Most Google Workspace administrators have never opened the setting that governs what Gemini can read. It is on by default, and it covers Gmail, Drive, Docs, Calendar, Chat and Meet. David Gewirtz documented the configuration for ZDNET and found the switch easy to reach but harder to apply selectively than most admins would expect.
This is not a leak, and it is not a change Google announced this week. Workspace content is not used to train models and does not travel outside the domain. What the default does is move the boundary for internal data from the platform to the admin console, where somebody has to make a decision about it.
What Workspace Intelligence Sources are
Google groups the six surfaces Gemini can reach under one label: Workspace Intelligence Sources. Chat, Drive, Docs, Calendar, Meet and Gmail are each treated as a source the assistant may consult when someone asks it a question from inside one of those products.
Google is specific about the limits. Workspace content is not used for training, prompts and generated responses are not shared with other users or other organisations, and no separate AI database is built from the documents and messages. That moves the question out of regulatory territory and into internal policy.
How Gemini actually gets the data
The mechanism is search, not training. Google already indexes Workspace content, and when Gemini receives a prompt it runs real time retrieval over that index, pulls back only what the user is permitted to see, and writes an answer from it.
The permission model is the part worth understanding. Gemini inherits the access the person already has, so it will not surface a document they could not open themselves. The exposure is not new access, it is new reach. Content that was technically visible but practically buried is now one question away.
Where the governance risk sits
Gewirtz points at three cases. Client contracts or sector rules that prohibit AI scanning of company information. Departmental isolation, where an HR complaint or a live deal sitting in a Doc could surface to somebody in another team. And employees, curious or otherwise, using the assistant to reach around a boundary.
None of those need a breach to become a problem. An innocent question is enough if the underlying permissions were always looser than the org chart implied. The same question applies to every assistant a team has connected, which is why a comparison of ChatGPT and Gemini increasingly turns on administration rather than model quality.

Why Australian businesses should check this
For Australian firms the pressure arrives from contracts before it arrives from regulators. Professional services agreements routinely carry confidentiality clauses written before generative AI existed, and a client is entitled to ask what your assistant can read.
The Privacy Act 1988 and the Australian Privacy Principles do not prohibit this configuration, but they do put accountability on the organisation holding the data. For a firm handling client files, that makes the admin console a compliance artefact rather than an IT preference, and the same logic reaches any document automation touching contracts or personal information.
Regulated practices have the sharpest version of the problem. Legal and accounting teams manage matter level confidentiality that Workspace permissions were never designed to express, which is why AI tools for lawyers tend to be assessed on isolation before capability.
The setting to look at first
The path is admin.google.com, then Generative AI, then Gemini in Workspace, then the Workspace Intelligence Sources block. From there a source can be switched off for the whole organisation in a few clicks.
The individual case is where it turns awkward. Gewirtz found the per user selector is marked view only, so restricting one employee means moving them into their own organisational unit or adding them to a new group. That converts a permissions decision into a directory restructure, which is exactly the kind of change that gets deferred.
It is worth settling before a client asks. The organisations that handle this well will be the ones that decided deliberately which surfaces their assistant should reach, rather than finding out during an audit.
