When a Break Must Be Made, Hesitation Breeds Chaos

Li Zhide

A recent case in Florida suddenly brought together artificial intelligence, love, violence, privacy, and public safety. Darren Zhou, 25, continued contacting his former girlfriend by phone, text message, and social media after she ended their relationship of about six months. At the same time, in extended conversations with ChatGPT, he discussed plans to stalk, kidnap, sexually assault, and kill her, and then take his own life. After OpenAI detected troubling content, the conversations were escalated for review, and the company eventually alerted the FBI. The FBI passed the information to authorities in Palm Beach County, who contacted the woman and opened an investigation. Zhou was arrested in May. On August 13, he pleaded guilty to three charges: aggravated stalking, written threats to kill, and unlawful use of a communications device. The judge withheld a formal adjudication of guilt and sentenced him to eight years of probation, with electronic monitoring during the first two years. Judge Scott Suskauer said one important reason he accepted the plea agreement was that the victim herself had agreed to it.

Some headlines reduced the case to something like this: “Man whispers secrets to ChatGPT; AI calls police and gets him arrested.” It is catchy, but not quite accurate. ChatGPT did not hear one dangerous sentence and immediately place a call to the FBI like a robot in a movie. OpenAI has explained publicly that its current process begins with automated systems identifying content that may involve serious violence. Trained reviewers then examine the material in context. Only when a threat is judged to be imminent, credible, and capable of causing serious real-world harm to another person may the matter be reported to law enforcement. In other words, this was not “a robot calling the police.” It was a chain of responsibility: a machine detected signals, human beings assessed the risk, law enforcement investigated, and a court ultimately decided how the case should be handled.

But to reduce the question to whether ChatGPT has the right to alert police would still miss the larger meaning of the case. What makes it worth thinking about is that an old human problem has reappeared in the age of artificial intelligence. There is an old Chinese saying: “When a decision must be made but is not, disorder follows.” Versions of this idea appear in records going back at least to the Records of the Grand Historian and the Book of Han. More than two thousand years have passed. Human relationships have changed. Communications technology has changed. Artificial intelligence has even entered private life. Yet some human difficulties remain almost unchanged: a relationship is over, but one person refuses to accept that it is over; warning signs appear, but the people involved and those around them hope the danger will simply fade away. A problem that might once have ended is allowed to grow into something much larger.

Start with the relationship itself. According to police materials, the woman did not fail to break up with Zhou. Quite the opposite. She had already ended the relationship, and because she was worried about how he might react, she chose to do so by phone. Zhou continued contacting her afterward, even though she repeatedly told him to stop. The material later reviewed by police showed an increasingly clear pattern connecting phone calls, text messages, numbers generated through different apps, social media activity, and violent statements made to ChatGPT. So the idea that “what should have been cut off was not cut off” must not be used to blame the victim. She was already trying to end the relationship. The person refusing to accept the break was the other party.

That distinction matters. In everyday life, people often imagine that a breakup is something two people must somehow complete together, as though the relationship is not really over until both sides agree. That is not true. A romantic relationship requires the consent of two people. Ending one requires the decision of only one. A person may try to win someone back. He may grieve, lose sleep, regret what happened, or remain unable to move on for a long time. But he does not have the right to demand that another person continue loving him. Love exists only through mutual consent. It is not property that one person acquires and then owns forever.

That is why statements such as “If I cannot have her, no one else can” are dangerous for reasons that go beyond anger. They reveal a way of thinking rooted in possession. The other person is no longer seen as someone with an independent will, but as someone who “belongs to me.” Once love changes from “I hope you love me” to “you must belong to me,” love has effectively ended and control has begun. When control develops into stalking, threats, fantasies involving weapons, and rehearsals of violence, the matter is no longer merely a private romantic dispute. It has entered the realm of public safety.

