Skip to content

Gemini Privacy: Before Connecting Candidate Email and Files

Check Gemini privacy before connecting candidate email and files: account protections, connected-app access, retained copies, and practical cleanup steps.

Ernest Bursa

Ernest Bursa

Founder · · 9 min read
A recruiting lead and IT administrator review paper records beside a closed laptop in an office.

Before connecting candidate email or files to Gemini, check the exact account, product surface, and applicable edition. Consumer Connected Apps and qualifying managed Google Workspace services have different privacy terms. Disconnecting an app does not delete earlier Gemini activity or every copy of the candidate information it helped produce.

That distinction matters when a recruiter asks a seemingly small question: “When can this candidate interview?” The source thread may also contain salary expectations, an accommodation request, or a private reference. Approving that task requires knowing which information crosses which boundary, and what happens afterward.

A Hacker News discussion brought Gemini’s handling of personal email and files into focus. The linked repository’s allegations do not independently establish a breach, full-file ingestion, or a new policy change. Google’s own documentation provides a firmer basis for the team’s decision. This guide compares those documented boundaries; it does not report a hands-on test of Google’s implementation.

Identify the Gemini surface before approving candidate data

“Gemini” does not identify a single data-handling arrangement. A chat at Gemini Apps, a Gemini feature inside Gmail, and an API request from your ATS can follow different terms, even when the same company supplies the model.

Record the actual signed-in account and where the task runs. A company email address, a paid subscription, or access to a Workspace mailbox does not settle the question. Google’s Gemini app administration guide distinguishes a protected core service under a qualifying Workspace edition from an additional service subject to general Google and Gemini Apps terms.

Use this table as a review starting point, not blanket approval:

Account and surface Documented distinction Recruiting-team decision
Personal-account Gemini Apps with Connected Apps or Personal Intelligence Consumer notices govern derived information, activity, improvement, and review. Check those conditions before connecting candidate sources.
Gemini inside Gmail or Docs with a personal Google AI plan Covered in-app features have separate commitments from sharing content into Gemini Apps. Verify the particular feature and its terms; do not copy conclusions from another surface.
Gemini app as a core service under a qualifying managed Workspace edition Enterprise commitments and administrator controls apply within their stated scope. Confirm entitlement, permissions, history, and internal approval.
Gemini app as an additional service without a qualifying Workspace edition General Gemini Apps terms may apply despite a managed identity. Pause identifiable candidate-data use until the owner confirms the boundary.
Account, surface, or edition unknown Applicable protections have not been established. Resolve the unknowns before proceeding.

These are operational recommendations. A row in the table is neither a legal conclusion nor evidence that a workflow meets your organization’s obligations.

What consumer Connected Apps can do with email and files

Google distinguishes training directly on an entire underlying inbox or library from using material brought into Gemini interactions. That distinction still leaves meaningful candidate-data exposure to assess.

Google’s Connected Apps explanation says that, with Keep Activity on, relevant summaries, excerpts, generated media, and inferences can contribute to model improvement. A short email or relevant file may effectively be reproduced in an interaction. Calling the result a “summary” does not establish that it contains no personal information.

Google also describes human review of a subset of interactions, measures to reduce identifying information and separate interactions from accounts, and exceptions related to abuse or harm. This does not prove that reviewers read every connected inbox. It does mean a recruiting team should assess information derived from the source, as well as the original source permissions.

Do not use “turn history off” as a universal answer. The Gemini Apps Privacy Hub describes exceptions and separate safety processing; disabling Keep Activity does not guarantee that no human can access information. Some integrations also depend on activity settings. Check whether your intended connection works under the configuration you actually approve.

There is an important personal-account exception. For covered Gemini features inside Workspace apps, Google’s Gmail privacy explanation states that content is not used to train or improve generative models, and prompts and output are not stored without permission. Sharing personal-account Workspace material into Gemini Apps or Search Personal Intelligence follows those other surfaces’ terms. Workspace Experiments also has separate terms.

The useful question is therefore specific: “Does this scheduling task run inside the covered Gmail feature, or does it send thread-derived material into Gemini Apps?” “Is Gemini private?” is too broad to support an approval.

What changes under qualifying Google Workspace

Qualifying managed Workspace services have enterprise commitments, but you still need to verify their scope and settings. Those commitments do not transfer automatically to a personal-account connection.

