Local project files and computation in Marx Max do not make its cloud AI service fully offline. When using cloud AI, the current request’s conversation, selected history, and tool-returned content are sent to the model service. A file remaining on your computer and its text entering a model request are separate facts.
Consider the particular task: what AI is asked to read, execute, or interpret, and how that content may be used next. Placing material in a project directory does not mean the entire project has automatically been synchronized to the cloud. Nor does local storage prove that reading and analysis transmit nothing.
What happens locally?
Source research files remain in the user’s project directory. Generated programs, manuscripts, figures, and presentations can be saved locally. Local Stata and Python environments perform calculations, and code and saved output can be inspected and revised.
When AI writes a program, investigates an error, or interprets output, it needs relevant task content. A calculation running on your computer does not mean the AI discussing it also runs locally. Saving an artifact and processing it with a model are different actions.
What may be sent to a model?
| Task | Content that may enter a model request |
|---|---|
| Discuss research with AI | The current message and selected conversation history |
| Ask AI to read a file | Text, excerpts, or other usable content returned by the reading tool |
| Check code and calculations | Code, errors, and returned computational results |
| Search local Zotero | Matching metadata; reading child notes also provides their text |
| Read a web page or paper | Source content actually obtained and returned to AI |
The scope is neither necessarily just one prompt nor necessarily the entire project folder. It depends on the request, history, and actual tool operations. Summary tables may also contain sensitive information. Small size or removal of names alone does not establish that material is safe to send.
Distinguish Planar from Zotero synchronization
Planar stores plans, tasks, and checkpoints for connected projects; Marx Max and the web interface use those cloud records. Understand sharing scope when writing plans or submitting artifacts. Viewing project progress does not automatically synchronize the entire research directory. Confidential content entered into plans or artifacts still requires attention to cloud processing.
Local Zotero access and Zotero’s own synchronization are also separate. If Marx Max adds a note under a selected item and Zotero sync is enabled, the note may synchronize under your existing settings. Zotero’s sync documentation distinguishes metadata and note synchronization from attachment file synchronization. Check the relevant settings and action.
Limit material before starting a task
Choose files needed for the task or practice with material you are allowed to process. Begin Stata practice with its bundled example data. When working with Zotero, search metadata first and then select the papers for further reading.
State the boundary in the request:
Check only variable names and execution order in analysis.do. Do not read data content, open other files, or execute the program. Explain the target and reason before expanding the scope.
This helps bound the task; it is not proof of a data-isolation guarantee. For material your institution prohibits external services from processing, confirm permission before providing it to AI. Projects needing stricter conditions should establish whether the actual service meets institutional requirements.
Is content used for training, and what does deletion remove?
Neither question can be answered from local storage or public website crawler rules. Allowing AI search to retrieve public articles does not authorize training on your research data. Website crawler settings do not replace processing terms for user material.
Training use, retention, recipients, and deletion conditions require verification against the actual service and applicable disclosures. This guide does not promise zero retention, no training use, or deletion of every remote copy when a local file is removed. Local deletion, server-side deletion, and deletion through third-party synchronization have different scopes.
If your research requires particular conditions, contact us to verify the processing arrangements before providing material. See the terms of service for separate usage conditions. Establish how the material may be processed before choosing a workflow.