OpenAI Alerted FBI After Man Used ChatGPT to Plan Violence Against Ex, Logs Helped Build Case

A 25-year-old Florida man’s private conversations with ChatGPT about harming his former girlfriend prompted OpenAI to notify federal authorities. Court records and news reports show those logs later helped local investigators build a criminal case that led to arrest and a negotiated plea.

What Happened

Darren Shida Zhou, a former financial analyst in West Palm Beach, used ChatGPT beginning in March to discuss the end of a six-month relationship with a 21-year-old woman. According to federal records reviewed by USA Today and local reporting, his messages to the chatbot described the woman's routines and people she saw and escalated into explicit threats and detailed planning.

"I'm gonna kill her by the end of this month. If I can't have her, then nobody can."

Escalation And Evidence

Court filings and investigators' summaries say Zhou’s messages included scenarios involving stalking, kidnapping, sexual assault, holding the woman hostage and murdering members of her family. He reportedly discussed waiting in a gym parking lot with flowers and talked about killing himself if rejected. Investigators also say Zhou sent the woman photos of firearms, zip ties and latex gloves; his attorney later told the court Zhou did not actually own firearms.

OpenAI’s safety processes routed the flagged conversations to human reviewers. When reviewers determined there was an imminent threat of serious physical harm, the company referred the matter to the FBI. Federal agents then provided roughly two months of Zhou’s ChatGPT message logs to the Palm Beach County Sheriff's Office. The material reportedly included Zhou's messages to the chatbot but not the chatbot’s replies.

Investigation, Arrest And Charges

After reviewing the logs, a deputy concluded Zhou's statements reflected repeated planning and rehearsal rather than isolated emotional outbursts. Deputies conducted a welfare check on the woman, who said she had ended the relationship because of what she described as Zhou’s "erratic, jealous, and controlling behavior." Investigators say Zhou continued to contact her despite requests to stop and used app-generated phone numbers after being blocked.

Zhou was arrested on May 29 and booked into the Palm Beach County jail on charges that include written or electronic threats to kill or cause bodily harm, aggravated stalking involving a credible threat, and using a two-way communication device to facilitate a felony. He spent two days in custody before posting $100,000 bail.

Plea And Sentence

In June, Assistant State Attorney Ana Cuskova filed three felony charges that together carried a potential combined sentence of up to 25 years. Zhou reached a negotiated plea and on Aug. 13 pleaded guilty to all three counts. Circuit Judge Scott Suskauer accepted the agreement, withheld adjudication and sentenced Zhou to eight years of probation.

• Probation Conditions: Electronic monitoring for the first two years, no contact with the victim, prohibition on possessing firearms, completion of a batterer’s intervention program, mental health evaluation and treatment, and abstention from drugs and alcohol.
• Withholding Adjudication: The court accepted Zhou’s guilty plea but did not enter a formal felony conviction under the terms of the plea agreement.

At sentencing the judge told Zhou, "Get over it and move on. There are other people in the world." Zhou’s attorney highlighted his lack of prior criminal history and academic achievements, including graduating magna cum laude from Northeastern University. Goldman Sachs later terminated Zhou after learning of the case; FINRA records list him as registered with Goldman Sachs & Co. LLC in West Palm Beach from March 2024 until June 2026 in Asset & Wealth Management.

Why This Matters

The case illustrates how content flagged on AI platforms can be escalated to law enforcement: when human reviewers judge a user poses an imminent threat of serious harm, companies like OpenAI may notify authorities. OpenAI’s public safety guidance says it may refer threats against others to law enforcement but generally does not refer self-harm conversations because of privacy concerns.

Local investigators used the ChatGPT logs alongside the victim’s saved messages and other evidence to support the investigation and the plea deal. The matter raises questions about platform safety practices, user privacy, and how online speech can lead to real-world interventions when it signals imminent danger.