Note that this is specifically Consumer Gemini with Connected Apps; Workspace Gemini (and Workspace Intelligence, the RAG thing for enterprise/Workspace) still claims not to be subject to human review:
So basically if you use consumer Gemini and specifically authorize the GMail app, for example, then it will sometimes use the GMail tool to load your mail from GMail, and that data (like all data in consumer Gemini) might be given to the human annotation pipeline.
This whole thing seems deeply stupid to me from Google's standpoint, although they've never been good at product management - there's _way_ too much nuance to what's going on and even if they aren't technically affected, this kind of thing will almost surely make their paying enterprise customers leery.
It allows Gemini to autonomously search anything in your Drive or email. If it deems it relevant to the conversation, Gemini can pull summaries, snippets, or even the file plaintext of the file into the context window, making it eligible for human review.
100% agreed on the Enterprise distinction—this disclosure applies specifically to Consumer Gemini with Connected Apps, not enterprise/edu Google Workspace domains with managed DPAs.
The core concern here is fourfold:
1. *The Walked-Back Commitment:* When Google launched Workspace extensions for consumer accounts in September 2023, they gave an explicit guarantee: personal content from Gmail, Docs, and Drive "is not seen by human reviewers." That isolation promise was quietly stripped in the October documentation updates without proactive notice or re-consent for users who originally enabled the integration under those terms.
2. *Opaque Consent Architecture:* Current onboarding modals for Connected Apps focus on utility while failing to clearly disclose that enabling the connector subjects pulled Drive/Gmail payloads to human review pipelines (as outlined in Help Answer 16836988).
3. *Ambient, Tag-Free Ingestion:* Many users assume Connected Apps only trigger when explicitly invoked via `@Gmail` or `@Drive`. In testing, ambient conversational prompts (e.g., asking about household bills or travel plans) autonomously trigger background Workspace retrieval tools, injecting raw email/document snippets into the session prompt—which then becomes eligible for the human annotation pipeline.
4. *Coupled Privacy Controls (All-or-Nothing Opt-Out):* Unlike OpenAI and Anthropic—which decouple model training/human review toggles from basic application utility like conversation retention—Google bundles human review directly into "Gemini Apps Activity." If a user disables human review, Google disables core features like chat history retention and personalized context.
You're spot on about the brand confusion. When an enterprise employee uses their personal Google account for contractor or side work, or when compliance officers see "Gemini" in headlines, the nuance gets lost fast.
> Google Photos data is used to infer your interests, relationships to people in your photos, and where you've been, including by associating your face with corresponding location data and timestamps. This data is also used to do things for you, like look up your info or add items to your grocery list, and to improve Google services, including training generative AI.
How does that work exactly if the images aren't even real and are generated by AI? When Google siphons up these photos to help them train their AI, the AI is now going to make associations as depicted in that photograph?
That you really were with that person, in that place, in that situation, on that date? As some sort of a fact?
Can someone else create that image and just put it in their library and then have that become a fact on your behalf when picked up, analyzed and trained on?
Really frustrating that our privacy eroded to the point anyone would think this is reasonable.
But it does explain some conversations with Gemini when I wonder why it's talking about particular things. Like asking it about consultation processes for changes to parks only to get a response that's uncannily tuned to the email I got prompting the enquiry.
There are 5 major frontier labs competing in the US alone with serious capital. Perhaps only 2 of them are are truly in contention for the top spot, but it's highly competitive. Then throw in the handful of very serious sub-frontier open weights labs, who are creating huge price competition for the frontier labs, and I have no idea how you can come to this conclusion.
Cloud is no different: 3 major cloud vendors, with 1 clear leader...and now a long tail of tier 2s and neoclouds that are also investing incredible sums to compete.
Where is the pre-structured duopoly? Where is the profit moating?
There are 2 US frontier AI labs. The others aren’t competing at the frontier hence no concern for national security or moating. The moat is 1 main, 1 backup/pusher.
Any tool having manual human reviewers or annotator like all of Gemini and notebooklm (now Gemini notebook) suffer from this. Your data has high chance being read by random people or being leaked for training because they might flag it then use it to "improve" the model.
I had a Google Meet Gemini transcription take well over 24hours once recently, rather than the usual <1hr. I could only assume manual intervention was the cause.
It might also explain why Google do not provide a page to view transcription job queues from meet. You just have to sit and wonder
There are explicit data retention contracts for Workspace (including Gemini transcription) and Google claim this data is never used for training or human review, so this is fairly baseless.
