Florida woman's Claude 'diary' entry led police to her door

Per the report, Anthropic scans for key phrases and threatening content. These messages were serious enough to be passed to a human review team, and that team got in touch with police. Deputies went to Heller's home, took her into custody without any trouble, and handed the case to an intelligence detective.

She was charged on September 30 under Florida Statute 836.10, which covers written or electronic threats to kill, injure, or carry out a mass shooting or act of terrorism. It's a second-degree felony. Gizmodo notes that in Florida that can mean up to 15 years in prison and a $10,000 fine. Tom's Hardware, which checked the court records, reports that her arraignment is set for November 2 and that a public defender has been appointed, with no attorney yet named on the docket. She hasn't been convicted, and what police describe is still an allegation.

Sheriff Carmine Marceno used the arrest as a warning. "Users need to understand that you are never truly anonymous," he said.

Anthropic hasn't commented publicly on this case, according to The Next Web and Tom's Hardware.

Can a chat with software count as a threat?

Much of the coverage treats the arrest as the end of the story. It may be just the start, because of how Florida's law is worded. Section 836.10 doesn't make it a crime to write down a violent thought. It makes it a crime to "send, post, or transmit" a threat in a writing or electronic record in a way that "may be viewed by another person." A paper diary in a drawer wouldn't qualify. A post on Facebook clearly would. A chat with an AI model sits somewhere in between, and that's the gap Heller's lawyers could aim at.

You can see both sides. A defense lawyer could argue she was talking to software, not to a person, and that she had no reason to think a human would read it. That fits her own description of Claude as a diary. A prosecutor could answer that the messages were sent over the internet to a company's servers, that the company's terms openly say staff may review and report content, and that a person at Anthropic did in fact read them. The second message, the one about getting a new gun, will likely matter too. It makes the entries read less like venting and more like a plan, at least as police describe them.

Lawyers who have weighed in so far sound cautious. Legal analyst and attorney Michael Raheb told The Mary Sue that proving cases like this can be a "very slippery slope," noting that the First Amendment protects speech only until it crosses into true threats or incitement. Heller's own side hasn't spoken publicly; her public defender isn't named in the court record.

The statute's text doesn't spell out whether software counts as the kind of audience it has in mind. In our searches, we didn't turn up a published Florida appeals ruling that settles whether a chatbot conversation falls under 836.10, and we're not predicting how a judge will rule. But this question will come up again as more of these cases reach court.

What Anthropic's rules actually say

Heller may have thought her chats were private, but Anthropic's documents never promised that. The company's consumer terms say it reserves the right, "at our sole discretion," to report information from or about a user, including what they type and what Claude replies, to law enforcement. Its privacy policy, effective September 10, 2026, allows disclosure when Anthropic has a good-faith belief that it's reasonably necessary to, among other things, prevent serious harm. Even if you opt out of having your chats used to train models, the policy says conversations flagged for safety review can still be used to enforce its rules and improve harm detection.

TechSpot reports that the company describes its emergency sharing more narrowly: limited cases where disclosure is needed "to prevent death or serious physical injury." That gap in wording matters. "Serious harm" in the privacy policy is a looser standard than danger of death or serious injury. And neither document tells users how often Anthropic contacts police on its own initiative or what a reviewer checks before making the call. Anthropic does publish twice-yearly government requests reports through its Transparency Hub, but they count requests it receives from authorities, not referrals it makes unprompted, so we found no public figure for how often it calls police itself.

This isn't the first time Anthropic has done this. Back in August, the San Francisco Standard reported, citing police records, that a Claude user claimed to have purchased an AR-15 and talked about targeting chief executive Dario Amodei. The company shut down his account and passed the matter to San Francisco police. He told the paper it was a joke, and no charges followed. What Anthropic said at the time could serve as its position on Heller too: "This is our safeguards process working as intended." The Next Web also reports an August case in San Antonio, where a 22-year-old was arrested on a felony terroristic threat charge after asking Claude about a shooting at an elementary school.

So in about two months, there were at least three Claude conversations that reached police, and they ended very differently: one man had no charges, two people were arrested.

Damned if they report, damned if they don't

To see why AI companies are leaning toward calling police, look at what happened when one didn't.

In June 2025, OpenAI banned the ChatGPT account of Jesse Van Rootselaar after she described gun violence scenarios, but decided not to tell police. Months later, she allegedly killed eight people in Tumbler Ridge, British Columbia. In April, Sam Altman wrote to the town: "I am deeply sorry that we did not alert law enforcement to the account that was banned in June." OpenAI says it has since loosened its criteria for reporting to police and set up direct contacts with Canadian law enforcement. Canadian officials are also considering new AI rules.

No AI company wants to write that letter. But the pressure goes both ways. Santa Clara University law professor Eric Goldman told the Standard that companies face liability if they miss a real threat, and that this pushes them to "over-disclose knowing that some of the people they identify for law enforcement shouldn't be targeted at all." The San Francisco man who said he was joking is one example of that cost. A police visit over a bad joke isn't a prison sentence, but it isn't nothing either.

There's a fair argument for what Anthropic did in Florida. A named target, a claim of a newly acquired gun and a "right now" timeline is the kind of pattern OpenAI was criticized for not acting on. There's also a fair worry: the threshold lives inside a company, the reviewers are anonymous, and users can't see the rules being applied to them.

For regular users, the practical lesson is simple, even if the policy isn't. A chatbot isn't a locked journal. Your conversations are stored on a company's servers, automated systems scan them, a person can read flagged ones, and the terms you accepted allow the company to share them with police. Most people venting about a boss or a breakup will never come close to that line. But "private" here means private from other users, not from the company.

The next date to watch is November 2, when Heller is due for arraignment in Lee County. If her case goes to trial or a motion to dismiss, it could become one of the first tests of whether typing to an AI counts as sending a threat under Florida law. And watch whether Anthropic or its rivals start publishing how often they refer users to police, the number critics most want to see.

Don't expect the big question to be answered on November 2, though. Under Florida's criminal court rules, an arraignment is where the charge is read and the defendant pleads guilty, not guilty or no contest, and a defense lawyer can file a written not guilty plea beforehand, which waives the hearing altogether. Any fight over whether a chatbot counts as "another person" would come later, in motions.