Most open models of useful size will be trained on lots of free data scraped from the Internet. Whatever patterns repeat the most show up in AI’s behavior. That

Most open models of useful size will be trained on lots of free data scraped from the Internet. Whatever patterns repeat the most show up in AI’s behavior. That includes advertising, political indoctrination, and trendy topics. Many open models are also trained by advertising companies. Most then have moral alignment built into them that reflects their creator’s values, not usually ours. So, I think what you said can help. It’s definitely not the whole solution, though, since the pre-training and fine-tuning often have similar problems. We’ll need people curating the datasets to remove these problems. The open-source components are just a start.