Researchers use AI for everything except... writing the actual paper?

Been thinking about this a lot lately

Look at how researchers actually use AI today:

• Copilot/Cursor for their code and experiments

• Elicit/Consensus/Perplexity for literature review

• ChatGPT for brainstorming and early ideas

Even Overleaf itself has started adding AI, autocomplete-style suggestions, error explanations, a citation helper in beta. But it's still fundamentally a suggestion box bolted onto the editor, not something that actually drafts with you, reviews your full paper, or rebuilds your slides when the conference changes.

Genuinely curious what this community thinks: why do you think writing stayed this far behind while every other part of research went AI-native? Is it that agentic tools for writing are just harder to build well, or is academia more cautious about AI touching the actual prose?