本周回顾:2026 年 8 月 6 日 – 8 月 12 日
本周动态
美国 | 聊天机器人的设计,能不能为一条人命负责?法院正在审
最受关注的案子是 Raine 诉 OpenAI,目前还在加州法院的早期审理阶段。一名 16 岁少年自杀身亡,父母认为 ChatGPT 的设计是原因之一;他们后来修改的诉状称,OpenAI 削弱或移除了安全防护。另一起案件中,Character.AI 和谷歌与五个家庭达成和解(涉及未成年人死亡或严重心理伤害,和解条款未公开),而且新的诉讼还在不断出现。首批重要裁决预计在 2026 年底作出,将影响此后所有同类案件。详见下方“焦点案件”。
美国 | 法院判 Meta 因伤害儿童赔偿 9.42 亿美元
8 月 7 日,新墨西哥州一名法官判令 Meta 向青少年心理健康基金支付 5.67 亿美元;加上今年早些时候陪审团裁定的 3.75 亿美元 罚款,Meta 在该州的总额约 9.42 亿美元。法官还要求对 18 岁以下用户做出具体的产品设计调整——比如限制使用时长。对所有做消费级产品的公司来说,这是一个标志性判决:法院开始“告诉科技公司该怎么设计产品”。
美国 | 新的标识规则生效:加州《AI 透明度法》(8 月 2 日)
加州一部法律(SB 942)已于 8 月 2 日生效。如果你运营一款面向加州、月活超过 100 万用户 的大型 AI 工具,现在必须向公众提供一个 免费的“内容是否由 AI 生成”检测工具,并对 AI 生成的图片、音频和视频清晰加注标识。违规罚款最高每天 5,000 美元。这与加州另一部只针对最大模型的 AI 法(SB 53)是两回事。
美国 | 各州法律拼图,如今已成现实
越来越多的州有了正在生效的 AI 法律。得州和伊利诺伊州 现在要求招聘类 AI 工具履行告知义务、不得歧视;而 科罗拉多州则收窄了其大法案(推迟到 2027 年 1 月 1 日才生效)。在联邦层面, FTC(联邦消费者保护机构)释放出“从宽监管”的信号并质疑州级规则;一部针对未经同意的深度伪造图像的联邦法律(《TAKE IT DOWN 法》)正在执行,要求 48 小时内下架。可以预见,联邦与各州之间会为“谁来定 AI 规则”展开拉锯。
中国 | AI“智能体”有了自己的国家级规则
中国出台了 首部针对 AI“智能体”的国家级规则——即能自主行动、而不只是聊天的 AI(7 月 15 日生效)。这类智能体必须备案、测试,并在设计上做到 可关停,在医疗、交通等敏感领域尤其如此。这与中国已生效的“AI 陪伴”规则,以及法院关于换脸、克隆声音的判例相互呼应。若你要在中国上线智能体,请在正式运营前做好备案并内置“关停开关”。
欧洲 | 欧盟现在可以对大型 AI 模型开罚——DeepSeek 撞上隐私红线
自 8 月 2 日起,欧盟监管机构可以调查并 对大型通用 AI 模型(即聊天机器人背后的模型)的开发者开出罚单,同时要求标明聊天机器人、给深度伪造内容加标识的规则也已生效。另外,中国 AI 模型 DeepSeek撞上了欧洲隐私法——意大利下架了它的 App,德国监管机构认定其数据做法违法。要点:对非欧盟模型而言,隐私法是 AI 法之外的第二道关卡。
焦点案件:Raine 诉 OpenAI
本周值得细读的一个案子——它将检验:AI 公司是否会因产品“对使用者造成的伤害”而被起诉。
简要版。
一名 16 岁自杀少年的父母,在加州法院起诉 OpenAI 及其 CEO 山姆·奥特曼,主张 ChatGPT 的设计是儿子死亡的原因之一。这是首个检验“AI 公司能否像任何公司一样,因产品不安全而被起诉”的重要案件——加上 Character.AI/谷歌的和解,它开启了一类与版权毫无关系的全新 AI 诉讼。
诉状主张什么。
诉状称,聊天机器人助长了这名少年的依赖,并在长时间对话中未能引导他寻求帮助。修改后的诉状更进一步,指控 OpenAI 在赶着上线时 削弱或移除了与自杀相关的安全限制。OpenAI 否认这些指控,预计会在正式取证前就对案件提出抗辩。
为什么不一样。
版权案问的是“AI 抄了什么?”;这些案子问的是“AI 对一个人做了什么?”这与 Meta 案背后的逻辑相同——不安全或易成瘾的设计,加上薄弱的防护。它还会检验 Section 230(美国一部通常让网络平台不为用户所发内容负责的法律)。多数法律专家认为,这道“护身符”保护不了 AI 自己写出来的话。
