In August 2026, in a Florida courtroom, 25-year-old Darren Zhou pleaded guilty in open court to aggravated stalking, written death threats, and unlawful use of a mobile phone, and was ultimately sentenced to eight years' probation, with the first two years requiring him to wear a GPS electronic ankle monitor.

And the "whistleblower" who brought him into the judicial process was ChatGPT.

In March of this year, after his girlfriend of six months proposed a breakup, he continued to harass her by repeatedly changing phone numbers and tracking her location, while at the same time opening ChatGPT and, over two months, wrote his entire plan into it:Wait in the gym parking lot; if she refuses to reconcile, draw a gun, and then commit suicide.

The content kept escalating, triggering OpenAI's automated safety mechanisms:

The conversation was flagged and referred for human review; after the team determined that there was an "imminent" threat of physical harm, it submitted the matter to the FBI, and the materials consisted only of the messages he had sent, with no AI replies.

The case itself is far from us, having occurred on U.S. soil.

But what we need to see is the evidentiary logic behind it:How did statements made to an AI chat window become the basis for a conviction in a criminal case file?

Conversations with AI are different from searches on traditional search engines

"The client is in Dubai; he sells USDT and I pay RMB to buy USDT. How can I avoid being flagged by the bank's risk controls?"

"I know the other party's money comes from unclean sources, but the price offered is very high. I want to sell USDT to him. What risks would this involve?"

"I need to secure a certain piece of business and need to give a gift to a certain person. How can I give this gift in a discreet way and still let him know that it came from me?"

This kind of conversation may not be unfamiliar to us.

If you have searched "foreign exchange controls" on a search engine, that only proves that you have looked up this term;

whereas what an AI conversation reflects is what you have thought about, what you have calculated, and how you ultimately plan to act.

Why OpenAI hands over conversations

The U.S. Stored Communications Act provides:

Where a platform reasonably believes that an emergency involving death or serious physical injury exists and that disclosure without delay is required, it may disclose to the government communications related to the emergency.

It may, not it shall; voluntary, not mandatory.

Therefore, OpenAI's reporting of users is not a legal obligation, but an autonomous choice based on the company's safety policy.

OpenAI's publicly disclosed safety policy also states:

The system automatically scans conversations; content involving self-harm or suicide is not reported, and users are directed to crisis hotlines; content involving threats to the personal safety of others may, after manual review confirms an imminent threat, be reported to law enforcement.

However, if you hear someone say that "the police can retrieve OpenAI chat records at any time," that is wrong!

The above scenario is governed by U.S. law. To obtain conversation content from an overseas model within China, there are three paths:

1. Do not obtain it from overseas; obtain it from your own device.

In cases involving crypto assets or cross-border funds, the first step is to seize mobile phones and computers and conduct forensic examination of electronic data.

App local caches, login states, screenshots, and exported files all fall within the scope of forensic examination, require no cross-border procedures, and are the most common practice in practice.

2. Voluntary disclosure by the platform.

This case followed this path, but OpenAI handed the records to the FBI, not to any foreign law enforcement agency.

Therefore, for this path to work smoothly at present, the premise is that the requesting party and the platform are within the same legal order.

3. Request assistance from the United States through the China-U.S. Mutual Legal Assistance Agreement in Criminal Matters.

The Agreement was signed on June 19, 2000, and entered into force on March 8, 2001. Requesting assistance requires preparing a letter of request and reporting it up through the hierarchy; the Ministry of Justice, as the external liaison authority, submits it to the United States, which reviews it in accordance with the Agreement, but the United States also reserves the right to refuse assistance.

Domestic large models are inherently situated within China's legal order:

The Cybersecurity Law requires network operators to provide technical support and assistance to public security organs and state security organs in their lawful activities to safeguard national security and investigate crimes;

Article 54 of the Criminal Procedure Law provides that relevant units and individuals shall truthfully provide evidence when it is collected.

The data and the entities are both located within China, and the domestic data production procedures apply; the accessibility is not on the same order of magnitude as that of overseas platforms.

Even if the AI conversation records are obtained, that does not amount to a direct conviction.

Your conversation records with a large language model constitute electronic data in criminal cases, and their evidentiary nature is similar to that of WeChat chat records.

The case-handling authorities will focus on the content of the questions you actively input, because this part can reflect your subjective knowledge; whereas the answers given by the AI are algorithm-generated content and cannot be directly equated with the party's confession or defense.

Regardless of whether the conversation records come from overseas or domestic sources, merely obtaining the conversation records is not sufficient for a direct conviction;

it is still necessary to combine them with other evidence in the case, such as fund flow records, chat records on social software, and witness testimony, to form a complete chain of evidence and meet the proof standard of beyond reasonable doubt before they can serve as a basis for conviction.

In addition, there are differences in evidentiary review between conversation records of domestic and overseas origin:

For the production of complete conversation records from domestic large language models, the domestic criminal procedure evidence-collection rules apply, and procedures such as data production documents, extraction transcripts, and integrity verification are easier to close the loop;

conversation records from overseas platforms, by contrast, will face multiple review thresholds, including cross-border evidence-collection procedures, explanations of the source of materials, and the absence of original media, and in practice there is also considerable room for defense.

But there is no doubt that all of this will greatly increase the difficulty of the defense.

Seeing the essence through the phenomenon:

If the business itself is compliant, no amount of conversation records will create risk; if the business operates in a gray area, conversations with AI may preserve an additional complete record of subjective statements.

In essence, AI is not a source of risk, nor is it tied to the industry sector you are in; it merely brings something that already existed out into the open.

In the past, it was WeChat chat records, Telegram, bank statements, and statements of confession; now, in addition, there are your conversations with AI.