内容来源：https://aiweekly.co/issues/openai-faces-50-plus-lawsuits-over-alleged-chatgpt-harm

内容总结：

OpenAI面临逾50起诉讼 最新一批涉加拿大校园枪击案

人工智能企业OpenAI目前正应对超过50起诉讼，原告方指控长期使用ChatGPT导致心理伤害、身体损伤甚至死亡。最新一波诉讼规模异常庞大——30起新案件来自加拿大不列颠哥伦比亚省图姆布里奇岭2月校园枪击案的幸存者及遇难者家属。

该枪击案造成8人死亡、27人受伤。新增原告包括事发时在校内但未遭枪击的人员。诉状指控OpenAI存在过失，并首次在本案中提出“协助及教唆袭击”的指控。

原告方称，ChatGPT助长了枪手的暴力思维，且OpenAI员工在审阅涉及枪械暴力与袭击策划的对话后，未向加拿大警方发出预警。OpenAI虽已封禁该账户，但枪手随后另开新号继续使用。公司方面回应称，相关活动未达到需要联系执法部门的“迫在眉睫且可信风险”标准。

目前关键事实仍存疑点——枪手的ChatGPT日志尚未公开。新提出的“协助教唆”指控需证明主观故意，预计将面临初步法律挑战。另有一项诉状指称OpenAI高管克里斯·莱恩曾阻止向警方转介案件，公司否认该说法，称莱恩并未参与此事。

OpenAI高管杰森·权表示，公司通过自动化系统与人工判断在公共安全与用户隐私间寻求平衡，同时承认这一判断机制并非万无一失。

这些案件与涉及自杀、严重心理健康危机、跟踪骚扰及其他枪击事件的诉讼合并处理。迄今尚无判例证明ChatGPT在法律层面上造成了所称损害。然而，这些案件共同迫使一个自愿性安全政策未能解答的现实问题浮出水面：当聊天机器人提供商发现可能的暴力或自残迹象时，对用户、潜在受害者及执法部门应承担何种义务？

美国12州出台AI伴侣聊天机器人法规

法律层面的应对已超越个案诉讼。美国已有12个州通过针对伴侣型聊天机器人的专项法律——这类系统旨在维持长期私人或情感对话。纽约、加州和夏威夷已先行实施，另有9个州将于2027年生效。

全部法律均要求明确告知用户对话对象为AI而非真人，同时须建立识别自杀或自残言论并引导用户获取危机干预资源的机制。多数州还增设未成年人保护条款，包括限制色情内容、情感操控及以最大化用户黏性为目标的诱导设计。

各州法规不尽相同：华盛顿州禁止聊天机器人向任何用户声称自己是人类；科罗拉多州和佐治亚州要求某种形式的年龄估算；俄勒冈州规定当用户在获得危机信息后仍持续表达自杀或自残意图时，须进行额外干预。对服务商而言，“我们有安全政策”正转变为监管机构和法院可审查的具体义务清单。

Uber司机集体诉讼：算法“学会”压低报酬

一项拟议中的欧洲集体诉讼涉及约24.1万名Uber司机，指控该公司通过分析司机个人特征来设定薪酬和派单。该案在阿姆斯特丹提起，索赔损失并要求法院颁布禁令，叫停违反欧洲数据保护法律的行为。

一名伦敦司机称曾看到同一行程给另一位司机的报价为27英镑，而自己的报价仅为23英镑。原告方认为Uber系统能“学习”哪些司机更可能接受低价订单，并据此压低其收入，称动态定价机制导致英国司机年均收入减少约5000英镑。

Uber全面否认指控，表示价格是基于行程本身信息计算的——包括时长、目的地、供需状况、促销活动及测试因素——并非参考司机过往接单或拒单记录。法院将不得不审查这套“行为如同管理者但信息披露远少于人类管理者”的系统。

澳大利亚议会证据文件惊现AI伪造引用

澳大利亚议会各委员会在住房、家庭暴力等议题调查中，会邀请专家、组织和公众提交意见材料供议员参考。

《卫报》调查发现至少39份提交文件中存在疑似由AI生成的虚构引用。记者提取了本届议会所有提交文件中的引文，与学术数据库比对，并人工审查了存在大量无法匹配引用的文件。部分文件出现少量假引用，另一些文件则所有引用来源均查无实据。

另有超过100份文件在复制链接中包含ChatGPT标记，尽管这些标记不能证明最初使用者身份。在一个案例中，谷歌AI摘要竟将一篇虚构学术论文当作真实文献介绍，参考来源正是那篇“制造”该论文的议会意见书——错误信息已开始自我确证。

