Quick observation from my actual attempted use of genAI (in this case, Notion's) -- It's not just the model, but the system designs, that seem to err towards more "creativity", for example putting more things in context. I wanted Notion to summarize just two documents, but it clearly used content from the parent document in its response. This poisoned the output, because the content was different.
Most use cases where I actually want to use the capabilities, I don't want expanded context.