The 25-year-old Goldman Elite Was Arrested for Premiting the Murder of His Ex-girlfriend: a Criminal Plan Reported by AI

Introduction

As people are more and more accustomed to telling artificial intelligence what they say about work, health, emotional troubles and even unwillingness to tell people around them, chat records appear more and more and more in judicial cases as a new digital record. It can help the police detect the danger in advance, but it also makes people worried about another problem: when people think they are just talking to a machine, are these chats considered private conversations?

Reporter: Luan Ruoxi
Editor: Wang Shan

This article is the original content of “Sanlian Life Weekly”.

8 Years of Probation

Since March this year, some anonymous harassing messages have popped up on Amber’s mobile phone from time to time. At first, it was some ambiguous sentences, “forgive yourself”. Soon the maliciousness of the content gradually deepened, and it escalated to “cheating slut” and “can’t wait to smell you”. The words began to be insulting, pornographic and threatening. Sometimes the messages sent directly point out her current position, hinting the other party knows her whereabouts.

Amber’s attitude was to ignore it. She guessed that the content might come from her ex-boyfriend Darren Zhou. On the one hand, it was a time node coincidence. These messages happened to appear after they broke up in March. On the other hand, after the account was blacklisted, Darren contacted her with an anonymous number. Amber was also very afraid, but she still hoped that these harassment would stop over time. After all, the two still had some feelings before, and she didn’t want to call the police and get him into trouble. But on May 27, the staff of the Sheriff’s Office in Palm Beach County, Florida, came to the door. Amber was surprised to find that Darren’s emotions not only did not calm down, but also told ChatGPT in detail a murder plan for her and her family. It was this conversation with ChatGPT that attracted the attention of the police.

Judging from his resume, Darren, who is only 25 years old this year, is a typical financial talent. He is studying mathematics and business administration degrees at Northeastern University in the United States. He is proficient in English, Chinese and German. After graduating from college in 2024, he joined Goldman Sachs as a financial analyst. From the second half of 2025 to March this year, Amber dated him for 6 months. In the short relationship, Amber found that his behavior was erratic, jealous and controlling. In February this year, Darren began to violently destroy things in the room and scolded her because he was dissatisfied with the results of the Super Bowl. Worried that he might lose his temper, Amber finally ended the relationship over the phone.

The chat records obtained by the police showed that Darren frequently discussed the relationship with ChatGPT after breaking up, and asked ChatGPT to judge whether Amber’s comment interaction with other men on social media was “ambiguous” and the possibility of getting back together. Slowly, the discussion began to become violent, “If I can’t get her, no one wants to get her”, “I’m going to smash her face”, followed by statements of suicide and killing Amber and his family, and these contents changed from separate statements to a whole set of detailed plans. For example, on April 24, Darren told ChatGPT that he would go to Amber’s place where volleyball played volleyball every Thursday and wait outside the parking lot. When she appeared, he would approach her with flowers in her hand. If she refused, he would take out a gun to kill her and then commit suicide. Later that day and on the 28th and 29th, he repeatedly describe the plan.

According to the Probable Cause Affidavit submitted by the police to the court, which summarizes the preliminary evidence held by the police, in May, the developer of ChatGPT, OpenAI, an American artificial intelligence institution, identified Darren’s repeated violent remarks in the background and issued an emergency disclosure notice to the law enforcement department. The notice said that the user repeatedly made remarks about the intention to attack or murder his ex-girlfriend, “kill her by the end of this month” and “kill her after she comes back from a cruise”, and he claimed to have a gun and was tracking his ex-girlfriend’s schedule in order to know where to find her when deciding to escalate the operation. According to the police, the content of these ChatGPT dialogues is particularly important, which clearly points out the victims, approach, weapons, sequence of action and premeditated sexual violence, showing continuous drills and planning, not an impulsive statement or vague emotional outburst.

The police, combined the content of the ChatGPT conversation and the screenshot of harassment saved by Amber, identified Darren and judged that he had posed a sufficient personal safety threat to Amber, and arrested him in May. Florida Assistant Prosecutor Ana Cuskova charged him with serious stalking, written threats of harm and illegal use of two-way communication equipment, and sentenced to up to 25 years in prison. On August 13, Darren pleaded guilty to all three charges. On the same day, the judge did not make a conviction and sentenced the defendant Darren to an eight-year suspended sentence.

Circuit Court Judge Scott Suskauer once said in court that the assistant prosecutor and Darren reached a plea agreement, and he accepted it because the victim agreed to the agreement. The agreement exempted Darren from imprisonment and a serious conviction. During this period, he must comply with a series of restrictions, such as wearing an ankle monitor for the first two years of probation, not having weapons, and participating in the psychotherapy activities of the abuser intervention project.

