This is just Meta having no idea how their own software works. You don't need to be a machine learning developer to see how current text generators work, yet Meta developers were completly blind. This has absolutely nothing to do with the training data, you could give it every fact that exists and you can still easily get it to output things that are not true. It has everything to do with the way current text generators generate their output.

Current text generators estimate the next token based on it's training data and input tokens. The newest tokens take precedence over older tokens so input data is given a higher priority for estimating the next token. This means whatever a user inputs heavily influences the output. The AI does not output facts, it outputs text it thinks the user would type in next.

There is a work around. Hidden text can be added after user input giving the AI instructions to ignore certain user input. However, if the user knows what the hidden text says they can craft input that works around the work around. If the hidden text says "only use facts" the user could give the AI false facts, and because input has higher priority over training data then the false facts given by the user becomes facts to the AI. It's like the three laws where the stories find ways to get around them and nobody knows that because nobody has ever read any of the stories.

To output only facts would require a different type of text generator that outputs text in a different way that's not based on estimating the next token from user input. Current text generators are very good at generating creative output and can't be measured by their ability to produce facts no matter what anti-AI people demand. And I bet a fact producing generator would be terrible at being creative, which of course proves that it doesn't work according to anti-AI people.

Meta took a tractor to an F1 race and was flabbergasted that it couldn't keep up because it's so good at pulling heavy things. Then all the anti-tractor people declare tractors are a failure and can never work because they can't keep up. In reality the tractor was never designed to go fast, and no amount of tweaks will ever change that. Take an F1 car to a tractor pull and you'll get a very different outcome that the anti-tractor people will ignore, and Meta developers will say this means tractors can beat F1 cars in a race and they just need to tweak it to make it happen.