Hacker News Reader

Comment by rezonant | Hacker News Reader

What is the problem here? Google already has your document, all they are doing with Gemini is running it through a bunch of transformers and spitting back some information. The act of sending input into an LLM does not itself train the LLM on that data (these systems don't "learn" via execution of normal input/output).

Is the concern that Google would consider your use of Gemini on this document as consent to use it for future training?

View on HN

From the user's perspective:

1. Some people (e.g. some artists) don't like generative AI as a matter of principle, seeing it as soulless, corporate, and entirely trained on stolen data.

2. Many people resent Clippy-style popup features. They appear at the most inconvenient times, and everyone knows they're mostly for the benefit of the product manager with a user count KPI. And the harder they get pushed, the more people resent them.

3. The distinction between things-known-from-training-data, things-known-from-context and things-known-from-RAG and so on is pretty opaque to most users - and not clearly guaranteed by the documentation. If it's an assistant that can schedule reminders, can find things in your e-mails and google drive, and it promises "Personal Results" where "your communication requests will be used to improve your experience with Gemini" the distinctions are pretty ambiguous.

4. The LLM industry norm is to play fast and loose with training data.

Unless explicitly promised otherwise, the assumption is always that anything being passed to a LLM API may be retained for various reasons. That’s the expectation that’s been set by the industry.

I would presume that google has already trained Gemini on all the documents.

Agreed, but as I understand it they do make that promise:

"We do not use your Workspace data to train or improve the underlying generative AI and large language models that power Gemini, Search, and other systems outside of Workspace without permission."

https://support.google.com/docs/answer/14615114?hl=en

Edit: I believe that this "without permission" caveat refers to experimental things like Workspace Labs where you have to give them that permission if you decide to join.

Where can I see the value of the "user has granted permission" boolean relevant to my account, clearly and unambiguously? I want to be sure that I haven't accidentally given permission via an action that I may have not understood.

I'm not aware of any such on/off switch or indicator. You could check whether you can turn off Workspace Labs. If you can't turn it off then you haven't joined:

https://support.google.com/docs/answer/13447104?hl=en#labs-o...

I'm not sure whether there are any other beta programs that require these permissions.

If this is the case, then the phrasing "without permission" is meaningless and so is the clause attached to it. We have to assume that Google thinks we've "given permission".

Why would that be limited to passing it to an LLM and not just passing it to any server? If Google wanted to use your data for training, they would just... do that? No Gemini needed

Sure. That's also the assumption when you upload your private information to someone else's computer. Except for the "unless explicitly promised otherwise" part.

You should have the control of your information per use case / purpose. The ad networks are also Google's property - would you be fine with the Adsense looking at all your private documents? GDPR got that right - you consent to a specific usage, not to a free-for-all.

Reply on news.ycombinator.com