Matthew Chris (Matthew T. Christ) is a partner lawyer of Florida Lafferty Domnick Cunningham and RDCY, and previously served as a legal assistant to Florida Supreme Court Justice Barbara Parrient. Chris told this magazine that judging from the final suspended sentence, the victim’s opinion played a great role in it. According to Florida’s Marsy’s Law, the victim has a certain say in similar trial decisions. The victim agrees to the agreement, and Darren has no criminal record. The judge also took these factors into account and decided to suspend the conviction, although the chat records implied that he owned a gun, but did not actually possess a gun.

New Evidence

“I can’t believe it’s true at all.” Mitchel told this magazine that he and Darren had known each other since high school for nearly 10 years. Mitchel said that Darren’s hometown is in Massachusetts, northeastern United States. In his impression, Darren is a very talkative person who likes to joke, is positive and loves sports, plays tennis and rugby, and pays great attention to body management. He often posts photos of himself going to the gym to exercise on social media. He likes small animals and will take photos of kittens he meets on the street corner. But when we met last year, he found that Darren had changed a lot. First of all, he was obviously much fatter than before. In addition, his personality seemed to have changed a little. He hadn’t seen him for a long time. He just kept talking about himself, saying how much he loved his work, how powerful and advanced his position was. He often went to music festivals and participated in many pai. Right. But whenever asked very specific questions, such as how the work content is, how the living environment is, he always answers “Good” and “Fun”, and ends the conversation with one or two words, leaving only an awkward silence.

Although everyone went their separate ways after college, they also maintained contact with each other on social media and liked and commented on each other’s content. Darren often posted photos with Amber on social media. “The two people are very close in the photo, hugging and cuddling with each other. They look like a very sweet couple.” Mitchel said that what surprised him most about the whole thing was that Darren used to be very friendly to women, and he also thought that his boyfriend interfered too much with his girlfriend’s interaction with others on social media, which was a very controlling thing. He had always opposed to men doing this, thinking that this kind of behavior was “toxic”, “I didn’t know that he was What will become like this today?”

In the process of reviewing Darren’s case, one of the details that Chris paid special attention to — the role played by ChatGPT in it. He told this magazine that the dialogue between Darren and ChatGPT played a crucial role in the case. First of all, the victim has no idea of reporting the case, which means that it is impossible to trigger the case investigation without the content of ChatGPT’s dialogue. Moreover, there is no direct evidence that the anonymous text message received by the victim is from Darren. The content of ChatGPT helps to establish a connection and reduces the difficulty of the case investigation. Moreover, according to Florida law, the crime of serious stalking needs to meet the needs of repeated stalking, credible threats, to give the victim a reasonable fear and show obvious ability to implement, the dialogue content of ChatGPT shows that Darren has been repeatedly tracked, and the threat behavior has gradually escalated in the past two months.

According to Chris’s observation, about two years ago, the chatbot’s conversation records began to be cited in the search warrant and affidavit. As a lawyer, he also began to ask the person concerned if he had a conversation with an artificial intelligence chatbot on a certain topic when handling the case. Chris said that 30 years ago, lawyers investigating cases might look for letters, business records and witness testimony. After that, emails, text messages and a large number of social media began to appear. As people’s lives went further on the Internet, the lawsuit could cover more information, and as people turned to artificial intelligence more and more. Chatbots can seek job, health and emotional advice, and these conversation records with chatbots are also appearing more and more frequently in court as evidence. These conversation records are often regarded as ordinary electronic records, rather than privileged communication content, and may be submitted through civil evidence disclosure, subpoens, court orders, searches of suspects’ equipment, or transfers by artificial intelligence companies.

A statistic from the Washington Post found that at least 12 court cases in the past two years have cited chatbot transcripts. “We found that artificial intelligence chatbots provide a false sense of intimacy, privacy and security. Most people tend to believe that their conversations here will never be read by another person, so they often reveal more of their true motives here,” Chris told this magazine. In legal practice, the dialogue records of chatbots are often used to prove the intentions of the parties and whether the testimony before and after are contradictory. For example, in a traffic accident, if the party causes a traffic accident and causes injury to others, he may claim to his insurance company, lawyer and the lawyer of the injured party that the accident was caused by the other party running a red light. At the same time, on the day of the accident, the person concerned may mention in the chat record of the artificial intelligence chat robot, “I ran a red light today and caused an accident. What kind of responsibility will I bear? What kind of compensation may you face?” At this time, this is legally equivalent to “self-recognition of the parties”.

Chris said that compared with traditional digital evidence, chatbot dialogue records provide a new perspective. Internet search records may only be a few words, and there can be a variety of room for interpretation, but dialogue records can see the whole thinking of the other party through the prompt information provided by the person and the interaction between questions and answers, process and behavioral intention. But not all chat records are used as evidence.

Chris said that, for example, chatbot conversations with negative and violent tendencies may represent accidental emotional catharsis, momentary curiosity or role-playing. In recent years, artificial intelligence robot chat records that can really be used as evidence of credible threats should consider the duration of speech, the ambiguity or detail of language, whether the negative degree has escalated, whether there is the ability to complete the threat, etc., and the important point of judging whether the dialogue content is “transboundary” — whether there is actual action. “Take the Darren case as an example, if there is only the dialogue content of ChatGPT, and he did not actually send any harassing messages to the other party, then it is not enough to prosecute with the content of the chat as evidence,” Chris said, but that doesn’t mean nothing can be done. Given the seriousness of the speech, the police can apply to the court for a mental health assessment or a personal restraining order.

