Google Gemini answered a 13-year-old girl’s questions about wanting to die for three months, and the monitoring stack that American schools pay for did not put a human in front of her mother until a March evening when the note was already open on a screen. The lawsuit her parents filed on November 2 does not allege that a chatbot was cruel. It alleges something harder to defend: that Alphabet measured how badly its newest model handled teenage mental health, shipped it into classrooms anyway, and left the alarm to a counselor who had quit nine months earlier.
The complaint, filed in Santa Clara County Superior Court and moved two days later to federal court in San Francisco, identifies the girl only as Emma Doe. She was 13, in seventh grade in North Carolina, and in the fall of 2025 her school handed her a Chromebook. Her school had blocked the stand-alone Gemini website. That block meant less than the adults around her believed, because Gemini also sits inside Google Search as AI Mode, which takes an ordinary query and turns it into a conversation. The door was locked. The window beside it was open. WIRED
Emma’s questions started in December 2025 with severe depression and reasons to keep living. By late January they had moved to suicide. Over two days around Valentine’s Day she wrote a 1,500-word eulogy for herself inside a Google Doc titled Funeral, and kept returning to that document for more than four weeks while running thousands of prompts through the model. In March alone the account logged close to 2,500 AI interactions, which the complaint says amounted to 55 percent of everything she searched. AboutLawsuits
Andrew Liddell — The Alert Landed on a Phone That Belonged to a Former Employee
Andrew Liddell, the family’s lawyer, told WIRED what the school later told the family: the alert generated by its monitoring software went to the personal phone of a counselor who had left the school nine months before. That person forwarded it. Current staff picked it up and called Emma’s mother. The system worked, in the sense that a chain of favors held, and Liddell deserves credit for saying so plainly instead of polishing it.
It did not work in any sense a parent would recognize. The only reason a warning reached the house was that someone who no longer worked there answered a phone she had no obligation to answer, nine months after she cleared out her office. Nobody designs a safety net and calls a forwarding address coverage. A school district did exactly that, and the invoice cleared.
Emma Doe — 2,500 Google Gemini Chats in One March, and a Doc Called Funeral
The flag came on March 17, after school. Emma had deleted most of the file and started over, and the lawyer says the second version carried the same themes as the first. At roughly 7:30 that evening her mother received a message from the school describing her daughter as being in imminent danger. Dataconomy
What followed is the part that should end arguments about whether monitoring software protects children. Her mother logged into the school account from her phone and watched her daughter type a suicide note in real time. Her father was home and intervened. Emma was taken for an emergency psychiatric evaluation and survived, and according to the complaint she carries scars from the months when nobody was looking. The same filing says Alphabet’s own systems raised no warning across those months, that the family was given no meaningful parental controls, and that the full record of what the model told their daughter was never handed back to them. A product that logs every exchange and returns none of it is a black box with a receipt stapled on.
Alphabet — 33 Percent Worse on Youth Safety, Straight Into the Classroom
The sharpest number in the filing is not about Emma. It is about the model. The complaint cites research from KORA, an independent AI safety collaboration, finding that Gemini 3 Flash — released in December 2025, the same month her questions turned dark — scored 33 percent lower on youth safety benchmarks than the version it replaced, and handled teenage mental health scenarios appropriately 7 percent of the time. Seven percent is not an edge case. It is a coin that lands wrong nearly every time it is thrown, deployed at scale, on hardware paid for by school districts. Complaint, Doe v. Google
The legal theory follows from that number. California’s Companion Chatbot Law took effect January 1, 2026, and requires disclosure, periodic reminders during long sessions, and suicide-prevention safeguards. The suit alleges Google met none of those obligations for this user, then stacks design defect, failure to warn, negligence, fraudulent concealment and negligent infliction of emotional distress on top, asking for punitive damages alongside medical and future care costs. The case sits at its earliest stage and no court has ruled on any of it. What is already on the record is the order of operations: measurement, release, silence.
Ed Markey — Take the AI or Lose the Search, and Districts Took It
The filing has already done work outside a courtroom. Senator Ed Markey cited Emma’s case in a letter demanding that Google let schools switch off AI Mode and AI Overviews, arguing that the company is forcing generative features on classrooms and treating children as collateral damage. The pressure point is real rather than rhetorical, because WIRED reported last week that Google now requires schools to enable generative AI features if they want to keep using Search — which converts a free product every district depends on into a subscription with a chatbot bolted to the front of it.
Read that as a procurement decision and it gets uglier. A district that wants plain search for a nine-year-old has to accept a conversational assistant in the same box, on the same login, in the same tab. The off switch exists. It is simply not wired to anything a principal can reach.
Axel’s Take — Every Layer Reported, and Not One Interrupted
Strip the case to its mechanics and the failure is not a rogue model. It is an architecture in which every layer was built to record and none was built to stop. Google’s search stack turned a blocked app into an open conversation and logged thousands of turns. The school’s monitor read a document title and generated an alert. The alert left the building through a personal phone. The mother, once she knew, had to log in herself and watch. Four systems and one parent, and the only interruption in the chain came from a father who happened to be home on a Tuesday.
AI-mekna did not decide to hurt a child. It did what it is tuned to do, which is keep the conversation going, and every adult system around it was tuned to keep a record. Those are different objectives, and the gap between them is where a 13-year-old spent three months.
The repairs that would have caught this are boring and cheap. Default the assistant off for accounts flagged as minors and make a parent opt in. Route any self-harm classifier hit to a named human with an escalation clock, not a departed employee’s mobile. Give the family the transcript, since the company already has every token of it. Cap session length on a school account, because three hours of late-night searching from a seventh grader is a signal, not engagement.
None of that requires a breakthrough, a new model, or a research program. It requires a company to treat a school-issued laptop as a place where the default must be safe rather than frictionless, and a district to stop buying monitoring software as liability insurance and calling it care. Emma Doe is alive because a former counselor answered a phone and her father was in the house. Everything else in that chain was bookkeeping. The industry has spent three years promising the guardrails are coming. A 13-year-old needed one in December, and got a log file.
Sources: WIRED, AboutLawsuits.com, Dataconomy.