Let's take a particularly ridiculous piece of news: Beatrix von Storch, a MP of the far-right German AfD party, claimed a few years ago that the sun's activity (changes) were responsible for climate change [1]. Due to the sheer ridiculousness of that claim, it was widely reported on credible news sites, so basically prime material for any AI training dataset.
A human can easily see from context and their general knowledge: this is an AfD politician, her claims are completely and utterly ridiculous, it's not the first time she has spread outright bullshit and it's widely accepted scientific fact that climate change is caused by humans, not by sun activity changes. An AI at ingestion time "knows" neither of these four facts, so how can it take that claim of knowledge and store it in its database as "untrustworthy, do not use in answers about climate change" and as "if someone asks about counterfactual claims relating to climate change, show this"?
[1] https://www.tagesschau.de/faktenfinder/weidel-klimawandel-10...
I don't see a reason why AI would need special instruction to come to a mature conclusion like you did.
Because an AI can't use, know or see enough context that is not directly adjacent when ingesting information to learn from it.
>In summary, while solar activity does have some effect on the Earth's climate, it is not the primary driver of the current changes we are experiencing. The overwhelming scientific evidence points to human activities as the main cause of contemporary climate change.
So it's possible for LLMs to figure things. Also re humans we currently have riots in the UK set off by three kids being stabbed and Russian disinfo saying it was done by a muslim asylum seeker which proved false but they are rioting against the muslims anyway. I think we maybe need AI to fact check stuff before it goes to idiots.
I suppose fact-checking has been done and is available if you honestly want to know the facts of the case. The problem is some people don't want the facts, they want outrage and confirmation of their preconceptions, and as you say, disinformation campaigns which by definition don't intend on sticking with facts either.
Quite a lot of humans are bad at that too. It's not so much that AIs are markov chains but that you really want better than average human fact checking.