Can an AI do research with no internet?
What a local research AI can do with the network unplugged, what it honestly cannot, and when offline is not a limitation but the point.
Yes. A research AI built on local models does its retrieval, its reasoning and its writing on your own machine, so it keeps working with the network unplugged. What it reads is the library you have already added: papers, PDFs, websites and notes. What it does not have offline is the open web, so its answers are grounded in your documents alone.
What still works offline.
Nearly the whole loop. Indexing new documents, computing the embeddings that match questions to passages, retrieving, and writing the answer with citations are all computations on your machine. None of them consult a server. If you spend an afternoon in a cabin, on a plane, or in a building whose network policy forbids uploads, the tool behaves exactly as it does at your desk, because the model lives on your GPU and the index lives on your disk.
The work this supports is real research work. Ask questions across a few hundred PDFs, pull every passage that bears on a claim, compare what two authors said about the same method, and draft sections with the citation attached to each sentence. The library is the corpus, and the corpus is local.
What the network was for.
Honesty about the loss: offline, the tool cannot discover sources you do not already have. It cannot run a web search, fetch a page, or read anything published after your last import. If a question can only be answered by the live web, a grounded tool should refuse and tell you that, because an invented answer on a plane is still an invented answer. The scope narrows to what is in the room, which is precisely what the grounding contract promises: every claim points to a passage you can open.
This is why the offline case pairs well with preparation. A researcher who syncs the reading list before the trip loses almost nothing; one who expects the tool to browse discovers the boundary immediately.
When offline is the point.
For some work, the disconnection is not a workaround but a requirement. Field sites with no reliable coverage, data governed by agreements that forbid uploads, interviews whose consent terms do not contemplate a cloud service, or simply the desire that a draft in progress never leaves the machine. In those settings a cloud notebook is not merely inconvenient; it may be disqualified before quality is even discussed. A local loop meets the constraint by construction: there is nobody to send anything to.