跨境人工智能与金融科技业务 — 美国 · 纽约 · 中国 · 香港 · 欧盟
两周回顾：2026年8月27日 – 9月9日
一、本期回顾 — 新法与新案
本期人工智能法律领域颇为活跃，而数字资产立法则相对平静。在美国，司法部（DOJ）就人工智能著作权诉讼首次表明联邦政府立场，提交意见支持 OpenAI 的合理使用抗辩；与此同时，Andersen 诉 Stability AI 案在旧金山开庭审理，成为美国首例由陪审团审议“模型即复制件”理论的案件。加利福尼亚州议会在结束2026年会期时通过了约三十项人工智能相关法案，现正等待州长签署。在中国，最高人民法院发布了首个关于人工智能纠纷案件审理的全国性司法规则，明确回应了深度伪造与人格权问题。在欧盟，通用人工智能（GPAI）提供者的首批系统性风险评估将于9月中旬到期。在 Web3 领域，美国证券交易委员会（SEC）的《加密资产条例》提案与《CLARITY 法案》仍在推进，但在本期窗口内均未形成新的正式立法。
第一部分仅涵盖本期新增事项。所有持续跟踪及即将到期的事项列于汇总表。
美国联邦 — 司法部支持 OpenAI 的合理使用抗辩（2026年9月1日）
司法部在合并审理的人工智能著作权诉讼中提交意见，主张以受著作权保护的文本训练大语言模型总体上构成合理使用。据行业媒体报道，这是美国政府首次在人工智能著作权案件中表明立场。政府的核心论点是：模型训练具有转换性——模型不会以可识别的形式再现原作品——且模型输出不会在市场上直接替代用于训练的书籍与文章，因此市场损害这一要素亦倾向于合理使用。该意见不对任何法院具有约束力，但在多起著作权诉讼即将进入判决动议与庭审阶段之际，明确释放了行政部门立场的重要信号。
实务评析
这是在天平上加了一枚砝码，而非终局裁判。司法部的观点具有说服力，将被援引于每一起在审的训练数据案件；但合理使用仍取决于个案事实，且训练语料的获取方式（如使用盗版数据集）是该意见无法弥补的另一重风险——须注意 Bartz 诉 Anthropic 案15亿美元和解的争点在于盗版复制件本身，而非训练行为。应提示人工智能开发方客户：联邦立场的转向并不等于可从盗版资料库取得数据，现在即应完善合法数据来源记录。对权利人客户而言，近期主战场将转向输出中的再现问题以及训练数据的获取方式。
美国联邦 — Andersen 诉 Stability AI 案开庭（2026年9月8日）
据报道，视觉艺术家集体诉讼 Andersen 诉 Stability AI 案已开庭，将首次由美国法院就“扩散模型本身即为其训练图像之侵权复制件”这一理论作出裁决。该案在侵权、合理使用及任何输出相似性问题上的裁决，将影响其他所有在审生成式人工智能案件的博弈筹码。在与法院原始记录核对确认前，相关表述均应视为初步信息。
加利福尼亚 — 州议会以约30项人工智能法案结束2026会期；州长签署期限为9月30日
加州议会在结束2026会期时通过了约三十项人工智能相关法案，内容涵盖聊天机器人安全、未成年人网络安全、心理健康与医疗用途、消费者权利以及前沿模型监管。其中最受关注的是《亚当法案》（SB 1119），该法案将强化面向青少年的人工智能聊天机器人的发布前测试与安全协议。州长纽森须在2026年9月30日前作出签署或否决决定；本会期法案目前均尚未生效。这些立法建立在该州现有的前沿人工智能制度（SB 53 / TFAIA）与陪伴型聊天机器人法（SB 243）之上，后两者均于2026年1月1日生效。
中国 — 最高人民法院发布首个全国性人工智能纠纷审理司法规则（2026年9月7日）
最高人民法院发布《关于审理人工智能相关纠纷案件的意见》，共五部分24条，是中国最高审判机关就人工智能作出的首个全国性裁判规则。该意见针对因人工智能训练与生成引发的人格权侵权，明确涵盖“AI 换脸”、声音克隆以及对已故者的“AI 复活”（近亲属享有相应权利），并涉及人工智能生成内容的责任、算法“大数据”差别定价（大数据杀熟）、自动驾驶、训练数据、开源软件以及人工智能生成证据的处理。有报道指出，该意见未就人工智能生成物的可著作权性问题作出结论。
实务评析
对于在中国经营或以中国为目标市场的客户而言，这是法院今后将实际适用的操作规则。任何克隆真实人物面部或声音、或生成可识别形象的服务，都需要有据可查的授权同意——最高法将责任直接与姓名权、肖像权和声音权相挂钩，并通过近亲属将保护延及已故者。应与国家网信办的拟人化人工智能陪伴服务规则（2026年7月15日生效）结合适用：同意、标识与防沉迷保障如今已由监管与司法双重强化。请在年底前重新审视同意流程、数据集来源以及个性化定价算法。
欧盟 — 首批 GPAI 系统性风险评估须于9月15日报送人工智能办公室
继欧盟《人工智能法》就通用人工智能的执法权于2026年8月2日生效之后，算力超过系统性风险门槛的通用模型提供者面临9月中旬的关键节点，须向欧洲人工智能办公室报送首批正式的系统性风险评估。该办公室预计将审查红队测试方法、能耗披露以及对标准化著作权训练数据摘要模板的使用情况。通用人工智能违规的罚款最高可达全球年营业额的3%或1500万欧元（以较高者为准）。
二、汇总表 — 新增与持续跟踪
标注“新增”的高亮行为本期事项；“持续”事项为此前已报道、保留以供跟踪。
2026.9.1 · 美国联邦
司法部提交意见支持 OpenAI 在著作权诉讼中的合理使用抗辩
状态：新增
实务影响：首个联邦立场；有说服力但无约束力——注意数据来源风险
2026.9.8 · 美国联邦
Andersen 诉 Stability AI 案开庭
状态：新增
