In October 2026 Google reopened a blog post from 2023 and added a parenthesis to it. The post is the company’s “responsible approach to building guardrails for generative AI”, and the sentence it qualifies is a promise: if you connect Workspace to Gemini, “your content from Gmail, Docs and Drive is not seen by human reviewers, used by Bard to show you ads, or used to train the Bard model.” The appended note, dated 16 October 2026, reads: “It can also be subject to human review in limited circumstances.”

That parenthesis is the whole story in one sentence, and it is on Google’s own page, unedited, directly beneath the promise it reverses. A write-up circulating on GitHub this week says Google “quietly reversed this commitment”. Half of that is right and the more interesting half is wrong.

How the scope is drawn

Google’s current position on this is not in the blog post. It is in a help page, About personalization with Connected Apps, and in the Gemini Apps Privacy Hub, last updated 26 October 2026. Between them those two pages carry fifteen separate clauses that set the scope of what a human reviewer can see. Seven widen it. Eight narrow it. Counting them one at a time is the only way to find out what the eight narrowing clauses are narrowing.

The widening seven are mostly about reach. The data shared with Gemini “includes things like your saved data from Search services and YouTube, emails, files, events, photos, videos, info about your contacts (such as phone numbers, addresses, and birthdays), info from Google Wallet”. It “may relate to topics you find sensitive, like race, religion, and health, or confidential info.” It is forward-dated: “if new data is added to these apps in the future, Gemini will have access to that data.” Disconnecting an app later “doesn’t delete any data from Gemini Apps Activity”, and deleting your activity does not reach anything already seen — that data can be deleted “unless it has already been reviewed by trained reviewers”, and reviewed material is “retained for up to three years”.

The axis the limits sit on

Now the eight that cut the other way, each with what it actually bounds:

  • “a subset of Connected Apps data” is reviewed. Bounds frequency. The subset is not sized.
  • “Gemini does not train generative AI models directly on your Gmail inbox, Drive, Contacts, Calendar or other Workspace apps.” Bounds a training pathway, which is a different pipeline from human review.
  • Privacy-preserving techniques “like differential privacy” are applied “before using them to train generative AI models”. Training again, and hedged twice: “take steps to” and “like”.
  • Turning off Keep Activity stops future review. Prospective only — and the Privacy Hub adds that “other Connected Apps like Google Workspace are unavailable if you turn off Keep Activity”, so the control and the feature are mutually exclusive.
  • Temporary chats “are not used to improve Google AI with help from human reviewers”. But the same Hub page says that even then, Google uses chats to protect its users “including with help from human reviewers”.
  • The 2023 blog update’s “limited circumstances”. Bounds frequency, undefined.
  • Humans “don’t view media from your Google Photos library”, except for abuse, harm, or submitted feedback.
  • Before review, Google takes steps including “disconnecting interactions from your Google Account and reducing personal information in this data, unless it’s to address abuse or harm.”

Six of those eight bound how often, through which pipeline, whether your name is attached, or when you can switch it off. Only the last two bound what is in front of a reviewer, and both carry the same abuse-or-harm exception. The Photos clause is the stronger of the pair and it is worth stating plainly: Google has committed in writing that a category of content is not looked at. That is a real limit, and it is the limit that the equivalent clause for mail and files does not contain.

Printed twice, qualified neither time

What the mail-and-files clause contains instead is this, from the training section of the help page: “In some cases, for example, where the email or file is brief or particularly relevant to respond to the prompt, the summary may just be the email or file.”

The same sentence appears again, verbatim, in the section headed “What data is included in human review”. Both times it is unqualified — no sampling rate, no length threshold, no redaction promise attached to it. The word doing the work in every reassuring sentence on the page is “summaries”, and this is the sentence that tells you a summary and a document can be the same object. Everything downstream inherits that: the reviewer sees “interactions, summaries, excerpts and inferences from your relevant emails and files”, and the summary may just be the email.

Google also writes, on the same page: “Please don’t connect your apps to Gemini Apps if they contain confidential information that you wouldn’t want a reviewer to see.” That instruction is only coherent if the preceding sentence means what it says.

Why “quietly” is the wrong word

The GitHub write-up’s accusation of stealth does not survive the documents. Google went back to a three-year-old page and dated its own correction, rather than leaving the obsolete promise standing. It published the sentence that undercuts its own framing, twice. It named sensitive categories explicitly rather than letting “data” cover them. None of that is what concealment looks like, and saying so is not a concession — it is why the rest holds up. The problem is not that Google hid the change. It is that the change is documented in a help page about personalization, while the promise it replaces is in a post about guardrails, and only one of those two is the kind of page anyone reads before clicking connect.

There is one gap running the other way, too. “Limited circumstances” in the blog update is softer than what Google’s own help page describes, which is a standing review programme with a three-year retention tail. The marketing page understates the documentation. That is the usual direction for this split, and it is the reason to read the second document rather than the first.

What the pages do not settle

The GitHub author’s central empirical claim is that Workspace retrieval fires without being asked — that “ambient mentions of personal topics (trips, bills, names, reservations) autonomously trigger background Workspace lookups”, producing anything from a summary to “the entire text dump of the email, including its headers”. We did not reproduce that. It is one person’s self-report with a stated method, no account was tested here, and nothing in Google’s three pages confirms or denies that retrieval is triggered without an explicit request. Treat it as untested.

What the pages do settle is narrower and does not need the reproduction to be true. A document you connect can be reviewed by a contractor; the reviewed copy survives your deletion by up to three years; the only way to withdraw consent prospectively also withdraws the feature; and the sentence governing how much of the document travels has no qualifier on it at all. The figure nobody publishes is the one that would make all of this measurable: how often “the summary may just be the email or file” is the case. Google says subset. Nothing on either page says subset of what.