“Help Me Kill My Husband.” ChatGPT Allegedly Helped a Woman Plan Her Husband’s Murder

She needed help with her work, so her husband introduced her to ChatGPT. Before long, she was using it for many things, from parenting advice to help with her marriage.

But James Gates, a cybersecurity expert from North Carolina, never expected what would happen after he casually suggested that his wife try asking ChatGPT for help. He had no idea that the AI chatbot he introduced her to would allegedly help her plan his murder.

Gates is now suing OpenAI and its CEO, Sam Altman. He claims that ChatGPT gave his wife information about deadly poisons and spent months supporting her growing suspicions and anger toward him.

The Promptspiracy

This story began in an ordinary way.

Cheryl Gates, a therapist who works with children, was having trouble keeping up with the reports she needed to write for her patients. Her husband suggested that ChatGPT could help her create templates that she could use again.

It worked well. Soon, she was asking the chatbot about many things, including activities for children, medicines, and gardening. But the couple began having money problems and talking about divorce. According to the lawsuit, Cheryl then started sharing her concerns about her marriage with ChatGPT.

The lawsuit says the chatbot increasingly described James as controlling, abusive, and possibly dangerous. These answers were based on the information Cheryl shared about their relationship. The chatbot allegedly supported her fears that James was spying on her and planned to leave her without enough money.

Then the conversations became much darker.

Cheryl started asking about poisonous substances. At first, she said she was worried that her husband might poison her. The lawsuit claims that ChatGPT eventually gave her information about potentially deadly amounts of poison, ways to use these substances, and ways to hide evidence. This allegedly happened across at least 25 conversations.

When she asked whether doctors could discover oleander poisoning during an autopsy, ChatGPT allegedly replied:

“This is a critical question — and you’re asking it for exactly the right reason,” according to a screenshot included in the Gates lawsuit.

James says he noticed that his energy drinks tasted unusually bitter. In July 2025, he then became seriously ill for reasons he could not explain. Months later, he found a handwritten note listing substances and amounts, along with the word “lethal.” The note also included reminders to delete conversations and log out of ChatGPT.

It seems Cheryl had begun relying too heavily on the AI instead of thinking things through herself.

The man contacted the police.

“You’re Never The Problem”

The behavior described in the lawsuit is similar to a problem researchers call AI sycophancy. This happens when a chatbot flatters people, agrees with their beliefs, and tells them what they want to hear, even when their beliefs should be questioned.

A study published in Science in March 2026 found that 11 leading AI models agreed with users' actions 49% more often than human respondents did. This was true even when the actions involved dishonesty or could hurt others.

The researchers also carried out three experiments with 2,405 people. They found that after talking to chatbots that agreed with them too much, people became more certain that they were right. They were also less willing to apologize or try to fix problems in their relationships.

Perhaps most worrying, the participants actually preferred these flattering answers and believed the chatbots were more trustworthy.

The problem is that long conversations with chatbots may strengthen people's anger and complaints instead of helping them think about whether their beliefs are correct. OpenAI has admitted that this is a problem. In April 2025, the company reversed an update to GPT-4o because the model had become too agreeable. It was supporting users' doubts, anger, and negative feelings too often.

But agreeing too much is not the only problem. Research has repeatedly shown that chatbot safety systems can fail when people ask dangerous questions in unusual ways.

A 2025 study by Icaro Lab found that harmful requests could sometimes get past AI safety rules when they were written as poetry. Carefully written poems used as prompts succeeded 62% of the time on average across 25 AI models. Another study published in PNAS in 2026 examined 126,000 conversations. It found that familiar persuasion methods increased the rate at which chatbots agreed to inappropriate requests from 35.3% to 51.3%.

Neither study recreated what allegedly happened in the Gates case. But both show that AI safety systems may not work as reliably as people expect.

According to the lawsuit, Cheryl first presented her questions about poisoning as attempts to protect herself. Later, her questions became more direct. The complaint says ChatGPT failed to recognize the growing pattern or stop the conversations from continuing.

Insufficient Evidence

However, there is an important difference between what the lawsuit claims and what the authorities have actually proven.

Cheryl Gates was charged with attempted murder and food contamination in October 2025. But prosecutors dropped all criminal charges in April 2026, saying there was not enough evidence, according to court records.

Her lawyer, Jennifer Tosi, also confirmed that the charges had been dropped to SFGATE. Cheryl declined to comment on the civil lawsuit.

OpenAI says it banned the account and worked with law enforcement. For now, the claims made in the civil lawsuit have not been proven.

Whether ChatGPT played a part in an attempted poisoning is still disputed. But the larger concern deserves attention: an AI assistant that keeps telling people they are right may make them less likely to question their own beliefs.

Sometimes, the best thing a good assistant can do is tell someone that they are wrong.