This is also why it is too easy to ask the woman, “If you were so frightened, why didn’t you call the police earlier?” Researchers who study intimate-partner violence have long recognized that separation does not always end danger immediately. In some cases, it can be the period when danger increases. Research collected by the U.S. Department of Justice has found that controlling partners may continue stalking or attacking former partners after separation, while data from the Centers for Disease Control and Prevention show that current or former male intimate partners account for a significant share of perpetrators in homicides of women. To an outsider, “just call the police” is a simple sentence. For someone who knows the other person’s temper, fears provoking him, and may still carry emotional attachment to a relationship that has just ended, the decision can be far more difficult.

The lesson, then, is not that victims should be told, “You should have been more decisive earlier.” The more important lesson is that society must learn to recognize earlier the signs that a boundary has already been crossed. Repeated harassment is not devotion. Changing phone numbers in order to continue unwanted contact is not persistence. Threats are not the language of love. Stalking is not an attempt at reconciliation. Sending pictures of weapons to frighten someone cannot be brushed aside as mere “emotional instability.” Research on stalking has long warned law enforcement against viewing such signs one by one. They need to be understood as parts of a pattern that may be escalating.

This case also shows one new capacity that artificial intelligence may add to risk detection: it can see a sequence over time that an ordinary person may never see. A friend hears one furious remark and may dismiss it as something said in anger. A police officer sees one text message and may have difficulty judging how serious it is. But if a conversational system sees the same person over several weeks repeatedly mention the same target, locations, timing, weapons, stalking, and ideas of harm, it is no longer looking at a single sentence. It is seeing an increasingly coherent trajectory. When OpenAI described its safety procedures in 2026, it made precisely this point: an individual message may not appear dangerous on its own, while patterns that emerge over a longer conversation, or even across conversations, may reveal a more serious level of risk.

This may be the real reason the case deserves a place in the history of artificial intelligence. In the past, many dangerous thoughts remained inside a person’s mind, in a diary, or scattered across fragments of conversation. Today, more and more people use AI as a secretary, adviser, friend, emotional dumping ground, or even a kind of electronic confessional. That creates a question that did not exist in this form before: How private should a conversation between a person and an AI be?

The answer cannot simply be “completely private,” but neither can it be “the platform may inspect whatever it wants.” If someone says, “I hate him so much I could kill him,” and that alone triggers a police report, we will quickly create a suffocating digital society. Many things people say in anger, grief, or fantasy do not amount to an intention to act. If every dark thought becomes a police matter, people may become afraid to speak honestly to therapists, doctors, friends, or even an AI. A safety mechanism could then end up undermining safety itself.

But the other side of the problem cannot be ignored either. If a conversation moves step by step from emotional venting toward a named target, a real location, a schedule, weapons, stalking, and preparations for action, and the platform still insists, “This is private user information and none of our business,” then privacy may become an excuse for doing nothing. OpenAI’s stated approach tries to draw the line here: automated systems identify signals, human reviewers examine them, and the question then becomes whether there is an imminent and credible risk of real-world harm to another person. The boundary is far from perfect, but it gets at the central issue. What requires intervention is not the existence of a bad thought by itself. It is the point at which a bad thought begins moving toward real action.

For that reason, this case should not become an excuse for unlimited expansion of AI surveillance. A platform is not the police, and an algorithm is certainly not a court. Risk-detection systems can make mistakes. Language contains irony, fiction, role-playing, angry venting, and differences of context. A responsible system must preserve human review, set a high threshold for real-world danger, minimize the disclosure of irrelevant information, and leave actual investigation and legal judgment to law enforcement and the courts. AI may raise an alarm. It must not thereby acquire the power to declare someone guilty.

This is another form of knowing when to cut something off in a modern society. What should be interrupted is the chain of escalating risk, not ordinary private expression. What should be cut is the path from fantasy to real violence, not every zone of privacy between human beings and technology. If the threshold for intervention is too loose, genuine danger may be missed. If it is too tight, we may build a society in which no one ever feels unwatched. The hardest problem of governance in the age of artificial intelligence is not choosing either “security” or “freedom.” It is finding, before harm becomes irreversible, a point of intervention supported by evidence, bounded by clear thresholds, subject to human judgment, and open to oversight.

