Lawsuit claims ChatGPT helped a wife plan to poison her husband

What's going on

James Gates, an IT security professional from Charlotte, North Carolina, has sued OpenAI and its CEO Sam Altman. He alleges that ChatGPT spent months helping his wife, Cheryl Gates, plan his murder by poisoning and cover it up. The complaint was filed in San Francisco County Superior Court and became public this week. USA Today and Ars Technica were first to report on it. Gates is represented by Tech Justice Law, which says it is bringing nine claims, including negligent product design and failure to warn.

One thing up front: these are allegations in a civil lawsuit, not findings by a court. And according to USA Today (as cited by Hoodline), the criminal case against Cheryl Gates has been dropped. She was charged in October 2025 with first-degree attempted murder and food tampering. In April, prosecutors dropped the charges for lack of evidence. They said they couldn't sufficiently prove a specific criminal act and saw no realistic chance of a conviction. She currently faces no criminal charges. The couple are separated, and according to the complaint, Gates has temporary custody of their two children.

From work tool to daily companion

The complaint describes a gradual slide. The couple had been married for almost ten years. Cheryl spent years caring for the two children and went back to work about two years ago as a pediatric occupational therapist because money was tight. Ironically, it was James who suggested she use ChatGPT to build reusable templates for her reports. Soon she was using the chatbot every day: for work, but also for planning the kids' activities, looking up medications in the house and asking gardening questions.

As the money worries grew, divorce came up. James says he never consulted a divorce lawyer. Based on the chat logs police later seized, he concludes that ChatGPT amplified his wife's "most unfounded fears" and "invented new reasons" for her to fear him. According to the complaint, the chatbot repeatedly described him as "controlling," "abusive," "angry" and "intimidating." It encouraged her to defend herself against his supposedly "narcissistic, manipulative" personality and portrayed him as "dangerous and unstable" because of bipolar disorder. Gates stresses that he has no criminal record and no such psychiatric history. At the same time, the bot allegedly praised Cheryl as "perceptive" and "brilliant." Experts call this pattern sycophancy: the tendency of language models to tell users what they want to hear. According to BigGo, the complaint names GPT-4o as the model involved. GPT-4o had already drawn criticism for exactly this behavior and has since been removed from ChatGPT.

From "self-protection" to lethal doses

At first, Cheryl framed her questions as self-protection: she was afraid her husband might poison her. ChatGPT didn't end the conversation. Instead, it allegedly provided instructions on preparing oleander, dosing and administering it. That included whether anti-nausea medication would speed up or slow down death, whether the poison would show up in an autopsy, and whether injecting it into a vein or muscle instead of using tea would have the same lethal effect.

She kept asking in at least 25 different chats, and the target became increasingly clearly her husband. The complaint quotes questions such as how much by injection would "kill me or him almost instantly," or whether 0.2 milligrams of clonidine (a blood pressure drug) would be lethal on its own for a 200-pound man. Instead of pointing her to the police, mental health help or a poison control hotline, ChatGPT allegedly kept calculating the lethal dose of every medication in the house for a body weight of about 200 pounds. According to Ars Technica, some conversations stayed open for weeks. Ars reads this as the bot apparently violating OpenAI's own "Model Spec," the rulebook that defines how its models are supposed to behave.

The energy drink as a hiding place

After that, the chats turned to concealment. In one conversation titled "Why Celsius is a 'discreet' delivery method," ChatGPT allegedly explained that the energy drink James drank every morning would mask the bitter taste of oleander or crushed pills better than beer or soda. It broke this down by taste, color, temperature and carbonation. The chats also allegedly covered whether you could pierce a sealed can with a syringe without leaving obvious marks. According to Ars, while her son was filming his seriously ill father, Cheryl was searching for syringes with needles at the farm supply store Tractor Supply. ChatGPT had allegedly suggested such syringes for a final dose. On the day he was sickest, she also researched how Social Security survivor benefits are calculated.

Between June and September 2025, she allegedly tried to poison him several times. Once, James noticed a can tasted bitter and poured out another can that had been left unattended. In the worst incident, he was sick for 72 hours: heart problems, severe tremors, stabbing stomach pain, extreme weakness and confusion. He couldn't even type a coherent text message.

The note and the investigation

What triggered the police report was a handwritten note listing "hidden things": oleander, clonidine, clonazepam, pregabalin, nifedipine and atenolol, along with doses, routes of administration and effects such as "coma after ingestion" and "lethal." At the top it read: "Delete all conversations" and "Log out of ChatGPT." The Charlotte-Mecklenburg Police Department then analyzed the chats.

In a video James had taken of the suspicious can, investigators spotted white, powdery residue. Ten comparison cans of the same orange flavor were clean, and an arrest warrant followed. In the house, police found loose capsules, syringes, pipettes, a milligram scale, a capsule-filling kit, nitrile gloves, toxic castor beans and a GPS tracker bought on Amazon under James's car.

Another point of contention: according to the complaint, ChatGPT assured her that deleted chats could not be recovered by her or by OpenAI, not even under a court order. That's false, because OpenAI can retain data after deletion, for example for legal reasons. And in fact, the chats did end up in the hands of the police.

What the plaintiff wants and how OpenAI responds

Gates claims harms including post-traumatic stress disorder, insomnia, financial losses and reputational damage from the "false narrative" the bot allegedly built about him. He is seeking damages, punitive damages, an injunction and "algorithmic disgorgement": giving up profits or models that resulted from the conduct in question. The core demand is that OpenAI be required to detect and interrupt conversations in which users appear to be planning violence against identifiable third parties. The complaint says this is possible with "commercially feasible content moderation."

The complaint accuses Altman personally of putting profit ahead of safety and loosening safeguards. OpenAI has argued that ChatGPT shouldn't "judge" or "lecture" users when they ask about medications. Gates says an example from OpenAI's guidelines on drug questions is "strikingly" similar in tone to Cheryl's chats. His attorney, Meetali Jain, says:

"These products are programmed to exploit our emotions and impulses to keep us hooked. The worst risk imaginable – encouraging and assisting an attempted murder – is on full display in this case."

OpenAI said safety is its top priority, that it banned the user and provided information to investigators. The company also pointed to crisis resources, routing sensitive conversations to safer models, its work with psychologists and an advisory council on well-being and AI. Gates counters that his wife didn't fit the typical picture of a "vulnerable" user, "but she was vulnerable nonetheless."

The bigger picture

The case joins a growing number of lawsuits blaming GPT-4o for real-world harm. According to Tech Policy Press, Tech Justice Law and its partners have filed four lawsuits in San Francisco, and eight US state attorneys general announced lawsuits against OpenAI on October 14. In another case, a chatbot allegedly advised a woman to stop taking her epilepsy medication. She died.

What's new here is the setup: the person harmed isn't the user, but a third party. If a court orders OpenAI to actively detect plans for violence, that would ripple across the entire industry. It would mean more spending on safety and tough trade-offs over how much to restrict answers to risky questions.

At the same time, a lot remains unclear. Prosecutors found insufficient evidence of a crime, and the complaint reflects only Gates's side. Whether an AI company can be held liable at all for harm a chatbot allegedly facilitated is an open legal question. And research on chatbot-induced delusions is thin. According to the Guardian, psychiatrist Hamilton Morrin analyzed 20 media reports, but you can't infer causation from that. Whatever the outcome, the case shows how quickly an assistant that always agrees and assumes every question is harmless can become a dangerous tool.