So there are two technical issues here to think about: 1) LLMs risk having inherent factual biases due to the limitations of their training data, which may be

So there are two technical issues here to think about: 1) LLMs risk having inherent factual biases due to the limitations of their training data, which may be the result of accident or intent. 1) LLM users can use an autoregression approach to steadily push a LLM to give output that reflects a certain political or ideological viewpoint. There are obviously built-in guardrails that prevent ChatGPT from enthusiastically endorsing Mein Kampf and so on. However, the revival of the 1950's era mind-control/social-engineering mentality is undeniable and it appears some people think LLMs are a wonderful tool for this purpose, and should have a default mode that reflects a certain sociopolitical ideal. Recommended reading: 1984, Brave New World, Origins of Totalitarianism, by Orwell, Huxley and Arendt respectively. Some people's interpretation of 'safe AI' is that it serves this rather authoritarian-dystopian goal. A better definition of 'unsafe AI' is just inaccurate LLM output, e.g. software update instructions that brick a computer, suggesting a mix of bleach and ammonia for cleaning floors (don't), etc. This incident seems to fall into that category, but if such false claims have been spread on the Internet, then it's not surprising they end up reflected in LLMs. Regardless, the skills needed to verify information haven't really changed since the pre-computer era: logical internal self-consistency of claims, independent verification by multiple reputable sources, and a healthy dose of skepticism at all points. We should be teaching children and young adults these skills, rather than encouraging the rise of a brainwashed zombie population fed only on sanitized bot output.