实务影响：美国首例陪审团审议“模型即复制件”理论；判决待定
2026.9.7 · 中国
最高法《人工智能纠纷审理意见》（24条）
状态：新增
实务影响：深度伪造、声音/肖像、AI 生成证据纳入司法规则
至2026.9.30 · 加利福尼亚
约30项 AI 法案通过；待州长签署（含 SB 1119）
状态：新增
实务影响：关注签署情况；新的聊天机器人/未成年人保护及前沿义务或于2027年生效
2026.9.2 · 美国联邦
CFTC 市场监督部向 Electron Exchange 出具不采取行动函
状态：新增
实务影响：救济范围有限；关注更广泛的数字资产市场结构信号
2026.9.15 · 欧盟
首批 GPAI 系统性风险评估须报送人工智能办公室
状态：新增
实务影响：系统性风险模型提供者须报送；罚款最高达营业额3%
2026.9.15 · 美国联邦
《CLARITY 法案》参议院程序性表决（cloture）拟定
状态：持续
实务影响：市场结构法案；鉴于会期紧张，通过前景不明
2026.8.18 · 美国联邦
SEC 提出《加密资产条例》提案
状态：持续
实务影响：60天征求意见期开启；量身定制的加密发行/安全港制度
2026.8.2 · 欧盟
《人工智能法》通用人工智能执法权生效
状态：持续
实务影响：人工智能办公室可要求提交文件、评估模型、处以罚款
2026.7.15 · 中国
网信办拟人化人工智能陪伴服务规则生效
状态：持续
实务影响：算法备案、安全评估、身份披露、防沉迷、未成年人保护
2026.4.10 · 香港
金管局发放首批稳定币发行人牌照（汇丰、Anchorpoint）
状态：持续
实务影响：港元稳定币将于2026年下半年依《稳定币条例》推出
2026.2 · 中国
重申加密/稳定币/RWA 禁令（含跨境）
状态：持续
实务影响：未经批准不得在境外发行与人民币挂钩的稳定币
2027.1.1 · 纽约
《RAISE 法案》生效
状态：持续
实务影响：前沿模型安全协议；重大事件须在72小时内向金融服务局报告
2026.1.1 · 加利福尼亚
SB 53（TFAIA）与 SB 243 生效
状态：持续
实务影响：前沿模型透明度/事件报告；陪伴型聊天机器人义务
三、合规提示
致人工智能企业
现在即建立训练数据来源记录。司法部的合理使用立场有助于回应“训练”这一环节，但对语料的获取方式并无帮助。应保存能够证明数据集合法取得的记录，避免使用盗版资料——迄今最大额的人工智能著作权赔付，正是取决于这一区分，而非训练行为的转换性。
重新完善肖像、声音与人格化功能的授权同意。中国最高法《意见》与网信办陪伴服务规则均以“同意”为核心。任何面向中国市场提供的换脸、声音克隆、数字人或“复活”功能，都需要有据可查的授权、清晰的 AI 生成标识，以及针对未成年人和防止过度依赖的保障。
梳理对加州本会期法案的风险敞口。跟踪州长在9月30日前签署的约30项法案中的哪些，尤其是面向青少年的聊天机器人测试（SB 1119）及任何前沿模型修订，并将可能于2027年生效的日期列入日程，以便提前规划产品与安全测试的调整。
准备好欧盟 GPAI 合规证据。若贵司提供达到或超过系统性风险门槛的通用模型，应确保系统性风险评估、红队测试记录、能耗披露及著作权训练数据摘要已备妥，可供人工智能办公室查阅；如今违规罚款最高可达全球营业额的3%。
致 Web3 / 加密资产企业
就 SEC《加密资产条例》提案积极提交意见。征求意见期已开启。发行方与平台应评估拟议的发行豁免与投资合同安全港是否适用于其代币；如适用，应在截止日前准备意见——最终规则将影响美国的代币发行结构。
跟踪《CLARITY 法案》参议院程序性表决（9月15日）。SEC 与 CFTC 之间的市场结构管辖权分配仍未确定；CFTC 已表示，若该法案搁置，将自行推进数字资产规则。应制定在任一结果下均可运行的合规方案。
避免涉及中国的人民币挂钩稳定币与 RWA 活动。重申的禁令延及跨境行为；未经批准在境外发行与人民币挂钩的稳定币被禁止。应据此筛查涉华产品与交易对手。
为香港持牌稳定币市场做好布局。随着金管局发放首批发行人牌照、港元稳定币将于2026年下半年推出，寻求进入香港市场的机构应使其储备、赎回与反洗钱安排符合《稳定币条例》要求。
本简报仅供参考，不构成法律意见。如需人工智能法律合规或 Web3 与数字资产合规方面的协助，请与我们的律师联系。
English Edition
AI & Web3 Biweekly Legal Newsletter
Cross-Border AI and Fintech Practice — United States · New York · China · Hong Kong ·
Two Weeks in Review: August 27 – September 9, 2026
This Period — New Laws & Cases
