Is Google Gemini trained on Google docs? One indie developer thinks so, after it told players about his unreleased game plans : privacy

[-] ExLisper@lemmy.curiana.net 64 points 3 weeks ago

The headline is wrong. They don't claim Gemini was trained on docs.

" It also seemed to be able to regurgitate accurate information related to unreleased mechanics that the developer says they had only written into GDocs the day before the player’s AI interaction."

If it was written only the day before there's no way the model was trained on that data. The claim is Gemini accessed his docs while generating the answer. We know Gemini can access public docs so I think the question here is if it accesses private docs in a breach of security, the doc had wrong permissions or it was all just a coincidence.

[-] elvith@feddit.org 9 points 3 weeks ago

I also wrote the same thing in a comment in another community. What you shouldn't forget in this case: Players love to speculate and argue about how to fix/improve/add a part of the game by introducing mechanic X. Other game studios may have tackled/expanded their games in a similar way. It could be an educated (or lucky) guess based on that or some general public description that was available.

Same for the character name - there may be some (unconscious?) pattern that decided the character names so far and the new character was a lucky guess based on that?

[-] ColeSloth@discuss.tchncs.de 4 points 3 weeks ago

Two lucky coincidences? Really?

[-] lambalicious@lemmy.sdf.org -1 points 3 weeks ago

If it was written only the day before there’s no way the model was trained on that data.

[citation needed]

[-] ExLisper@lemmy.curiana.net 3 points 3 weeks ago

You want citation about how time works?

[-] Kolanaki@pawb.social 1 points 2 weeks ago

How does time work? 🤔

[-] ExLisper@lemmy.curiana.net 1 points 2 weeks ago* (last edited 2 weeks ago)

You have 3 arrows of time: one points from the smaller universe to bigger universe, second points from universe with lower entropy to higher entropy and third points from what we perceive as past and remember to the future that we don't remember. Right now those 3 arrows of time are aligned and points in the same direction. It's possible that first arrow will change direction at some point, second will not and we don't know about the third. So on universe's level it's complicated. On human level we simply remember events "behind on" on the timeline and don't remember the ones "in front of us". It's not clear why or if this can change. Maybe it's different for other organisms.

[-] lambalicious@lemmy.sdf.org 1 points 2 weeks ago

Do any of us understand how time works? I still feel like it's past Friday!

I'm asking for citation on the idea that it's not possible for the model to have been trained on that data from the day before if it was already readily available for them.

[-] ExLisper@lemmy.curiana.net 0 points 2 weeks ago

https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-versions

[-] lambalicious@lemmy.sdf.org 1 points 2 weeks ago

That only says when a new given model is released; not whether the model can use data that finds ~~today~~ yesterday on the internet to fine-tune, which is the current issue. If you have two secrets, and an internet thing guesses both of them, it's not irrational to suspect there is a leak (or a "leak") the internet thing is grabbing from.

[-] Hexarei@beehaw.org 3 points 3 weeks ago

Model training takes a long time

[-] Abyssian@lemmy.world 29 points 3 weeks ago

Of course it is. Every company's AI is trained on every bit of data they can possibly access. All the laws against it and privacy options and toggles are meaningless. They only exist to make you feel like you have some control and a way to say no.

There is no such thing as a magic cloud, only computers that rich people own. Anything you put on them they will use for whatever they want.

[-] dev_null@lemmy.ml 1 points 2 weeks ago

And this story is not an example of this happening, for the simple reason that model training takes longer than 24 hours and there were no new Gemini model releases between the doc was written and the chat happening.

[-] Abyssian@lemmy.world 1 points 2 weeks ago

Didn't even notice that part. Training is definitely the wrong word to use. Checked out the Reddit thread and the screenshot of the conversation where it came up and it seems like total bullshit.

The person talking sends an image showing the character name at 2:44am and then between the two of them there are 8 more lines in that same minute, including the dev saying that the descriptions are bullshit. So we're supposed to think dude read several paragraphs of text and then him and his buddy sent those 7 other messages back and forth all within the same minute the image arrived?

