
IT之家 10 月 5 日消息,美国佛罗里达州李县(Lee County)一名 30 岁女子因涉嫌在与 Anthropic 旗下聊天机器人 Claude 的对话中写下针对当地警长办公室的暴力威胁而被捕。
据当地警方通报,邦尼塔斯普林斯居民卡莉 · 米歇尔 · 赫勒(Carli Michelle Heller)于 9 月 26 日在对话中称要对李县警长办公室实施枪击(shoot up)。相关内容随即触发 Anthropic 的安全风控机制,并被 Claude 上报给人工审核及执法部门。
据当地媒体 WINK News 报道,警长卡迈恩 · 马尔切诺(Carmine Marceno)告诉该媒体,赫勒后来称她把 AI 当作“日记”使用。
调查人员称,Anthropic 借助安全机制监测关键词和潜在威胁内容,相关信息视严重程度被升级至人工审核,随后由人工团队确认后联系当地执法部门。
接报后,当地警员前往其住所,并将该女子“无冲突拘留”,案件随后由李县警长办公室一名情报警探接手。赫勒告诉调查人员,当下她面临一项书面暴力威胁指控,但尚未被法院定罪。
佛罗里达州法规第 836.10 条规定,任何人借助可以被他人查看的书面或电子记录,威胁杀人、实施身体伤害、开展大规模枪击或实施恐怖行为,均可能构成二级重罪。现行法条同时明确,相关规定适用于电子记录,并规定向他人发送、发布或传播包含上述威胁的记录可能构成犯罪。至于本案中的 Claude 对话是否满足具体法律构成,以及检方能否在司法程序中证明犯罪事实,则仍需由法院裁定。
马尔切诺表示,AI 可能改变了威胁的产生方式,但保护李县居民的承诺不变,当有人利用 AI 实施或协助实施威胁时,警方必须认真对待。
同时,该案件也引发了关于聊天机器人隐私边界的讨论。Anthropic 当下公开的政府信息请求政策表明,公司通常只有在收到有效法律程序后才会向政府披露使用者信息,但如果公司认为某种紧急情况可能导致迫在眉睫的身体伤害或死亡,且及时给出相关信息可能避免危险,则存在紧急披露例外。
Anthropic 在 2026 年 7 月更新的透明度页面也表明,公司使用安全团队负责外观和实施用于执行使用政策的检测与监控机制,并处理执法机构和政府提出的涵盖紧急披露在内的数据请求。由此来看,将 Claude 作为私人日记使用,并不意味着相关内容必然只会由使用者本人看到;平台可能依据安全和使用政策对部分内容开展检测和审核,在符合相应条件的情况下,还可能向执法机构给出信息。
另外,Anthropic 协议条款还提到,与 AI 对话并不受类似律师与委托人之间的保密特权保护,但 Anthropic 尚未详细说明此次对话的审核细节。
除了这起案件外,加拿大不列颠哥伦比亚省近期在美国旧金山联邦法院起诉 OpenAI 及其 CEO 萨姆 · 奥尔特曼(Sam Altman),指称 OpenAI 此前已标记 2 月塔姆勒岭(Tumbler Ridge)校园枪击案枪手与 ChatGPT 的暴力相关对话,却未向警方通报,本可阻止该事件。
据 TechSpot 报道,佛罗里达州 6 月已起诉 OpenAI 和奥尔特曼,指称 ChatGPT 造成了现实伤害,其中涵盖 2025 年佛罗里达州立大学枪击案(IT之家注:上述均为诉方指控,法院尚未认定相关责任)。
本文来源:https://www.ithome.com/1/009/795.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
IT Home, October 5 — A 30-year-old woman in Lee County, Florida was arrested for allegedly writing violent threats against the local sheriff’s office in a conversation with Anthropic’s Claude chatbot.
According to local police, resident Carli Michelle Heller of Bonita Springs said on September 26 that she would "shoot up" the Lee County Sheriff’s Office. The content triggered Anthropic’s automated safety mechanisms, and Claude escalated it to human review and law enforcement.
Local outlet WINK News reported that Sheriff Carmine Marceno told the outlet Heller later claimed she was using the AI as a "diary."
Investigators said Anthropic monitors keywords and potential threat content through its safety systems; depending on severity, the material is escalated to human review, and once confirmed by the human team it is passed to local law enforcement.
Responding officers went to her home and detained her without conflict; the case was then taken over by a detective in the Sheriff’s Office. Heller told investigators she now faces a written threat charge but has not been convicted.
Florida Statute 836.10 provides that anyone who threatens to kill, inflict bodily harm, carry out a mass shooting or commit terrorism through a written or electronic record viewable by others may commit a second-degree felony. The statute explicitly applies to electronic records and states that sending, posting or distributing such threatening records may constitute a crime. Whether the Claude conversation meets the elements of the offense, and whether prosecutors can prove it, remains for the court to decide.
Marceno said AI may have changed the way threats are produced, but the commitment to protecting Lee County residents is unchanged: when someone uses AI to carry out or assist in a threat, police must take it seriously.
The case has also sparked debate about the privacy boundaries of chatbots. Anthropic’s published government information request policy indicates it generally discloses user information only after receiving valid legal process, with an exception for emergencies where the company believes imminent bodily harm or death could occur and timely disclosure could avert danger.
Anthropic’s transparency page, updated in July 2026, states that the company uses safety teams to build and operate the detection and monitoring mechanisms that enforce its usage policy, and that it handles data requests from law enforcement and governments including emergency disclosures. This suggests that using Claude as a private diary does not mean the content is seen only by the user: the platform may examine and review some content under its safety and usage policies, and may provide information to law enforcement when the conditions are met.
Anthropic’s terms also note that AI conversations are not covered by confidentiality privileges of the kind that exist between lawyers and clients, though the company has not detailed the review of this particular conversation.
Beyond this case, the Canadian province of British Columbia recently sued OpenAI and CEO Sam Altman in the U.S. District Court in San Francisco, alleging OpenAI had flagged violent ChatGPT conversations tied to the February Tumbler Ridge school shooting but failed to alert police, which could have prevented the incident.
TechSpot reported that Florida sued OpenAI and Altman in June, alleging ChatGPT caused real-world harm, including the 2025 Florida State University shooting. (IT Home note: these are plaintiff allegations; the courts have not established liability.)