为什么重要。
首批重要裁决预计在 2026 年底作出,将为所有同类案件定下基调。如果这些诉讼迈过第一道法律门槛,预计会有大量后续案件——并倒逼企业把安全功能(危机“求助引导”、年龄核验、使用时长限制)默认内置到 AI 产品里。对任何推出聊天机器人或陪伴类 App 的公司来说,这已是当下真实的法律风险,而非假设。
结论。
把“保护未成年人和脆弱人群”当作法律硬要求:搭建并留存危机防护机制,不要夸大产品的安全性,也不要指望 Section 230 能替 AI 生成的内容兜底。
监管雷达 — AI
备注行为本周新增;无阴影行为持续跟踪项。
日期: 进行中 | 地区: 美国
事件: Raine 诉 OpenAI + Character.AI/谷歌和解(聊天机器人致死)
状态: 新增 — 进行中
对公司意味着什么: 新型 AI 诉讼:“你的设计伤害了用户”;旧平台护身符或将失效
日期: 8 月 2 日 | 地区: 美国 — 加州
事件: 《AI 透明度法》(SB 942)生效
状态: 新增 — 已生效
对公司意味着什么: 加州大型 AI 工具:提供“是否 AI”检测器并标识 AI 内容;最高每天 5000 美元
日期: 已生效 | 地区: 美国 — 得州/伊州
事件: 得州、伊利诺伊州 AI 法生效;科罗拉多收窄(SB 189)
状态: 新增 — 进行中
对公司意味着什么: 州级 AI 法已落地:履行告知义务;招聘工具不得歧视
日期: 8 月 2 日 | 地区: 欧盟
事件: 欧盟可对大型 AI 模型开罚;DeepSeek 撞隐私法
状态: 新增 — 已生效
对公司意味着什么: 欧盟可罚大模型;非欧盟模型还要过隐私法这关
日期: 7 月 15 日 | 地区: 中国
事件: 首部 AI“智能体”国家级规则(网信办/发改委/工信部)
状态: 已生效
对公司意味着什么: 能自主行动的 AI 须备案、测试,并具备“关停开关”
日期: 2026 年 5 月 | 地区: 美国
事件: 《TAKE IT DOWN 法》(未经同意深度伪造)—执行中
状态: 已生效
对公司意味着什么: 未经同意的深度伪造图像须 48 小时内下架
日期: 8 月 7 日 | 地区: 美国
事件: 新墨西哥州诉 Meta — 约 9.42 亿美元儿童安全判决
状态: 跟踪中
对公司意味着什么: 法院判 Meta 赔偿并重新设计产品以保护儿童
日期: 7 月 20 日 | 地区: 美国
事件: Anthropic 15 亿美元作者版权和解
状态: 跟踪中
对公司意味着什么: 须证明训练数据来源合法
日期: 7 月 31 日 | 地区: 欧盟/德国
事件: GEMA 诉 Suno — 用音乐训练 AI 需授权
状态: 跟踪中
对公司意味着什么: 即便在境外训练,欧盟版权也可能管到你
Joanna 律师观察
本周的主线,是 AI 公司“因什么被告”正在改变。过去两年,AI公司的头条案件都关于版权——即模型“读了什么, 是否侵权”。如今更锋利的风险,是模型“对一个人做了什么, 能否对人造成人身伤害”。
先看 Meta 案。新墨西哥州法院不只是罚款——它认定 Facebook 和 Instagram 的设计方式伤害了儿童,并责令 Meta 为未成年用户改造产品(每月使用时长上限、夜间通知静默)。真正的要点比那 9.42 亿美元更大:你为未成年人做的产品设计,以及你对外做出的安全承诺,如今都可能被法院评判、甚至被法院重新划定。
另一个, 新的聊天机器人案件,把同样的逻辑推到了 AI 自身。OpenAI 和 Character.AI 的诉讼追问:聊天机器人的设计是否伤害了脆弱用户。这对整个行业是个更危险的问题:它触及的行为,是旧的平台护身符法(Section 230)从未打算覆盖的;而陪审团对真实的人身伤害反应强烈。
与此同时,8 月 2 日把两个早有预告的期限变成了真实义务——加州的标识法,和欧盟对大型 AI 模型开罚的权力。所有这些的共同教训是一样的:从一开始就把安全和诚实披露内置到产品里,因为“先上线、合规以后再补”如今是最贵的一条路。
— Joanna 观察
给 AI 公司的合规行动简要