This gist is specifically about the consumer version of Gemini used in Gemini Apps. Basically if you connect your consumer Gemini (which always sends stuff to the training pipeline) to your personal Google Apps account, it gets tools that it can invoke to read your Google Apps data, which then might appear back in your consumer Gemini transcripts that are getting sent to the training pipeline.
Doing this is a stupid product move from Google (per the usual) because it confuses their enterprise customers into thinking Really Bad Stuff like this is happening.
I am in the process of moving my stuff from Google to selfhosted alternatives. Google Photos => Immich, Google Drive => probably Next Cloud or some better alternative.
I am also looking into how to selfhost my email (and yes I know, the horror of inboxing from self hosting. But I am at the point of not only willing to give it a tr, but actively saying "fuck you" to everybody that does not receive my email, because they are on Outlook or O365 (or Google, bt inboxing into Google is easier imho).
Especially with my paid Google now daily having spam in my inbox, while valid emails are so often in spam, that I have to check spam at least once per week.
As someone who self hosts many services, I would encourage you to switch to self hosted email last. Even though self hosting has been made much easier thanks to really incredible FOSS tools, you are going to have random learning pains, some delayed by a few months, and it would really suck to lose email for even the shortest durations. The others are easy to live without while you debug during the learning process.
Btw immich is incredible. I recommend Nextcloud AIO instead of raw Nextcloud. Absolutely great tools
I think the low cost reasonable option right now is to self-host a receiving address and then pick the email provider that looks the least trash and forward your stuff to them to actually handle sending/browsing. It gives you the option to go full self-hosted in the future if your mind changes with minimal headaches but costs practically no effort for the short term.
I’m looking to set up immich on my server. How do you back up your media? If you get hit by a bus how do you leave instructions and mechanisms to your family to restore the backup? My wife won’t move off of iCloud until I have a great answer for those.
I've spent the last decade building a Dropbox alternative that works on top of any storage protocol (SFTP, S3, SMB, IFS, GIT WebDAV, local disk, and literally everything else [1]) and its on github [2]
But it isn't - not for doing stuff like self-hosting files? It's actually incredibly cheap, most people probably have a commodity PC lying around and can shove a few cheap HDDs into it and RAID stripe it if they're ambitious... or just full mirror their files if they aren't. And if you're really scared of loss find some host provider and rent a cheapo instance somewhere geographically far from you to park a mirror of the data.
Of course they do. Now that all the freely available training material has been consumed, every website ingested, every image classified, every YouTube video watched, and even books scanned and destroyed, the insatiable hunger for more data must be fed.
"Rock Solid" Google was known to be privacy invasive. Idk how people trust Google for privacy when it comes to Ai. Look at this thread https://news.ycombinator.com/item?id=40965892
As far as securing the system -- yes there are about 25 layers of isolation and hardening to get through, to pwn the system. Even if you got through all of it, the data can just be restored on a previous system. Here are just some:
6) All network egress goes to a second Safebox, whose job it is to actually enforce policies before signing transactions / sending emails / using your credentials. So even if you totally compromised the first Safebox somehow, or even a single app on it, the policies still get enforced
7) All credentials / keys encrypted and split between Safeboxes
https://knowledge.workspace.google.com/admin/generative-ai/g...
So basically if you use consumer Gemini and specifically authorize the GMail app, for example, then it will sometimes use the GMail tool to load your mail from GMail, and that data (like all data in consumer Gemini) might be given to the human annotation pipeline.
This whole thing seems deeply stupid to me from Google's standpoint, although they've never been good at product management - there's _way_ too much nuance to what's going on and even if they aren't technically affected, this kind of thing will almost surely make their paying enterprise customers leery.
It’s not clear what clicking yes does. Is it my entire drive and Gmail? Do I have to give permission per-file/folder/email?
So I click no. And it just gives me the popup again later
The core concern here is fourfold:
1. *The Walked-Back Commitment:* When Google launched Workspace extensions for consumer accounts in September 2023, they gave an explicit guarantee: personal content from Gmail, Docs, and Drive "is not seen by human reviewers." That isolation promise was quietly stripped in the October documentation updates without proactive notice or re-consent for users who originally enabled the integration under those terms.