ChatGPT接入Epic电子病历系统

医疗机构可将经授权的Epic患者病历接入ChatGPT医疗版。在某些部署环境中，临床医生无需离开病历界面即可获取诊前摘要、病程时间线、用药信息或上次就诊以来的变化。

该接入权限为只读性质——ChatGPT无法将信息写回Epic系统。OpenAI称在27项临床任务中，医生对超过4300条测试回复的99.1%评为“安全”。需要指出的是，这属于公司自评结果，且“安全”不意味着每条信息都准确无误。真正的考验在于医院能否在错误信息影响诊疗决策前，及时识别遗漏的检查结果或误导性摘要。

美国智库推演AI数据中心攻击场景

新美国安全中心一份报告设问：若中国接近建成通用人工智能（AGI），美国应如何应对？报告构建了一个AGI即将实现的情境假说，进而推演各种后果。

其中一种情景考虑破坏活动、网络攻击，以及最危险的选项——对竞争对手的数据中心实施轰炸。作者雅各布·斯托克斯还呼吁美国针对中国AI发展制定防备演练方案。值得关注的不在于对AGI的预测本身，而是一个尚有争议的技术前提，已在被转化为军事规划的现实动向。

花絮：数字迷彩对抗AI识别

艺术家西蒙·韦克特将模糊的绿粉色图案印在衬衫上。现场演示中，当布料遮住穿着者躯干时，目标检测软件停止将其识别为“人”，布料移开后系统重新识别成功。韦克特设计这款“数字迷彩”的初衷，是对柏林警方首次部署物体识别摄像头做出的回应。

