A solo indie game developer said that Gemini gave out "frighteningly accurate" details about his project's upcoming content, even though this information was not published anywhere.
A solo indie game developer said that Gemini gave out "frighteningly accurate" details about his project's upcoming content, even though this information was not published anywhere.
It all started in the game's Discord community, where one of the players decided to ask the AI "silly questions" about Operation Octo for fun. After receiving surprisingly accurate answers about the game, the developer, out of curiosity, asked the player to ask the AI about future content.
To everyone's surprise, the AI revealed the specific name of the new character — "Vantage Tripod."
"I never mentioned this name to anyone. As far as I know, the only place digitally where this information existed was in one of my own Google Docs. And it wasn't me who was talking to the AI!" shared the creator of the tower defense game Operation Octo under the nickname KlubKofta on Reddit.
The developer added that his game is quite small, so there is no “noise” or rumors online that could confuse the AI. This allowed the algorithm to produce a real leak without unnecessary guesswork.
In the comments under the post, users suggested that Google automatically uses files from Google Drive to train its models. However, in a comment to Polygon, the company representatives categorically denied this:
"Google does not scan your private content in Workspace (including Drive and Docs) to train our underlying AI models (including Gemini). Links to public documents can only be indexed if they have been published in the public domain somewhere."
Despite this statement, the indie developer insists that the document with the character's name was never distributed anywhere, so he still doesn't understand how the AI learned this data.
The author of the post later updated the post to mention another developer under the nickname kallico. She conducted her own experiment: she asked Gemini to generate ideas based on the title of her game and compared the results with her private design document, which was also not published anywhere.
The Google AI’s responses matched her document content with astonishing accuracy. When she asked the same questions to Claude’s neural network (which doesn’t have access to the Google ecosystem), the responses were standard and didn’t contain any privately described ideas.
This incident has sparked a new wave of concern among content creators about how private the information they store in cloud services is.
