Based on my usage of Linux, Windows 10 and Windows 11, I have no doubt that Windows 11 uses at least 1-2gb more of ram. It runs like crap on older PCs while Windows 10 runs okay. That's the true test.

I've dealt with AI and know the history with hallucinations. Today Hallucinations are more of an issue when it's a controversial topic (politics, religion, or issues the political influencers have taken control of in the media).

They're fed training data pushing a narrative preferred by the AI developers, who leave out the contradictory info. It's a lot like social media moderation. Ultimately, there's a group of individuals at the company who think a certain way and think the rest of the world should think like them too. It's more of an issue after the R&D model has been turned into a commercial variant at the company by the engineers there and it's out of the hands of the Ph.Ds who originally made it.

As soon as you prompt a basically decent AI to not rely on a training data or a superficial web scrape, and it goes looking for contradictory info, then it re-thinks what it told you. This is more of an issue with the commercial corporate subscription AIs and the older small LLMs from years ago that didn't have web scraping capability. Things have definitely changed in the last year or two where the dense models with web scraping capability are concerned. At least, if you go with the uncensored or ablitterated "Thinking" versions of them available on the Huggingface site, or you use a third party service to access a large open weights model that's been made uncensored by whoever maintains it.