I just asked GPT-4 if the government lied when they claimed face masks didn’t prevent COVID-19 early in the pandemic. It evaded the question, and said that masks weren’t recommended because there were shortages. But that wasn’t the question. The question was if the government was lying to the public.
I’m going to guess a Chinese model would have a different response, and GPT-4 has been “aligned” to lie about uncomfortable facts.
Why would you ask an LLM whether a government was lying? It’s a language model, not an investigative body with subpoena power charged with determining whether someone made an intentionally false representation.
e.g. Google's Gemini's output had some rather embarrassing biases in its outputs; to the point where asking it to "draw a 1943 German soldier" resulted in images of women and black soldiers. https://www.nytimes.com/2024/02/22/technology/google-gemini-...
I wouldn't put that on the same level as "refusing to talk about massacre of civilians"; but I wouldn't put it to the level of "free and unbiased" either.
I'm not sure it's avoiding biases so much as trying to have the currently favoured bias. Obviously it got it a bit wrong with the nazi thing. It's tricky for humans too to know what you are supposed to say some times.
I did not mention historical events in my comment.
And in any case, whatever model you train is going to have the biases of the training datasets, and if you make heavy use of Wikipedia you will have the footprint of Wikipedia in your output, for good or bad.
"Similarly censored for similar topics" implies heavily that hypothetical events such as Tiananmen Square would be similarly surprised by English large language models.
