Don't Use LLMs for News.
Full. Stop.
This is great research and a nice post explaining their high level results. They show very clearly that the standard Chatbot results and even the more advanced reasoning chatbots cannot be trusted to give you factual answers about the world.
“Ultimately, the only way forward is better cognition, including systems that can evaluate news sources, understand satire, and so forth. But that will require deeper forms of reasoning, better integrated into the process, and systems sharp enough to fact check to their own outputs. All of which may require a fundamental rethink.”
From this post:
This has been true might since the beginning of them large language model age since these models don't actually know the difference between true and false, fact and fiction. But this study shows that it's much, much worse than people think. That’s because for topics where there are people in the world who want to manipulate information these models are easy to manipulate so that the stories are telling us are what someone else wants us to hear.
This problem seems to be actually worse for the reasoning models than the original LLM chatbots since they're searching the internet for news stories that can be manipulated.
The only disagreement I would have with this article is that they don't state things strongly enough and that their conclusions seem to indicate that that's a way to fix it.
TL;DR
So let's be crystal clear: these results are completely disqualifying for using LLM reasoning chatbots as a news source in any way.
That's academic speak for saying their garbage and they're dangerous you should never, ever use them to find an answer about what's going on in the world.
The Way Forward
The way forward for searching for information online about anything is not, as the researchers say, “better reasoning with these AI systems”. The way forward is going backwards to searching and finding reputable resources that you the human user can verify and decide to trust or not yourself.
This approach of aggregating answers into one combined answer is inherently flawed as long as we're using systems that cannot do fundamental reasoning about cause and effect, time, fact versus fiction.
Timelines
Which brings me to the second minor thing I’d emphasize differently from the article.
There is no indication or reasonable ti
meline for having reasoning systems that can do all of these things despite what the billionaire is running or companies trying to get you to pay for the chatbots say. We are not close to having general agents that can reliably understand cause and effects, integrate all of the context of human beings do to understand that for example news stories from are propaganda, and inherently care about the difference between things being true or false. That's not how any of these systems with a large language model at its core work.
Is it possible to build such a system? Yes, my brains can do it we can make a computer do it someday. Many people are working on ways to do a little parts of that everyday, it's called artificial intelligence research.
But no amount of hacking and bolting on top additional loops of language generation, not even simple reinforcement learning loops over the same sense of words, will give us this common sense information aggregation oracle that everyone seems to want in the short term.
Back to the “Good-Old Days”?
What worked for almost 20 years was Google returning a list of links to actual websites created by actual humans so that you could decide to trust them or not.
That can still work, while using AI to aggressively investigate and filter which sites are real and which are fluff.
There isn't anything else, there isn't going to be anything else anytime soon. Just accept that, adjust your web browsing, and sell or buy stocks appropriately based on who admits this who doesn't.