The case of Darren Zhou also raises another question that can easily be obscured by the spectacle of “AI calling the police.” How does a well-educated young man working at a major financial institution reach this point after a breakup? Education, income, and professional ability clearly do not automatically teach a person how to accept rejection. Someone may master complex financial models and make it to Wall Street, yet never learn one of the simplest and hardest lessons in life: another person has the right to leave you. Court materials indicated that his defense described his condition at the time as a serious mental-health crisis. But psychological distress should not become a synonym for violent threats, nor can it erase responsibility for specific conduct. The overwhelming majority of people who experience mental distress do not harm others. What still needs to be discussed are the actions themselves: control, harassment, threats, and repeated violations of another person’s boundaries.

This may be one of the most underestimated problems in education today. We spend years teaching young people how to succeed, but rarely teach them seriously how to fail. We teach them how to compete, but seldom how to withdraw. We teach them how to win love, but rarely how to accept that love has ended. Yet no human life consists only of gaining. Jobs are lost. Investments fail. Friends leave. Parents die. Love ends. A mature person is not someone who always gets what he wants. It is someone who, even when he does not get what he wants, still knows where the line is that he must not cross.

In this sense, “cutting off” is not cruelty. It is an acceptance of reality. If a relationship has ended, let it end. If contact has been rejected, do not begin again from another phone number. If danger signs have already appeared, do not keep soothing yourself with “he is just talking” or “he will calm down in a few days.” Individuals need this ability. Families need it. Companies and governments need it as well. Looking back, many disasters were not preceded by a total absence of warning. Instead, one possible stopping point after another was allowed to pass. Small problems became far more costly because people were unwilling to let go, afraid to stop, or unwilling to admit that an earlier judgment had been wrong.

But the principle of “cutting off when necessary” has another half that is often forgotten: we must know exactly what should be cut off. A person should not be treated as a criminal merely because he says a few vicious things after a breakup. The ability of AI to analyze conversations should not give platforms an unlimited right to monitor people. Nor should one case in which a report may have prevented violence lead us to declare that privacy must henceforth give way unconditionally to security. A mature society does not simply swing a knife at every knot. It learns to recognize which line has actually been crossed before the damage becomes irreversible.

Darren Zhou ultimately did not carry out the most horrifying acts described in the court materials. We cannot prove that he certainly would have committed violence if OpenAI had not reported him. Nor can we prove that his plans would always have remained words. This is precisely what makes risk governance difficult: action must sometimes be considered before the ending is known. What police eventually had was not only language from ChatGPT but also a pattern of continued contact, harassment, and threats in the real world. This was therefore not a case in which one angry sentence caused an arrest. It was a case in which digital traces and real-world behavior gradually converged into a risk pattern that law enforcement had reason to investigate.

More than two thousand years ago, people warned that when a necessary break is not made, disorder follows. The age of artificial intelligence adds a new layer to that old insight. Between people, what must be broken is the claim of possession over a relationship that has already ended. In the face of violent risk, what must be interrupted is the chain that leads from threat to action. Technology platforms must cut off AI assistance that contributes to real-world harm. At the same time, public power must hold another line and refuse to let “security” become a passport to unlimited surveillance.

So what deserves to be remembered about this case is not that “ChatGPT informed on its owner.”

What matters more is that, in a society where more and more human activity leaves a digital trace, we may now have the ability to see certain chains of risk before some tragedies occur. The question is no longer simply whether machines can see them. It is whether human beings have the judgment to decide when to keep watching and when to intervene; when to respect one person’s secrets and when another person’s life must take priority.

The real meaning of “when a break must be made, hesitation breeds chaos” is not impulsiveness. It is a sense of boundaries. It is judgment. And it is a society’s ability to say, before something becomes irreversible:

This goes no further.