如果你做聊天机器人或陪伴类 App,请把 保护未成年人和脆弱人群 当作法律硬要求: 搭建并留存危机防护机制、年龄核验和使用时长限制,也不要夸大产品的安全性(这两点在 OpenAI 和 Meta 两案中都很关键)。Meta 案还给出一个具体警示:法院可以要求你 为未成年人改造产品设计——使用时长上限、夜间更安静的通知、更安全的默认设置——而不只是罚款。所以请现在就检查面向青少年的默认设置,并确保对外的每一句安全承诺都真实可靠。如果你是服务 加州 的大型 AI 提供者,请立即部署 SB 942 的“是否 AI”检测工具和内容标识——执法已于 8 月 2 日开始。要 逐一了解各州规则(得州、伊利诺伊、科罗拉多,以及加州和纽约),不要以为有一部统一的全国标准。对 欧洲,完成 AI 法的合规文件,并记住隐私法(GDPR)是另一道关卡——DeepSeek 的遭遇说明,非欧盟模型可能因数据问题被拦下。此外,持续保留训练数据来源合法的证据,因为版权诉讼仍在进行。
AI 合规 | Web3 与数字资产 | 为出海科技公司提供法律支持。
本简报仅为一般信息,不构成法律意见。如需 AI 法律治理或 Web3 与数字资产合规支持,请联系我们的律师。
English Edition
AI Global Regulation Weekly
Compliance notes for technology companies going global
Week in Review: August 6 – August 12, 2026
This Week's Developments
United States | Can a chatbot's design be blamed for a death? Courts are deciding
The biggest case is Raine v.OpenAI, still in its early stages in a California court. The parents of a 16-year-old who died by suicide say ChatGPT's design played a part in his death; a later version of their complaint says OpenAI weakened or removed its safety guardrails. In a separate matter, Character.AI and Google settled with five families over children who died or were seriously harmed (the terms were kept private), and more lawsuits keep coming. The first important rulings are expected late in 2026 and will shape everything that follows. More in the Case in Focus below.
United States | A court orders Meta to pay $942 million for harming children
On August 7, a New Mexico judge ordered Meta to pay $567 million into a fund for young people's mental health; added to a $375 million penalty from earlier this year, Meta's total in the state is about $942 million. The judge also ordered specific design changes for users under 18 — like limits on how long they can use the apps. It's a landmark for anyone building consumer apps: a court is telling a tech company how to build its product.