I'm a big reader and type quickly, but that's not at all realistic. After all of that in one minute, only one more single minute passes before the dev sends a cropped screenshot showing that same name and saying it was the only place he'd ever typed it. So he got in to Google Docs and got the right document open to the right place then snapped and pasted the image in one minute.

And... the "proof" image that starts everything and is supposedly from a conversation with an AI doesn't show the interface at all. All we see is bulleted lists. That could be word, or it could be another screenshot from Google Docs. Nothing in any of it shows any AI interface saying anything.

Seems like a weak attempt at getting publicity for his game, which of course he links to the sales page for.

[-] dev_null@lemmy.ml 2 points 2 weeks ago

Nice, I didn't even notice the timing discrepancies. Google is obviously stealing all data it can to train AI, but in this instance it seems the developer is just making stuff up to market his game.

[-] onlinepersona@programming.dev 24 points 3 weeks ago* (last edited 3 weeks ago)

Come on, you can safely assume that anything and everything you pass to or through google is used to train AI. This isn't even news

[-] elucubra@sopuli.xyz 12 points 3 weeks ago

Just as I have proof that Android phones are listening and sending our conversations constantly, I'm willing to bet that Google is scraping all our Google docs, files in drive, and Gmail emails.

[-] infeeeee@lemmy.zip 17 points 3 weeks ago

What proof lol.

Google already had all the docs, they never said it was encrypted, as they also scan all the emails for "spam detection". It would have been more surprising if they didn't use it for training

[-] elucubra@sopuli.xyz -2 points 3 weeks ago* (last edited 3 weeks ago)

Proof about phones listening in.

"lol"

[-] infeeeee@lemmy.zip 5 points 3 weeks ago

Could you elaborate on that? Because it was refuted countless times that current technology can't do that, and I want to be up to date on the latest unhinged conspiracy theories.

[-] MadhuGururajan@programming.dev 8 points 3 weeks ago* (last edited 3 weeks ago)

"They are always listening" isn't exactly unhinged. Infact it's a bit unhinged to think they can't listen through current technology when that shit literally listens to people talking in the vicinity.

"The majority of the commands we issue to smart speakers are innocuous, perhaps asking to play a particular radio station or set an alarm. But whistleblowers have expressed concern that devices are waking up randomly and processing accidental recordings that should never have been made.

These are known as "false accepts" and data collected by one Google whistleblower shows that about 15 per cent of recordings occur by accident. This error rate indicates that smart speakers – essentially microphones in our homes – may not be as benign as we've been led to believe. "These technologies live in your home, they become part of your daily fabric," says Schaub. "When you phone a call centre, you typically get a message explaining that the call may be recorded for training purposes, but that explicit acknowledgment is missing from smart speakers."

https://www.thenationalnews.com/arts/are-big-tech-companies-actually-listening-in-on-us-via-our-phones-or-smart-speakers-1.898574

This is one source. 15% is too high even for accidental activations.

Edit: Found the most relevant article. I heard about this one maybe 6 months ago:

https://www.zdnet.com/article/new-report-claims-your-phone-tv-and-smart-speaker-are-spying-on-you-but-is-it-true/

[-] infeeeee@lemmy.zip 8 points 3 weeks ago

Yes, I know about this, but bugs in wake word detection is not "listening and sending our conversations constantly".

And on device STT capabilities will get better with each new generation of TPUs, so we are not far from ditching wake word detection as a technology, and phones will be able to do type everything they hear. But currently we are not there yet.

[-] ColeSloth@discuss.tchncs.de 1 points 3 weeks ago

The tech is there for certain. That part is super easy. It would just be putting the software in place. Also, it would be far more covert and energy efficient to simply record the radio onto the phone itself and then send it as a file every so often. Only recording\saving when voice is being picked up could be done in as little as 30MB a day.

[-] ludrol@programming.dev 3 points 3 weeks ago

Interesting that the claim couldn't be replicated. It could just be false positive.

[-] LaunchesKayaks@lemmy.world 2 points 3 weeks ago

I hope Gemini likes my really bad smut fanfics, cause that is the only thing I kept in docs when I used it lmao

[-] Marija_@programming.dev 2 points 2 weeks ago

PersonalIP isn't optional in the AI era.

Privacy