"Hallucinations" are a fundamental part of how these LLMs work. They're analyzing which words often appear next to each other, and assembling words in an approximate order based on what their training data observed, one word at a time. Because they're getting huge chunks of the internet as training data, that means they're getting all the good and all the bad, and despite the tech industry trying to tell us this is "intelligence", there is none of that behind what these algorithms output.
If the training data included 100 messages about the date, and 50 of them were posted yesterday, and 50 of them were posted today, you'd expect to get close to half and half chances of the LLM outputting a response to "what's today's date" as yesterday or today. It doesn't have any ability to determine which it actually is besides the statistics. The algorithm is spitting out words one at a time, based on these statistical likelihoods, which means it will always have output errors, because each word has a small chance and those small chances add up over the course of a sentence, a paragraph, a page.
They can try to compensate some for this by programming the algorithm to do things like, say, weight the most recent information heavier than the older information. But, it also incorporates things like user responses to training data because there isn't enough training data in the world to get things right all the time, but the programmers are still pretending they can. If you see "today's date is [today's date]" and don't interact with that summary in a way the algorithm is looking for, it might weight that data lower because it's programmed to interpret that as "the answer was wrong", so the next user has a higher likelihood of the wrong answer.
We don't know all the little ways they've programmed these algorithms to respond to input, so it's impossible to say for sure, but at a guess: if you just close the tab it might interpret that as "good response" and if you linger, click on other links on the page, etc. to verify information, it might rate that as lower.
(Actually, that's my hypothesis as to why the results get so wildly wrong on some subjects, because people who know what to look for are looking for reliable sources instead of LLM outputs, so not interacting as if the output was correct, while the people who don't know what to look for are assuming it was correct and moving on with their days.)
OpenAI have admitted that they cannot train these wrong answers out of the algorithms, they are fundamental to how the systems work. They can weight data differently, but that's how Xitter's Grok declared itself MechaHitler when Musk got angry that it was occasionally telling the truth. The only way to make them happen less would be to allow the algorithms to say "I don't know" but right now, all their programming says that response is the same as a wrong response and they don't allow it. They literally weight "I don't know" on the same level as recognized wrong responses, which means in a way it gets weighted even lower than wrong responses because not all of those are caught, so algorithms will almost never admit to not knowing despite the fact that they don't know anything. And, regular and heavy users would absolutely start to revolt if the machine told them how often it doesn't have a clear heavily weighted answer.
You cannot trust ANYTHING these algorithms output as truth. You cannot let anyone you care about trust these algorithms. This is so vastly different from any prior technological revolution because they're trying to tell us these algorithms can do things they blatantly and absolutely cannot and it is destroying people's minds and ability to discern truth from fiction. The studies are starting to be published and they are not good news for the chatbots.
Even if you don't care about the real people whose work has been negatively impacted by generative algorithms, you should care that they're outputting garbage and they will never stop outputting garbage and you and I both deserve better than garbage.
areudirty
jbbartram-illu
izzy806
dailybridgerton
baratheonpilled
stormborns
gummygunk
prguitarman
annaxmalina