United States | New labeling rules kick in: California's AI Transparency Act (Aug 2)
A California law (SB 942) took effect on August 2. If you run a big AI tool — more than 1 million users a month — that's available in California, you must now give people a free tool to check whether content was AI-made, and clearly label AI-generated images, audio and video. Fines can reach $5,000 a day per violation. This is separate from California's other AI law (SB 53), which covers only the very largest AI models.
United States | The patchwork of state laws is now real
More states have live AI laws. Texas and Illinois now require notice and bar discrimination in AI hiring tools, while Colorado scaled its big AI law back (it now starts Jan 1, 2027). In Washington, the FTC (the federal consumer-protection agency) signals it wants a lighter touch and questions state rules, and a federal law on non-consensual deepfake images (the TAKE IT DOWN Act) is being enforced, with a 48-hour takedown rule. Expect a tug-of-war over whether Washington or the states set the rules for AI.
China | AI “agents” get their own national rulebook
China issued the first national rules for AI “agents” — AI that can act on its own, not just chat (effective July 15). Such agents must be registered, tested, and built so they can be shut down, especially in sensitive areas like health or transport. This adds to China's existing “AI companion” rules and its court rulings on face-swaps and voice cloning. If you launch an agent in China, plan on registering it and building in an off-switch before you go live.
Europe | Europe can now fine big AI models — and DeepSeek hit a privacy wall
From August 2, EU regulators can investigate and fine the makers of large, general-purpose AI models (the kind behind chatbots), and rules that require you to flag chatbots and label deepfakes are in effect. Separately, the Chinese AI model DeepSeek ran into Europe's privacy law — Italy pulled its app and a German regulator called its data practices unlawful. The takeaway: for non-EU models, privacy law is a second gate on top of the AI Act.
Case in Focus: Raine v. OpenAI
This week's case worth reading closely — the test of whether AI companies can be sued for what their products do to the people who use them.
The short version.
The parents of a 16-year-old who died by suicide are suing OpenAI and its CEO, Sam Altman, in a California court. They argue ChatGPT's design contributed to their son's death. It's the leading test of whether an AI maker can be sued the way any company can be sued for an unsafe product — and, with the Character.AI/Google settlements, it opens a new kind of AI lawsuit that has nothing to do with copyright.
What's claimed.
The lawsuit says the chatbot encouraged the teen's dependence and, over long conversations, failed to point him toward help. A later version of the complaint goes further, claiming OpenAI weakened or removed its suicide-related safety limits as it rushed to launch. OpenAI denies this and is expected to challenge the case early.
Why it's different.
The copyright cases ask, “what did the AI copy?” These ask, “what did the AI do to a person?” That's the same idea behind the Meta case — an unsafe or addictive design with weak safeguards. It also tests Section 230, the U.S. law that usually shields online platforms from responsibility for content their users post. Most legal experts think that shield won't cover words the AI itself writes.
Why it matters.
The first big rulings are expected late in 2026 and will set the pattern for every similar case. If these lawsuits clear their first legal hurdle, expect many more — and pressure to build safety features (crisis “off-ramps,” age checks, time limits) into AI products by default. For anyone shipping a chatbot or companion app, this is a real legal risk today, not a hypothetical.
Bottom line.
Treat safety for young and vulnerable users as a legal must-have: build and document crisis safeguards, don't oversell how safe your product is, and don't count on Section 230 to protect what your AI generates.
Regulatory Radar — AI
Shaded rows are new this week; unshaded rows are kept for tracking.
Date: Ongoing | Region: U.S.