2. *Opaque Consent Architecture:* Current onboarding modals for Connected Apps focus on utility while failing to clearly disclose that enabling the connector subjects pulled Drive/Gmail payloads to human review pipelines (as outlined in Help Answer 16836988).
3. *Ambient, Tag-Free Ingestion:* Many users assume Connected Apps only trigger when explicitly invoked via `@Gmail` or `@Drive`. In testing, ambient conversational prompts (e.g., asking about household bills or travel plans) autonomously trigger background Workspace retrieval tools, injecting raw email/document snippets into the session prompt—which then becomes eligible for the human annotation pipeline.
4. *Coupled Privacy Controls (All-or-Nothing Opt-Out):* Unlike OpenAI and Anthropic—which decouple model training/human review toggles from basic application utility like conversation retention—Google bundles human review directly into "Gemini Apps Activity." If a user disables human review, Google disables core features like chat history retention and personalized context.
You're spot on about the brand confusion. When an enterprise employee uses their personal Google account for contractor or side work, or when compliance officers see "Gemini" in headlines, the nuance gets lost fast.
> Google Photos data is used to infer your interests, relationships to people in your photos, and where you've been, including by associating your face with corresponding location data and timestamps. This data is also used to do things for you, like look up your info or add items to your grocery list, and to improve Google services, including training generative AI.
How does that work exactly if the images aren't even real and are generated by AI? When Google siphons up these photos to help them train their AI, the AI is now going to make associations as depicted in that photograph?
That you really were with that person, in that place, in that situation, on that date? As some sort of a fact?
Can someone else create that image and just put it in their library and then have that become a fact on your behalf when picked up, analyzed and trained on?
Where does that all lead?
But it does explain some conversations with Gemini when I wonder why it's talking about particular things. Like asking it about consultation processes for changes to parks only to get a response that's uncannily tuned to the email I got prompting the enquiry.
No mention of I just read your emails etc .
in case the reply is not clear, I mean to emphasize and amplify the alarming Google behavior described above
There’s nowhere to go.
Cloud is no different: 3 major cloud vendors, with 1 clear leader...and now a long tail of tier 2s and neoclouds that are also investing incredible sums to compete.
Where is the pre-structured duopoly? Where is the profit moating?
You can’t build a fighter jet in your garage.
It might also explain why Google do not provide a page to view transcription job queues from meet. You just have to sit and wonder
This gist is specifically about the consumer version of Gemini used in Gemini Apps. Basically if you connect your consumer Gemini (which always sends stuff to the training pipeline) to your personal Google Apps account, it gets tools that it can invoke to read your Google Apps data, which then might appear back in your consumer Gemini transcripts that are getting sent to the training pipeline.
Doing this is a stupid product move from Google (per the usual) because it confuses their enterprise customers into thinking Really Bad Stuff like this is happening.
Unfortunately, the alternatives are all so bad...
I am also looking into how to selfhost my email (and yes I know, the horror of inboxing from self hosting. But I am at the point of not only willing to give it a tr, but actively saying "fuck you" to everybody that does not receive my email, because they are on Outlook or O365 (or Google, bt inboxing into Google is easier imho).
Especially with my paid Google now daily having spam in my inbox, while valid emails are so often in spam, that I have to check spam at least once per week.
Btw immich is incredible. I recommend Nextcloud AIO instead of raw Nextcloud. Absolutely great tools
[1] https://www.filestash.app/docs/plugin/#storage
[2] https://github.com/mickael-kerjean/filestash
Self hosting email is a nightmare and can bite you in the ass big time.
Clearly the person who bakes my bread should've known better and learned Linux a decade ago to be prepared for this. Sucks to be them I guess.
I have a strong 'told you so' feeling right now.
You mean the default prompts we use in safebots? https://safebots.ai/safebots.html
As far as securing the system -- yes there are about 25 layers of isolation and hardening to get through, to pwn the system. Even if you got through all of it, the data can just be restored on a previous system. Here are just some:
1) Immutable startup disk, nix-based builds
2) M-of-N signatures on any supply chain updates
3) ZFS based disk encryption and sync
4) Safecloud for storage (https://safebots.github.io/Safecloud)
5) gVisor isolation for all Docker containers
6) All network egress goes to a second Safebox, whose job it is to actually enforce policies before signing transactions / sending emails / using your credentials. So even if you totally compromised the first Safebox somehow, or even a single app on it, the policies still get enforced
7) All credentials / keys encrypted and split between Safeboxes
I could go on...