The manuscript you can’t paste into ChatGPT.

• Reads long PDFs. Open a 60-page review article, an unpublished manuscript under embargo, an IRB protocol, and ask questions cited back to the page. The PDF stays on your Mac.
• Summarises peer-review packages.The submission, reviewer comments, and your draft response, all summarised in one pass without any of it landing in a vendor’s log.
• Drafts grant prose. Specific Aims, Significance, Approach sections drafted from your own outline, edited as tracked changes in Word. The unfunded grant idea stays on your machine.
• Anonymises qualitative data. Strip subject identifiers from interview transcripts, clinical case reports, and ethnographic field notes before sharing with collaborators or depositing in a repository. Each redaction logged in a Word comment with the original value.
• Translates literature. Read a French, Russian, or Japanese-language source without sending it to an online translator that caches the file.
• Drafts IRB protocols and consent forms. Templated language adapted to your study design, with the actual study details never leaving the machine.

NIH’s notice NOT-OD-23-149 is unambiguous: peer reviewers may not use generative AI on grant applications or critiques, because “no guarantee exists explaining where AI tools send, save, view, or use grant application, contract proposal, or critique data.” NSF’s equivalent guidance from late 2023 is now in the 2025 PAPPG. The same logic applies to journal peer review and to any pre-publication manuscript you have agreed to keep confidential. Pasting it into ChatGPT or Claude is a disclosure to a third party.

The retraction wave is a separate hazard. Retraction Watch has catalogued ~100 papers with explicit ChatGPT-text evidence (the giveaway phrases like “As an AI language model…” left in the submitted text). 22.7% of AI-related retractions occurred in 2024 alone. Drafting on your machine means you actually see what you submitted, not what an opaque chat-history pipeline did to it.

For researchers in the EU and UK, the Schrems II transfer-impact assessment is a real piece of work to file every time identifiable subject data crosses into US-cloud AI. Keeping the data on the machine means the transfer never happens, which means the assessment never needs to be filed. ERC, Wellcome, and Horizon Europe data stewardship plans accept on-device processing without additional safeguards review.

• Reading and summarising long manuscripts and review articles.
• Peer review of submitted manuscripts under journal confidentiality.
• NIH, NSF, Wellcome, ERC grant drafting from outline.
• IRB protocol and consent form drafting.
• Anonymising interview transcripts and qualitative field notes.
• Literature review across foreign-language sources.
• Drafting cover letters, response-to-reviewer letters, retraction notices.
• Pre-print preparation under embargo.

These are useful for the discovery side of the workflow: search a literature, find papers, get a structured answer with citations. They are also cloud services that ingest the user’s uploaded PDFs into their infrastructure. Read the terms; the line between “processing for your query” and “training on your uploads” is usually not as sharp as the marketing suggests.

Muet sits at the other end of the workflow: once you have the PDFs, what you do with them stays on your machine. Use the cloud tools to find the literature; use Muet to read, draft, and write against it.

Cheaper than a single page-charge on most journals, and a fraction of a typical grant-writing software subscription. Most labs recoup it on the first round of peer review.