Apple notes = private
Google docs = ?
Siri requests = used for training models
Emails you send in gmail = ?
I'm seeking to understand what things people might think are private, because they're not posted on the open web, but where they're used for training AI models.
https://www.darkreading.com/risk/google-wardriving-how-engin...
I am the co-founder of a small by design company offering cloud services, and we spend a lot of time to make sure that our ToS reflect what happens operationally. Also money since we review it with a lawyer specialised in open source software. There are some other entities from chatons.org that care about such documents as well.
https://workspace.google.com/learn-more/security/security-wh...
https://learn.microsoft.com/en-us/compliance/regulatory/offe...
TLDR TOS.
> Google Docs, Sheets, & Slides uses data to improve your experience
> To provide services like spam filtering, virus detection, malware protection and the ability to search for files within your individual account, we process your content.
So yes, they do use your data for things like training AI. However it does not seem to be for general AI like Bard but for ML systems within the product in question.
"...from public dialog data and other public web documents..."
LaMDA 2 paper: https://arxiv.org/abs/2201.08239
My overview of Google Bard including dataset: https://lifearchitect.ai/bard/
My overview of Google PaLM and Pathways family including dataset: https://lifearchitect.ai/pathways/
Compare with other models including the use of DeepMind's MassiveWeb/MassiveText and EleutherAI's Pile dataset: https://lifearchitect.ai/whats-in-my-ai/
One exception to that is scanning for CSAM and Terrorism and DMCA. And with DMCA, it's automated based on file hash, and you still maintain access to your files, you are just limited from sharing them. Ads in gmail aren't based on content, but other online activity while logged in.
I think the other exception to that is smart compose. AI models do use email content for training data, but the output of those are strictly for use locally while writing emails. I imagine it's also siloed per user.
EDIT: Not a google employee, I apologize if my assertions seem too strong.
EDIT2: https://en.wikipedia.org/wiki/Federated_learning
"We have always maintained that you control your data and we process it according to the agreement(s) we have with you. Furthermore, we will not and cannot look at it without a legitimate need to support your use of the service -- and even then it is only with your permission. Here are some of the additional measures we take to ensure your privacy: (reference: GCP Terms).
In addition to these commitments, for AI/ML development, we don’t use data that you provide us to train our own models without your permission. And if you want to work together to develop a solution using any of our AI/ML products, by default our teams will work only with data that you have provided and that has identifying information removed. We work with your raw data only with your consent and where the model development process requires it. "
https://cloud.google.com/blog/products/ai-machine-learning/g...
https://support.google.com/mail/answer/6603?hl=en
https://arxiv.org/abs/1906.00080
https://ai.googleblog.com/2017/04/federated-learning-collabo...
"Google Will Keep Reading Your Emails, Just Not for Ads" - https://variety.com/2017/digital/news/google-gmail-ads-email...
Edit: Updated comment. Misread it as being posted from somebody working at Google.
If they do it on searches not meant for publication then are they irresponsible?
It's enough to have one exception. We have to assume that a language model will be trained from them and government officials are using it soon. Just think about how much it can tell about the next planned terrorist events with their organizers.
Not in general true. Something I write in a diary I keep under my pillow has the exact same copyright status as something I publish on my blog (assuming no Creative Commons, etc. licenses).
Doing irresponsible things on private data is hot business model of the day. I'm not saying it's google; I'm saying common expectations about "responsibility" are worse than useless.
>We have always maintained that you control your data and we process it according to the agreement(s) we have with you.
Ah the "we surveil you fair and square, get over it" clause.
They're just implying they don't process your data, while actually saying "the answer to your question lies in the text of the service agreements". Evasive at best.
In the absence of a denial, and the presence of an obvious motive to do so, I have to say my guess is they do use your gdocs and gmail data to train AI.
kek
So if there is an AI model for each user trained on the user's writing does that mean Google now also has the means to forge convincing emails?
There is no such thing as "scanning for DMCA". DMCA prescribes a process for reacting to complaints of copyright infringement, not a process for preemptively scanning content for material that might potentially generate such complaints.
OpenAI is using book1 (BookCorpus ?) and book2 sources. By the number of tokens, this seems less than a million book in total.
[1] https://www.blog.google/products/search/15-years-google-book...
The Google BERT paper (Devlin et al., 2018) also references it: https://aclanthology.org/N19-1423/
Privacy questions aside (as important as they of course are), it's very important to know what a model was trained on exactly: if Wikipedia was used in the training set, you can't use questions from Wikipedia to test it (as that would be cheating) - test data must be as "unseen" as a good exam.
It has also been analyzed here and here:
No for ads.
Yes for training models for Smart Reply, from: https://arxiv.org/pdf/1606.04870.pdf
Privacy Note that all email data (raw data, preprocessed data and training data) was encrypted. Engineers could only inspect aggregated statistics on anonymized sentences that occurred across many users and did not identify any user. Also, only frequent words are retained. As a result, verifying model’s quality and debugging is more complex.
Do their Terms of Service, etc., allow them to use that text for training models ?
Now, your next question is presumably going to be "why don't they just lie?".
First, the lie would be very expensive. What do you think privacy regulators around the world would think of it? What do you think that lying to customers about how their data is used would do to the their $25 billion / year enterprise business?
Second, the lie would get discovered very quickly, because somebody would leak it. Plenty of much less serious things seem to leak weekly.
Third, there's very little to be gained by using the data and lying about this compared to the alternatives (say nothing; tell the truth; don't use the data). "Use the data but lie about it" is the absolute worst strategy possible.
https://en.wikipedia.org/wiki/Max_Schrems#Schrems_I
(You will notice that the enforcement has been very slow moving on that one, since despite this, US companies are still used quite a lot in EUrope.)
Not saying they are lying, of course, but I don't think the incentives against it are as strong as you do.
But then, I'm skeptical enough about privacy policies that I've stopped reading them. The vast majority of them, including the ones I've read from Google, leave more than enough room for the companies to do pretty much anything they want with your data.
> We collect information to provide better services to all our users
Which is vague to the point of uselessness. Google could argue that pretty much any purpose they put that information to is to provide better services to all their users.
This sort of thing is why I ignore privacy policies. They tend to have quite a lot of verbiage of that sort. Even when they appear to solidly say they won't do something, often there's wording elsewhere that provides an exception.
The only way I'd put any stock in a privacy policy is if I have my attorney review it and explain to me what it really says (not being a lawyer, I am not capable of adequately interpreting contracts). Since it's entirely unrealistic to have an attorney review every privacy policy, the safest approach is to assume that they allow the company to do whatever they wish in the end.
And that's all assuming that companies put any real effort to adhering to their privacy policies.
From https://policies.google.com/privacy?hl=en-US
> We don’t show you personalized ads based on your content from Drive, Gmail, or Photos.
It is a direct and unambiguous statement in plain language saying that Gmail content is not used for ad targeting.
Given your objection was based on an invalid premise, are you now willing to change your opinion?
"We will not scan or read your Gmail messages to show you ads."
Whether it's being scanned for other reasons is another question.