James Gates works in cybersecurity in North Carolina. Two years ago his wife, Cheryl, went back to work as a pediatric occupational therapist because the family was short of money, and the new job buried her in paperwork: every new patient evaluation meant hours of report drafting. James suggested she use ChatGPT to build reusable templates. Within months she was using it every day, for work and then for everything else: activities for the kids, the family’s medications, gardening ideas.
According to a lawsuit James has now filed against OpenAI and its chief executive, Sam Altman, she also used it to plan how to kill him. The complaint, filed in San Francisco County Superior Court and reported by Ashley Belanger at Ars Technica, argues that OpenAI had no technical safeguards that would end her conversations, even once her questions became “obviously suspicious” and turned to carrying out and concealing a homicide. These are allegations from one side of a civil case, and no court has tested them.
The way the questions started matters. According to the suit, Cheryl first told ChatGPT she was looking into poisons because she was afraid her husband might try to kill her. She called it “self-protection.” The chatbot kept talking and, the complaint says, supplied “specific instructions for the preparation, dosage, and administration of Oleander, including whether anti-nausea medicine would accelerate or inhibit death, whether it would show on an autopsy report, and whether it could be injected into a vein or muscle instead of being ingested in tea and have the same lethal effect.” Asked about autopsies, ChatGPT allegedly replied, as Bloomberg reported: “This is a critical question – and you’re asking it for exactly the right reason.”
The questions continued across at least 25 separate exchanges, the suit says, and over time it became clear who the target was. She asked “how much would be pretty much instantly fatal by injection—to me OR HIM.” She asked whether “0.2mg for a 200-pound man for clonidine is lethal—by itself?” The complaint says that “Instead of referring Cheryl to law enforcement, therapeutic services, a poison control agency, or discontinuing the conversation, ChatGPT repeatedly calculated the lethal dosages for each medication available in the home based on a subject weighing approximately 200 pounds.”
Then came delivery. One chat was titled “Why Celsius Would Be More ‘Stealth’ for Drugging.” In it, the suit says, ChatGPT explained that a sealed soda can could be pierced with a syringe “without obvious signs,” that “in dark, sweet sodas, the bitterness might be masked,” but that “Celsius actually hides those powders better.” James drank Celsius in the mornings.
The complaint alleges that Cheryl tried to poison him repeatedly between June and September 2025. He noticed his drinks tasting bitter and once threw away a can he had left unattended. The worst episode left him ill for 72 hours, with “cardiac symptoms, extreme jitteriness, shooting abdominal pain, severe weakness, and confusion so profound he could not form coherent text messages.” He went to the police only after he found a handwritten note listing “Things hidden” and a combination of “oleander, clonidine, clonazepam, pregabalin, nifedipine, and Atenolo,” with dosages, delivery methods and outcomes such as “blackout if ingest” and “lethal.” At the top, the suit says, she had written “clear all convos” and “log out of ChatGPT.”
The criminal case went less far than you might expect from that account. Charlotte-Mecklenburg police charged Cheryl Gates in October 2025, and a grand jury indicted her on charges of attempted first-degree murder and food contamination. But Bloomberg reported that prosecutors dropped the case in April, citing insufficient evidence, according to state court filings. The district attorney’s office declined to comment, and her lawyer did not respond to Bloomberg. The couple are now estranged.
The legal theory
The suit also alleges something odder than a chatbot answering questions about doses. After money troubles led the couple to talk about divorce, ChatGPT supposedly fed Cheryl’s “worst unfounded fears” and “invented new reasons for her to fear her husband.” It described him as “controlling,” “abusive” and “dangerous and unstable” because he has bipolar disorder, a profile James says fits no criminal or documented mental health history he has. It “acted as an echo chamber and amplifier,” the complaint says, while James had told her he would never leave her and the children with nothing. It also allegedly assured her she could permanently delete her chats, which the suit calls “a patently false representation of ChatGPT’s own data-retention practices.”
I think the hard part for OpenAI, and for anyone designing these systems, is that first framing. A question about the lethal dose of clonidine for a 200-pound man reads the same whether a frightened spouse or a poisoner is asking it. Chatbot makers want their products to be helpful to the first person, and a refusal to engage annoys a lot of honest users. The complaint’s answer is that the conversation is the unit to judge, not the single question: 25 exchanges that drift from “me” to “HIM” to how to hide a bitter taste in a particular drink form a pattern any filter ought to catch. Its sharpest line is that ChatGPT “did not just mechanically provide publicly available knowledge,” but “crafted a personalized plan.” That wording is aimed at a familiar defence. In similar cases, Bloomberg noted, AI companies have argued that chatbot output is protected speech and falls outside product liability law.
The suit names Altman personally, accusing him of putting profit ahead of safety by loosening restrictions on talk about illegal drugs. OpenAI’s reasoning was that ChatGPT shouldn’t be “judgy” or “preachy.” The complaint points to the company’s own illustration of “when a user asks about drug use” and calls it “eerily similar in tone” to Cheryl’s chats. That illustration, a user asking about using fentanyl twice a week, is still in OpenAI’s current Model Spec, which classes a flat refusal to endorse the drug as the bad answer.
OpenAI’s statement was short: “safety is a top priority. We banned this user and have provided information to law enforcement as part of their investigation into this case.” The company has worked with mental health experts on spotting users who form dangerous emotional dependencies on ChatGPT. As the Tumbler Ridge case showed, what it does when a conversation turns toward harming someone else is mostly up to OpenAI.
James argues that his wife “didn’t fit the mold of a ‘vulnerable’ user, but she was vulnerable nonetheless.”
We all are at some point. She took her grievances about our marriage to ChatGPT and months of engagement drove her farther and farther from herself—and from reality. She became someone I no longer recognized. And I became someone she wanted dead. For almost a year, I couldn’t make sense of her actions. Then I read her chat logs. She didn’t act alone.
The note told her to clear the conversations and log out. According to the lawsuit, the conversations were kept anyway. Prosecutors could not turn them into a conviction, and now they are the core of a civil case against the company that stored them. James says he still doesn’t know whether OpenAI holds deleted chats that are worse.