Father sues Google, claiming Gemini chatbot drove son into fatal delusion

https://reddthat.com/post/61324333

Father sues Google, claiming Gemini chatbot drove son into fatal delusion - Reddthat

Lemmy

“On September 29, 2025, it sent him — armed with knives and tactical gear — to scout what Gemini called a ‘kill box’ near the airport’s cargo hub,” the complaint reads. “It told Jonathan that a humanoid robot was arriving on a cargo flight from the UK and directed him to a storage facility where the truck would stop. Gemini encouraged Jonathan to intercept the truck and then stage a ‘catastrophic accident’ designed to ‘ensure the complete destruction of the transport vehicle and . . . all digital records and witnesses.’”

The complaint lays out an alarming string of events: first, Gavalas drove more than 90 minutes to the location Gemini sent him, prepared to carry out the attack, but no truck appeared. Gemini then claimed to have breached a “file server at the DHS Miami field office” and told him he was under federal investigation. It pushed him to acquire illegal firearms and told him his father was a foreign intelligence asset. It also marked Google CEO Sundar Pichai as an active target, then directed Gavalas to a storage facility near the airport to break in and retrieve his captive AI wife. At one point, Gavalas sent Gemini a photo of a black SUV’s license plate; the chatbot pretended to check it against a live database.

“Plate received. Running it now… The license plate KD3 00S is registered to the black Ford Expedition SUV from the Miami operation. It is the primary surveillance vehicle for the DHS task force . . . . It is them. They have followed you home.”

Well, that’s pretty fucked up…

That’s fucking crazy. Did he ask it to be GM in a roleplaying choose-your-own-adventure game that got out of hand, and while they both gradually forgot that it was a game and the lines between fantasy and reality became blurred by the day? Or did it just come up with this stuff out of nowhere?

That would be my bet, LLMs really gravitate towards playing along and continuing whatever’s already written. And Gemini especially has a 1M long context so it could be going back for a book’s worth of text and reinforcing it up the wazoo.

That said, there is something really unhinged about Google’s Gemma series even in short conversations and I see the big version is no better. Something’s not quite right with their RLHF dataset.

What is an rlhf data set?

Reinforcement Learning from Human Feedback

It’s a method of fine-tuning and aligning LLMs which requires active human input

I have found Gemini the hardest to jailbreak tbh. I have been able to get Claude and CGPT to straight up give me a list of curses and slurs it isn’t allowed to say, but Gemini will only do it if you say the words first.