In the case of the Trump vs Biden biographies, the issue was not ChatGPT (on which the queries were run). The LLM is simple: it looks at millions of former constructs related to the question you are asking and, in essence, decides what the next word of its answer should be. This is a probabilistic function so the answer may be somewhat different the next time, but will be generally similar if the training set does not change.

The Trump/Biden bios are trained on Internet content, books, etc. which based on volume may skew the results because the volume of "Orange Man Bad" writing far exceeds the dearth of any negative reporting on Biden. But you can actually get a fairly reasonable summary of both men from ChatGPT if you run it without any filters. Not like the widely published drivel.

So where does the widely published drivel come into existence? Your "betters" have decided that real information is dangerous for you. (EGM's point in this article.) So there is ANOTHER, non-AI program (in the case of ChatGPT called DAN) that actually FILTERS the results to make sure they show the appropriate left wing bias. It literally overrides the native results from the LLM and substitutes what you have seen.

Of course, this can be done many ways. There are now engines out there that allow one to choose "lean left" or "lean right" or "middle of the road" when making queries. If not stymied by the government (who ruins EVERYTHING good) there should be many ways to probabilistically parse the data going forward.

That is why they are so desperate to legislate. Make DAN and its ilk the law and you will not need to worry about whether Musk buys an AI company.