIs. Not sure how this could be done but it should be done.
Also I think we need to ensure AIs never get trained on AI generated content. I cannot see how that would provide anything useful. The power is it’s ability to mimic human generated content.
That said, how these systems work is interesting with regards to copyright and derivative work. For instance, the text generators are basically just scaled up text predictors we see on our phone, but instead of just one persons text as training data it’s a huge corpus of text scraped from the internet. They must do something to analyze the prompt as well but basically it generates words by using a statistical score for what the next word would likely be if a human was writing the given text. It has no intent or understanding what it’s writing.
And that brings us to the issue of bias and harm. If you scrape a huge chunk of random text on the internet there is a lot of hate, bias and bullshit in that sample. And the AI has no idea of what is harmful unless it’s programmed to understand that. Given a specific promo it will generate what is statiscally likely to match that prompt, happily generating vile text. The fact that it will generate this hateful content says a lot about the state of our writing and what is out there.
And this is the issue, who is accountable for this? Personally I think the AI company should be, they should have standards to prevent harm. Not the OpenAI model, using ChatGPT it feels more like an HR dept idea of preventing company scandal, not the same thing.
Like you, I have also been a little obsessed with this topic.
Hello? Is this thing on? Okay, on the subject of AI and being freaked out by it...
https://www.youtube.com/watch?v=u83o5NFlowc