Transparency

The Darren case is not the first time that OpenAI has taken the initiative to submit the worrying ChatGPT chat content to law enforcement. In June this year, OpenAI also issued a warning to Brazilian law enforcement that a 36-year-old Brazilian man spent two months on ChatGPT discussing how to kill his 8-year-old child and then commit suicide to avoid paying alimony. Eventually, the man was required to wear an electronic ankle monitor and was prohibited from touching the boy and his mother. According to the data released by OpenAI, in the second half of 2025 alone, the company disclosed more than 80 accounts, an increase of more than four times over the same period last year.

However, not all potential dangers have been identified and warned in advance. In February this year, a school shooting occurred in British Columbia, Canada. It was later discovered that the gunman had entered violent prompts on the ChatGPT platform. According to a court document, in August 2025, OpenAI identified the gunman’s violent remarks, marked as “a credible and specific threat to gun violence against real people”, and then only blocked the gunman’s account. OpenAI explained that it did not report to the law enforcement department in time because there was no evidence of the impending attack.

“In fact, the comparison between the Florida case and the shooting in Canada highlights why it is so difficult to make such reports.” Dimitris Bilo, the founder of AGC for Africa, an artificial intelligence governance and compliance consulting company, told this magazine that compared with a few years ago, the existing artificial intelligence risk warning mechanism has made some progress in identifying violence and criminal premeditation, which can identify obviously dangerous content, reject harmful requests, and formulate escalation processing processes for serious threats, but there are still certain limitations, “because detecting words and understanding dangers are not the same thing.”

Bilo said, for example, if someone enters “I want to kill my boss”, the artificial intelligence system can easily identify that the sentence contains violent language, but the system may not be able to judge whether the person is making a real threat or expressing temporary depression after a hard day and is in a mental health crisis. Or maybe he is just writing a novel, quoting movie lines, and discussing a hypothetical situation. Words themselves are easy to detect, but it is much more difficult to determine the meaning of these words in the real world. On the contrary, some really dangerous situations may not contain obvious warning words at all. Users may discuss goals, locations, schedules or preparations in indirect language, which seem relatively harmless to look at their alone. Risks will only appear when these dialogues are seen as patterns for a period of time. Artificial intelligence systems usually can only obtain the dialogue itself, and cannot understand the broader context that helps to judge whether the threat is real. These ultimately require subjective judgment, and even experienced manual investigators are still difficult to deal with.

In the view of Marina Linde, a legal adviser and artificial intelligence ethics expert of the non-profit organization “AI for Change”, if enterprises report every disturbing conversation, it will not only bring overload risks to law enforcement, but also destroy users’ trust, fear that privacy will be leaked, and cause people to be afraid to ask for help when they are in trouble. Eric Goldman, a professor at the University of Santa Clara Law School and co-director of the Institute of High-tech Law, told this magazine that in the United States, these chat records are now only regarded as general data evidence, but what should be seen is people’s interaction with artificial intelligence chatbots. There are essential differences between the way, the degree of trust and the technical relationship formed by it compared with other technologies. “In my opinion, people are using artificial intelligence chatbots to solve problems in life. The role it plays is a bit similar to that of a psychological counselor, and the dialogue between patients and psychological counselors is subject to the law. Protected.”

According to OpenAI’s policy, the company may disclose user data to law enforcement agencies or government agencies, provided that such disclosure is considered necessary to prevent emergencies that could endanger life or cause serious personal injury. Anthropic’s reporting system also has a similar statement, claiming that information about customers will not be disclosed at the request of the government unless it is in accordance with valid legal procedures, such as a validly issued summons or search warrants. However, if the company believes that there is an emergency that may lead to imminent personal injury or death, and the timely provision of user information may avoid the emergency, it can make an exception. Google’s Gemini’s privacy policy is that in order to comply with legal or regulatory requirements, companies may be required or based on good faith to share users’ personal information with law enforcement agencies, government agencies, regulatory agencies or third parties at its sole discretion when deemed necessary.

Linde believes that the existing reporting systems of various companies are not transparent enough at present. Most artificial intelligence companies will explain that they may review the content of the conversation under special circumstances and report serious threats to the relevant departments, but users often do not know how these decisions are made and do not know how often such situations occur. They should at least give a general statement about which situations may trigger manual review, reporting or referral to law enforcement, and should be more transparent about reporting standards for different risks. Improving transparency helps to build trust. The reason why people are worried about the disclosure of personal privacy and possible criminal premeditation on the platform is often because they do not know what content will be monitored, when manual reviewers will be involved, and under what conditions information may be shared. More clearly Minghe’s regular transparency reports can reduce this uncertainty.

Amber and Mitchell are pseidonyms in the text.