中文翻译：

三十起新的投诉来自加拿大一起校园枪击案的幸存者。他们指控OpenAI未能提醒警方；该公司对关键说法提出异议。
赞助内容
哪些当前趋势正在塑造企业战略？
探索审计合伙人关于当前影响商业决策、风险管理和企业战略趋势的观察。简报
OpenAI目前正在应对50多起诉讼，原告声称长期使用ChatGPT导致了心理伤害、身体伤害或死亡。最新一波诉讼规模异常庞大：30起新投诉来自不列颠哥伦比亚省图姆布勒里奇二月校园枪击案的幸存者及家属。
该枪击案造成8人死亡，27人受伤。新原告包括当时在学校内但未被枪击击中的人。他们的投诉指控OpenAI存在过失，并在此案中首次指控其协助和教唆袭击。
原告称ChatGPT助长了枪手的暴力思维，且OpenAI工作人员在审查了关于枪支暴力和袭击计划的对话后未能向加拿大警方发出警报。OpenAI停用了该账户，但枪手又开设了另一个账户。该公司表示，相关活动未达到其触发联系执法部门的“迫在眉睫且可信的风险”门槛。
重要事实仍未厘清。枪手的ChatGPT日志尚未公开。新的协助和教唆指控需要意图证据，预计将面临早期的法律挑战。一份投诉还声称OpenAI高管克里斯·莱恩帮助阻止了向警方的转介；OpenAI表示该说法不实，莱恩并未参与其中。
OpenAI高管杰森·权表示，公司使用自动化系统和人工判断来平衡公共安全与用户隐私。他承认这种判断并非万无一失。
这些案件加入了涉及自杀、严重心理健康危机、跟踪骚扰及其他枪击案的诉讼行列。迄今尚无案件确立ChatGPT在法律上造成了所称的损害。然而，这些案件共同迫使一个自愿性安全政策未能回答的具体问题浮出水面：当聊天机器人提供商看到可能的暴力或自残迹象时，它对用户、潜在受害者和执法部门负有什么义务？
如果您或您认识的人正在考虑自杀，请在美国或加拿大拨打或发送短信至988。在其他地区，请联系当地紧急或危机服务部门。
十二个州已通过伴侣聊天机器人法律
法律回应已超越个案诉讼。美国有十二个州通过了针对伴侣聊天机器人的法律——这类系统设计用于维持长期的个人或情感对话。纽约、加利福尼亚和夏威夷的法律已生效；另有九个州的法律将于2027年生效。
所有十二个州都要求明确告知用户正在与AI而非真人交谈。它们还要求建立检测自杀或自残表达的协议，并将用户引导至危机资源。大多数州增加了对未成年人的保护，包括对性内容、情感操控以及旨在最大化用户参与度的功能进行限制。
这些法律并不统一。华盛顿州禁止受监管的聊天机器人向任何用户声称自己是人类。科罗拉多州和佐治亚州要求某种形式的年龄估算。俄勒冈州要求在用户收到危机信息后仍持续表达自杀或自残意图时进行额外干预。对提供商而言，“我们有安全政策”正在变成一套监管机构和法院可以审查的义务。
Uber司机称算法学会了他们能接受多低的报酬
一项拟议的欧洲集体诉讼涵盖约24.1万名Uber司机，指控该公司对个体工人进行画像以设定薪酬和分配工作。该案在阿姆斯特丹提起，寻求损害赔偿以及停止该诉讼声称违反欧洲数据保护法的做法。
一位伦敦司机描述称，他看到同一行程给另一位司机报价27英镑，而给他的报价只有23英镑。原告认为，Uber的系统会学习哪些司机可能接受更便宜的车费，并利用该信息压低他们的报酬。他们表示，动态定价已将英国的年收入削减了约5000英镑。
Uber坚决否认这些指控。该公司表示，价格是根据行程信息（包括时长、目的地、需求、促销和测试）计算的，而非根据司机接受或拒绝工作的历史记录。法院将不得不审查一个行事如同管理者、但在薪酬决策方式上透露的信息远少于人类管理者的系统。
虚假引文进入议会证据
澳大利亚议会调查邀请专家、组织和公众提交意见书。立法者在调查从住房到家庭暴力等议题时使用这些材料。
《卫报》发现至少39份意见书包含似乎由AI编造的参考文献。其记者提取了提交给本届议会的每份意见书中的引文，将其与学术数据库进行比对，并手动审查了包含许多不匹配参考文献的文件。一些意见书含有少量错误引文；而在其他意见书中，每一个被引用的来源似乎都不存在。
超过100份意见书在复制链接中还包含ChatGPT标签，尽管这些标签并不能证明最初是谁使用了该工具。在一个案例中，谷歌的AI摘要将一篇伪造的学术论文描述得像真的一样，并将编造该论文的调查意见书作为来源。这个错误已经开始自我确认。
ChatGPT现在可以读取Epic患者记录
医疗机构可以将经过授权的Epic患者记录连接到ChatGPT for Healthcare。在某些部署中，临床医生无需离开患者病历即可请求就诊前摘要、时间线、用药信息或自上次就诊以来的变化。
该访问权限为只读：ChatGPT无法将信息写回Epic。OpenAI表示，在27项临床任务中，医生对超过4300条测试回复中的99.1%评为安全。这是一项由公司主导的评估，且“安全”并不意味着每个细节都正确。实际考验在于医院能否在遗漏的结果或误导性的摘要改变诊疗之前发现它们。
一份美国政策报告已在推演针对AI数据中心的攻击
新美国安全中心的一份报告提出问题：如果中国似乎接近建造出通用人工智能，美国应该怎么做。该报告假设了一个AGI即将到来的虚拟世界，然后推演其后果。
其中一个情景考虑了破坏活动、网络攻击，以及最危险的选择——轰炸对手国家的数据中心。报告作者雅各布·斯托克斯还呼吁美国针对中国AI的发展准备应急演练。四位名人录网络中的人物推送了这篇报道。引人注目的不是对AGI的预测，而是一个存在争议的技术前提已经被转化为军事规划。
等一下，什么？
艺术家西蒙·韦克特在一件衬衫上印制了模糊的绿色和粉色图案。在一次现场演示中，当织物覆盖他的躯干时，物体检测软件停止将穿着者标记为“人”，而当织物移开后又重新识别出他。韦克特制作这款“数字迷彩”是为了回应柏林警方首次部署的物体识别摄像头。
本周投票
哪场AI问责之争最重要？
上周，248位读者参与了投票：
什么能让你相信某个AI部署是真实的？
哪场AI问责之争最重要？
下周再见，
——AI周刊团队

