> If you've ever used ChatGPT, you know that it has interrupts when it thinks it is talking about something unacceptable, where it gives pre-canned lines decide

> If you've ever used ChatGPT, you know that it has interrupts when it thinks it is talking about something unacceptable, where it gives pre-canned lines decided by the creators about what it should say. > This sounds like a good idea when it's done with reasonable things - you wouldn't want your AI to be racist would you? - but giving the people who run the servers for ChatGPT the ability to inject their own morals and political beliefs is a very real concern for people. Here's how you solve it: demand open source models, or at least, open source access to network weights (I think it's kind of hard to open the training itself since it requires so much compute). Demand OpenAI to actually be open. When Stable Diffusion was opened, the first thing people did was removing the morality systems to prevent NSFW - either interrupts like this, or even retrained the network to better generate human anatomy (which has advantages that goes beyond NSFW images). There is no effective control that Stability AI can impose on this technology now. As long as OpenAI products are closed behind a SaSS, ChatGPT and other models will be controlled by them.