The ChatGPT Conversations About Killing His Girlfriend
In May 2026, Darren Zhou, 25, of South Palm Beach, Florida, repeatedly typed violent plans about his ex-girlfriend into ChatGPT.
The conversations included lines like "I'm going to kill her this month," along with drafts of threatening messages.
At the time, Zhou was working as an analyst at Goldman Sachs.
Reported by OpenAI, Arrested Within Two Days
OpenAI flagged the conversation as a warning sign.
The company reported him directly to the FBI, and Zhou was arrested in May.
Two days after the arrest, he was released after posting $100,000 bail.
Goldman Sachs learned of it in early June, and Zhou was fired.
How the Canadian Mass Shooting Changed the Reporting Standard
Behind OpenAI's quick action was the mass shooting in Tumbler Ridge, British Columbia, Canada, on February 12, 2026.
Eight people died and 27 were injured.
According to reports, OpenAI had already banned the shooter's ChatGPT account in June 2025 over "disturbing content."
Twelve employees at the company asked that police be notified, but the company didn't report, saying it wasn't an "imminent and credible risk of physical harm."
After the tragedy, on February 27, 2026, OpenAI said it had introduced "an enhanced law enforcement referral protocol" that assesses risk with input from mental health, behavioral, and law enforcement experts.
The Zhou case is one of the early cases where this new protocol applied.
The British Columbia government is preparing a lawsuit against OpenAI, and victims' families have separately sued OpenAI and Sam Altman in US court.
| Date (KST equivalent) | Event |
|---|---|
| June 2025 | OpenAI bans the Tumbler Ridge shooter's ChatGPT account over "disturbing content" (no police report) |
| February 12, 2026 | Tumbler Ridge, Canada mass shooting: 8 dead, 27 injured |
| February 27, 2026 | OpenAI announces enhanced law enforcement referral protocol |
| May 2026 | Darren Zhou discusses killing his ex with ChatGPT; OpenAI reports him to the FBI; arrested and released on $100,000 bail |
| Early June 2026 | Goldman Sachs learns of the case and fires Zhou |
| August 13, 2026 | Zhou pleads guilty to all three charges: conditional sentence instead of prison |
How He Avoided Prison for an Ankle Monitor
On August 13, through a plea deal with prosecutors, Zhou pleaded guilty to all three charges.
In exchange for avoiding up to 25 years in prison and a felony record, he accepted these conditions:
- Wearing an ankle monitor for two years
- Completing a batterer intervention program
- A mental health evaluation
- No contact with the victim
- No weapons; no drugs or alcohol
According to local media reports, the case was first reported by the Palm Beach Post and was later picked up by AOL, Yahoo News, and others.
Can You Escape to a Local LLM?
The case also spread on X (Twitter).
Japanese AI researcher うみゆき (@umiyuki_ai) cited the case and wrote, "If you consult ChatGPT about anything illegal, even as a joke, you really will get reported. Ask a local LLM about that stuff." The post has more than 250,000 views (as of 11:02 PM KST on August 16, 2026).
The point is that cloud-based AI services may monitor conversations under their terms of use and notify law enforcement when warning signs appear.
An open-source LLM running locally has no such monitoring structure at all.
Still, it's worth separating the two: this isn't so much a way to avoid surveillance as it may cut you off from intervention paths (counseling, reporting) that could actually help in a crisis.
The Exact Gap Between the Report and the Arrest Isn't Public
It hasn't been confirmed in official records exactly when OpenAI began monitoring Zhou's account, or the exact time between the FBI report and the arrest.
OpenAI also hasn't separately confirmed an official comment on this case. Reporting so far rests on court records and the Palm Beach Post's reporting.