How to write without research data leakage

Typeahead processes sentence completion locally, so the writing around your cursor stays on the Mac instead of being sent to a remote completion service. Typeahead processes writing locally. Its only network activity is license activation, optional update checks and the one-time download of your AI model.

Classify each passage by the rules attached to its contents, then use only a writing path whose processing location, recipients, retention and contracts satisfy those rules. For sensitive passages, Typeahead removes the remote-completion disclosure path by using a local model. Tool choice is part of research governance.

A draft can require protection because it contains identifiable participant information, controlled-access data or derivatives, confidential peer-review material, collaborator-owned information or patentable details. The pre-publication stage includes documents with very different restrictions. One passage can carry several duties, and each duty can set a different permitted route.

Route each passage by its governing rules

Classify each passage by the rules attached to its contents, then use only a writing path whose processing location, recipients, retention and contracts satisfy those rules. For sensitive passages, Typeahead removes the remote-completion disclosure path by using a local model.

Removing names does not necessarily de-identify participant material. HHS recognizes Expert Determination and Safe Harbor as HIPAA de-identification methods. Safe Harbor identifiers must be removed wherever they occur, including free text. Dates, locations, occupations, rare events and other context can make an information-rich narrative identifiable even after obvious names disappear.

Sending text to a remote AI service is therefore a data-governance decision. Where the GDPR applies, transmission and disclosure are processing, while Article 5 limits processing to specified purposes and requires data minimization. NIST recommends evaluating third-party AI systems and their contracts, including secondary data use, liability and incident responsibility.

Trace the possible route for typed passages and pasted excerpts through processing, retention and reuse. Our guide to what happens when an AI tool reads your typing provides a concrete map.

Keep a lightweight record beside the project protocol. Record the passage class, governing rule, selected route, approved recipients, retention terms and reviewer. Update it when a tool, contract or project stage changes.

Some rules decide the route for you

Project-specific restrictions control even when a vendor offers strong general privacy terms. NIH bars controlled-access genomic and associated data from public generative-AI tools because supplying that material through prompts or other interfaces violates the Non-Transferability provision of its Data Use Certification Agreement. The applicable data-use agreement decides the route before convenience enters the decision.

Confidential review material can carry a separate prohibition. NIH peer reviewers may not upload grant-application content, original concepts, proposals or critiques to online generative-AI tools. A 2026 Nature Methods editorial similarly states that reviewers may not upload manuscripts into generative-AI tools because the upload can compromise peer-review confidentiality.

Potentially patentable detail needs review before external disclosure. WIPO warns that public disclosure before filing can destroy novelty in many jurisdictions, while grace periods and the legal effect of a particular disclosure vary. Technology-transfer staff or patent counsel should assess the planned route. Researchers should not assume that a private upload is either public disclosure or legally harmless.

Use three routes. Approved cloud services fit passages whose rules, recipients, retention and contracts permit remote processing. Local processing fits passages allowed on the secured device but not approved for remote completion. No AI is the correct route when consent, an IRB, a data-use agreement, review rules, collaborator terms or patent advice prohibits both.

Verify the path in the exact app

Local sentence completion removes the remote-completion disclosure path from the drafting flow. Typeahead completes user-started sentences as you type. We built the interaction around the cursor: you begin the sentence, a suggestion appears in a supported field and Tab accepts it. Nearby text is processed by the local model.

Chat is a heavy interface for the next few words. Moving a passage into another field also creates another copy and another possible recipient. Typeahead keeps the completion interaction at the cursor. Local AI autocomplete changes the architecture of that interaction.

After license activation and the model download are complete, verify the behavior yourself:

1. Turn off Wi-Fi on the Mac before opening the app used for the draft.
2. Open a supported text field in the app you use for research material.
3. Begin an unfinished sentence and continue typing as you normally would.
4. If a suggestion appears, confirm that Wi-Fi remains off, then press Tab to accept it.

We use this test to establish where sentence completion runs. IRB approval, consent, device security, data-use terms, collaborator permissions and publication policy remain the governing controls for the project. Apply them to the content before selecting its route.

Before pasting any passage into a tool, label what governs it: consent, IRB terms, a data-use agreement, peer-review confidentiality, collaborator ownership or patent review. Route it accordingly.