Google DeepMind published a 57-page report, From AGI to ASI, and the first-page reading instruction is more interesting than the formatting gimmick.
The report explicitly tells human readers to ask an AI assistant or agent for a summary tailored to their interests and background. Immediately after that, it gives instructions for the assistant itself: what to summarize, what not to compress, and which tensions in the report should remain visible.
That is the actual news.
The paper is not only about future AI systems. It is also written with the assumption that an AI system may be the first reader, the filter, and the interface between the document and the human.
This is already how many technical people read papers in practice: drop the PDF into Claude, GPT, Gemini, or another assistant, ask for the gist, then drill into the sections that matter. The difference here is that the behavior is no longer informal. It is acknowledged directly by the authors.
My read: scientific publishing is starting to adapt to AI-mediated reading. The next useful paper format may not only be optimized for human skim-reading, citation graphs, and PDF layout, but also for structured AI ingestion.
The open question is what this does to authorship and trust. If papers begin including instructions for AI readers, those instructions become part of the document’s interface. Good versions will preserve nuance. Bad versions will become marketing prompts embedded inside research.