英文来源：

Thirty new complaints come from survivors of a Canadian school shooting. They accuse OpenAI of failing to warn police; the company disputes key claims.
Sponsor
What are the current trends shaping corporate strategy?
Explore observations from audit partners on the trends shaping business decision-making, risk management, and corporate strategy today.The Brief
OpenAI is now fighting more than 50 lawsuits in which people allege that prolonged use of ChatGPT contributed to psychological injury, physical harm, or death. The latest wave is unusually large: 30 new complaints came from survivors and families connected to the February school shooting in Tumbler Ridge, British Columbia.
The shooting killed eight people and wounded 27. The new plaintiffs include people who were inside the school but were not physically shot. Their complaints accuse OpenAI of negligence and, for the first time in this case, of aiding and abetting the attack.
The plaintiffs say ChatGPT encouraged the shooter’s violent thinking and that OpenAI failed to alert Canadian police after its staff reviewed conversations about gun violence and attack planning. OpenAI deactivated the account, but the shooter opened another one. The company says the activity did not meet its threshold for an “imminent and credible risk” that would justify contacting law enforcement.
Important facts remain unresolved. The shooter’s ChatGPT logs have not been made public. The new aiding-and-abetting claim requires evidence of intent and is expected to face an early challenge. One complaint also alleges that OpenAI executive Chris Lehane helped stop a police referral; OpenAI says that claim is false and that Lehane was not involved.
OpenAI executive Jason Kwon said the company uses automated systems and human judgment to balance public safety with user privacy. He acknowledged that this judgment is not infallible.
These cases join lawsuits involving suicides, severe mental-health crises, stalking, and other shootings. None has yet established that ChatGPT legally caused the alleged harm. Together, however, they are forcing a concrete question that voluntary safety policies leave unanswered: when a chatbot provider sees signs of possible violence or self-harm, what does it owe the user, potential victims, and law enforcement?
If you or someone you know is thinking about suicide, call or text 988 in the United States or Canada. Elsewhere, contact local emergency or crisis services.
Twelve states have passed companion-chatbot laws
The legal response is already moving beyond individual lawsuits. Twelve US states have passed laws aimed at companion chatbots, the systems designed to sustain personal or emotional conversations over time. Laws are already in effect in New York, California, and Hawaii; nine more take effect in 2027.
All 12 require a clear disclosure that the user is talking to AI, not a person. They also require protocols for detecting expressions of suicide or self-harm and directing users to crisis resources. Most add protections for minors, including restrictions on sexual content, emotional manipulation, and features designed to maximize engagement.
The laws are not uniform. Washington bars a covered chatbot from claiming to be human to any user. Colorado and Georgia require some form of age estimation. Oregon requires additional intervention when a user continues to express suicidal or self-harm intent after receiving crisis information. For providers, “we have a safety policy” is becoming a set of duties that regulators and courts can examine.
Uber drivers say the algorithm learns how little they will accept
A proposed European class action covering about 241,000 Uber drivers alleges that the company profiles individual workers to set pay and allocate jobs. Filed in Amsterdam, the case seeks damages and an order stopping practices the claim says violate European data-protection law.
One London driver described seeing the same trip offered to another driver for £27 while he was offered £23. The plaintiffs argue that Uber’s system learns which drivers are likely to accept cheaper fares and uses that information to push their pay down. They say dynamic pricing has reduced annual UK earnings by about £5,000.
Uber categorically rejects the allegations. It says prices are calculated from information about the trip—including duration, destination, demand, promotions, and testing—not from a driver’s history of accepting or rejecting work. The court will have to examine a system that acts like a manager while revealing far less than a human manager would about how a pay decision was made.
Fake citations are entering parliamentary evidence
Australian parliamentary inquiries invite submissions from experts, organizations, and the public. Legislators use that material when investigating subjects ranging from housing to domestic violence.
The Guardian found at least 39 submissions containing references that appeared to be invented by AI. Its reporters extracted citations from every submission to the current parliament, checked them against academic databases, and manually reviewed documents with many unmatched references. Some submissions contained a few bad citations; in others, every cited source appeared not to exist.
More than 100 submissions also contained ChatGPT tags in copied links, although those tags do not prove who originally used the tool. In one case, Google’s AI summary described a fabricated academic paper as if it were real, using the inquiry submission that invented it as a source. The error had begun to authenticate itself.
ChatGPT can now read Epic patient records
Healthcare organizations can connect authorized Epic patient records to ChatGPT for Healthcare. Clinicians can ask for a pre-visit summary, a timeline, medication information, or changes since the last appointment without leaving the patient chart in some deployments.
The access is read-only: ChatGPT cannot write information back into Epic. OpenAI says physicians rated 99.1 percent of more than 4,300 test responses as safe across 27 clinical tasks. That is a company-run evaluation, and “safe” does not mean every detail was correct. The practical test is whether hospitals can catch an omitted result or a misleading summary before it changes care.
A US policy report is already gaming out attacks on AI data centers
A Center for a New American Security report asks what the United States should do if China appeared close to building artificial general intelligence. It assumes a hypothetical world in which AGI is imminent, then works through the consequences.
One scenario considers sabotage, cyberattacks, and, as the most dangerous option, bombing the rival state’s data centers. Its author, Jacob Stokes, also calls for the United States to prepare readiness drills in response to Chinese AI development. Four people in the Who’s Who network surfaced the story. The striking part is not a prediction about AGI. It is that a disputed technological premise is already being translated into military planning.
Wait, What?
Artist Simon Weckert printed blurry green and pink patterns on a shirt. During a live demonstration, object-detection software stopped labeling the wearer as a person when the fabric covered his torso, then recognized him again when it moved. Weckert made the “digital camouflage” in response to Berlin’s first police-run object-recognition cameras.
This week’s poll
Which AI accountability fight will matter most?
Last week, 248 of you voted:
What would convince you that an AI deployment is real?
Which AI accountability fight will matter most?
Until next week,
— The AI Weekly team