Google’s Workspace Privacy Hub says that content is not human reviewed or used for generative-model training outside your domain without permission. Existing Workspace permissions govern access. Preserve those qualifications when documenting the protection; “Google never retains or reviews anything” would be a different claim.

Retention also varies by surface and administrator configuration. For example, Google’s administration guide describes Gemini-app conversation-history settings of 3, 18, or 36 months. Turning history off does not immediately remove earlier history, and the service can still retain conversations for up to 72 hours to provide the service. History off also disables Workspace integrations for that surface. Those details are not universal settings for every Gemini feature in Gmail or Docs.

Ask the administrator to confirm the actual service entitlement, applicable organizational group, and retention configuration. Record the answer with a dated source link. If someone changes edition or moves a task to another surface, review the workflow again.

Disconnecting is only one part of cleanup

Removing a connection, deleting an AI conversation, and deleting the original email are separate operations. Treat cleanup as a review of copies, rather than a single switch.

Google’s Connected Apps documentation says disconnecting does not erase Gemini Apps Activity. Deleting that activity does not erase data held by other services. A summary pasted into a shared document has its own lifecycle. Google’s consumer Privacy Hub also says deleting activity does not delete reviewed chats. Reviewed chats can be retained for up to three years; that is a particular retention exception, not a universal retention period for all Gemini data.

For a candidate-data request or routine cleanup, inspect at least these locations:

Copy or access route What to check
Original email or file Source-system access, retention, and the action required there.
Connected-app authorization Whether the access route remains enabled and who can revoke it.
Gemini conversation or activity Applicable deletion controls and documented retention exceptions.
Generated summary or draft Whether it remains in a chat, an ATS field, or a saved document.
Shared or exported output Recipients, destination owners, and their separate cleanup process.

Assign an owner to each row. Record completed actions and unresolved retention limits separately. Do not promise that disconnecting retrospectively retracts information already returned, or that deleting the candidate’s original attachment removes every derived copy.

Record one recruiting task and its remaining copies

A useful approval describes a bounded task, its necessary information, and its cleanup owner. “AI use approved” leaves too much unspecified.

Consider this hypothetical: a recruiter wants an availability summary from an applicant thread containing an unrelated accommodation request. The scheduling task needs available times and perhaps a time zone. It does not need the accommodation details. Prefer selecting the relevant material before processing, where the chosen workflow permits that selection, rather than asking the model to ignore sensitive information after receiving it.

Removing a name can reduce exposure, but it does not guarantee anonymity. A rare role, location, employer, or quoted conversation can still identify someone. Check the selected content, including attachments and earlier replies, against the task’s approved data categories.

Keep a short record with these fields:

  • Task and data: the requested result, necessary fields, and excluded information.
  • Account and surface: signed-in account, exact Gemini feature, and verified edition or service status.
  • Access and approval: connected source, access scope, task owner, and internal approver.
  • Terms and settings: dated policy links, improvement and review conditions, and applicable history settings.
  • Copies and cleanup: original, conversation, output destinations, responsible owners, and known retention limits.
  • Review trigger: an account, connector, surface, edition, or policy change.

This record helps a teammate retrace the decision and locate the resulting copies. It also exposes a blocked approval clearly: if nobody can establish the service status or account owner, the workflow is not ready for identifiable candidate material.

Where Kit fits in the review

An ATS connection introduces another data boundary to document. Review what the assistant can request, which account grants access, and where returned records can remain. Our guide to AI providers and hiring-data isolation explains why authorized access and downstream provider handling need separate attention.

Kit’s documented Gemini Apps custom-app route currently requires a personal Google account, US availability, English, age 18 or older, and Keep Activity enabled. Work and school accounts are excluded from that route. You cannot solve its consumer-account boundary by assuming the same connector can move to a protected Workspace account. Gemini CLI is a separate client; its MCP support does not establish Gemini Apps availability.

Kit’s API-based AI features are separate processing routes too. Consumer Gemini Apps documentation does not establish their API data-use terms. Likewise, revoking a Kit connection can stop future authorized calls, but it cannot retract records already returned to a client. Local candidate anonymization does not prove deletion from Google conversations, exported summaries, or provider-held material.

Before approving your next recruiting task, fill in the account, surface, necessary fields, and copy owners. If one of those is unknown, resolve it before connecting candidate records. Use the same review for your ATS and every AI client around it.

Related articles

Try Kit for 30 days.

Hiring, security reports, and training in one account, for teams where none of it is a full-time job. Free for 30 days, card required. Cancel before it ends and you pay nothing.

Get started free