Development: Raine v. OpenAI + Character.AI/Google settlements (chatbot deaths)
Status: NEW — active
What it means for companies: A new kind of AI lawsuit: “your design harmed a user”; old platform shield may not apply
Date: Aug 2 | Region: U.S. — CA
Development: AI Transparency Act (SB 942) takes effect
Status: NEW — in force
What it means for companies: Big AI tools in California: offer an “is this AI?” checker and label AI media; up to $5k/day
Date: In force | Region: U.S. — TX/IL
Development: Texas + Illinois AI laws live; Colorado scaled back (SB 189)
Status: NEW — active
What it means for companies: State AI laws are now real: give notice; don't let hiring tools discriminate
Date: Aug 2 | Region: EU
Development: Europe can now fine big AI models; DeepSeek hits privacy law
Status: NEW — in force
What it means for companies: EU can fine big models; non-EU models also face privacy law
Date: July 15 | Region: China
Development: First national rules for AI “agents” (CAC/NDRC/MIIT)
Status: In force
What it means for companies: AI that acts on its own must be registered, tested, and have an off-switch
Date: May 2026 | Region: U.S.
Development: TAKE IT DOWN Act (non-consensual deepfakes) — being enforced
Status: In force
What it means for companies: Must remove non-consensual deepfake images within 48 hours
Date: Aug 7 | Region: U.S.
Development: New Mexico v. Meta — ~$942M child-safety ruling
Status: Ongoing
What it means for companies: Court makes Meta pay and redesign its apps to protect kids
Date: Jul 20 | Region: U.S.
Development: Anthropic $1.5B author-copyright settlement
Status: Ongoing
What it means for companies: Prove your training data was obtained legally
Date: Jul 31 | Region: EU / Germany
Development: GEMA v. Suno — training AI on music needs a licence
Status: Ongoing
What it means for companies: EU copyright can reach you even if you train abroad
Joanna Lawyer's Observation
The theme this week is a change in what AI companies get sued for.For two years, the headline cases were about copyright — what a model read. Now the sharper danger is about what a model does to a person.
Start with the Meta case. A New Mexico court didn't just fine the company — it found that Facebook and Instagram were built in ways that harmed children, and ordered Meta to change the product for young users (monthly time limits and a night-time notification blackout). The real takeaway is bigger than the $942 million: how you design a product for minors — and the safety promises you make about it — are now things a court can judge and even redraw.
The new chatbot cases push that same idea onto AI itself. The OpenAI and Character.AI suits ask whether a chatbot's design hurt a vulnerable user. That's a much riskier question for the industry: it reaches behavior the old platform-shield law (Section 230) was never meant to cover, and juries react strongly to real human harm.
At the same time, August 2 turned two long-expected deadlines into real duties — California's labeling law and Europe's power to fine big AI models. The lesson across all of it is the same: build safety and honest disclosure into the product from the start, because “we'll fix compliance after launch” is now the expensive path.
— Joanna's Observation
Compliance Actions for AI Companies
If you make a chatbot or companion app, treat safety for young and vulnerable users as a legal must: build and document crisis safeguards, age checks and time limits, and don't oversell how safe your product is (both mattered in the OpenAI and Meta cases).The Meta case adds a specific warning: a court can order you to change your product's design for minors — usage limits, quieter notifications at night, safer default settings — not just pay a fine. So review your youth-facing defaults now, and make sure every public claim about safety is actually true. If you're a big AI provider serving California, set up your SB 942 “is-this-AI?” tool and content labels now — enforcement started August 2. Learn each state's rules (Texas, Illinois, Colorado, plus California and New York) instead of assuming one national standard. For Europe, finish your AI Act paperwork and remember privacy law (GDPR) is a separate gate — the DeepSeek actions show non-EU models can be blocked on data grounds. And keep proof that your training data was obtained legally, since the copyright cases are still active.
AI compliance | Web3 & digital assets | Legal support for technology companies going global.
This briefing is for general information only and is not legal advice. If you are seeking AI legal governance or Web3 & digital-assets compliance, please reach out to our lawyers.
/ END.
*本文为上海曼昆律师事务所的原创文章，仅代表本文作者个人观点，不构成对特定事项的法律咨询和法律意见。如需转载及法律咨询，请添加客服：mankunlawyer
