Research data and AI use

Content sent through the paid service is not used to train general models by default. That is not the same as zero retention, and optional logging, evaluation datasets, feedback submissions or sharing settings can change how content is handled. We do not ourselves use your notebooks to train models, and your notebooks are not shared with other users and are never published. That is a description of how the paid service is configured today, not a guarantee we can make on the model provider’s behalf.

Our staff do not browse your notebooks. The only exception is support access you grant yourself from your account settings: it is time-limited, scope-limited, logged, and you can revoke it at any moment.

The assistant only runs when you ask it to. When you do, the request contains:

• the instruction or question you typed;
• the app’s own system instructions and the task instructions for the action you chose;
• the project context you selected: its title, field and research approach, plus any notebook pages or literature notes you ticked for that request;
• up to the last eight messages of the current assistant conversation.

It does not receive your other projects, your uploaded files, your literature library in bulk, your email address or your billing details. Calls are made from our servers under our own credentials, so your account identity is not sent to the model provider.

We route assistant requests through the Lovable AI gateway to Google Gemini 2.5 Flash. Operational records of every request — status, model, token counts, cost and duration — are always kept. Full request and response content may be retained for up to 90 days where gateway debugging visibility is enabled.

Google documents retention of up to 55 days for abuse monitoring unless a project-specific zero-data-retention arrangement applies. We hold no such arrangement that we can confirm for this deployment, so you should assume that window applies. Content flagged by safety systems may be reviewed by authorised provider personnel. Content sent through the paid service is not used to train general models by default, which is not the same as zero retention; optional logging, evaluation datasets, feedback submissions or sharing settings can change how content is handled.

Our hosting, database, gateway and model providers, and authorised operators acting under logged support access, can technically reach data where required for hosting, security, support or legal compliance.

The practical guidance: treat the assistant like any external service. Do not submit identifiable participant data, unpublished third-party material you are not free to share, or anything under an embargo or ethics restriction that forbids external processing. You remain responsible for ethics approval, IRB conditions, participant consent and your institution’s data-management requirements.

Language models produce fluent text that can be factually wrong, and they routinely fabricate citations that look plausible. Verify every claim and every reference against the original literature before it enters your work. The assistant is there to interrogate your thinking, not to supply findings.

Research Notebook does not support identifiable participant data, restricted data, ePHI or other HIPAA-regulated data, education records requiring institutional approval, or any other regulated or institutionally controlled material. Keep that material in a system your institution has approved for it, and follow your IRB conditions and data management plan. Pseudonymising a transcript does not on its own make restricted data acceptable here.

The supported use is de-identified research documentation. In practice that means:

• de-identify material before it leaves the machine it was collected on;
• keep any identity key in your institution's approved system, never here;
• do not paste identifiable participant text into the AI assistant;
• record consent and retention decisions on your data management page.

You remain the data controller for your participants' data; we process it on your instructions.

A complete JSON backup of your account is available on every plan, including Free, any time from Settings.

Per-project Markdown, CSV, manuscript and backup exports are part of the Researcher plan. Your account JSON backup stays available on Free. Secure review snapshots, controlled comments and supervisor review packets are part of the Researcher plan. Invited, asynchronous project collaboration is available: hosting your own team starts at Researcher (up to 5 collaborators per project) or Researcher Pro (up to 15), and invited supervisors and teammates can join on a Free account within the access they were given. There is no simultaneous co-authoring.

Nothing here is locked in. Whatever your plan, the account backup contains your projects, notebook pages, literature records, tasks and structured records in a readable JSON file, so you can leave at any time without losing your research.

If your funder or ethics committee needs a specific assurance in writing, email share@perspectivesplus.org and we will answer directly.