A busy fortnight for AI law and a quieter one for digital-asset legislation.In the United States, the Department of Justice took the federal government's first-ever position in the AI copyright wars, filing a statement backing OpenAI's fair-use defense, while the Andersen v. Stability AI trial opened in San Francisco — the first U.S. jury to weigh the “model-as-copy” theory. California's Legislature closed its 2026 session having passed roughly thirty AI bills now awaiting the Governor's signature. In China, the Supreme People's Court issued its first national judicial rules on AI disputes, squarely addressing deepfakes and personality rights. In the EU, the first systemic-risk evaluations from general-purpose AI providers fall due mid-September. On the Web3 side, the SEC's Regulation Crypto Assets proposal and the CLARITY Act both remain in play but produced no new law inside the window.
Section 1 covers only what is new in this period. All ongoing and upcoming items are in the Summary Table.
U.S. Federal — DOJ backs OpenAI's fair-use defense in AI copyright litigation (Sept. 1, 2026)
The Department of Justice filed a statement in the consolidated AI copyright litigation asserting that training large language models on copyrighted text is generally fair use.Per trade-press reporting, this is the first time the U.S. government has taken a position in the AI copyright cases. The government's core argument is that training is transformative — the model does not reproduce source works in recognizable form — and that model outputs do not directly substitute in the market for the books and articles used in training, so the market-harm factor also favors fair use. The filing does not bind any court, but it is a significant signal of the Executive Branch's stance as multiple copyright suits head toward dispositive motions and trial.
PRACTITIONER'S COMMENTARY
This is a thumb on the scale, not a holding.The DOJ's view carries persuasive weight and will be cited in every pending training-data case, but fair use remains fact-specific and the acquisition of the training corpus (e.g., pirated datasets) is a separate exposure the brief does not cure — recall the $1.5B Bartz v. Anthropic settlement turned on pirated copies, not the training itself. Advise AI-developer clients that a favorable federal posture does not license sourcing from shadow libraries; document lawful data provenance now. For rights-holder clients, the near-term battleground shifts to reproduction in outputs and to how training data was obtained.
U.S. Federal — Andersen v. Stability AI goes to trial (Sept. 8, 2026)
Trial reportedly opened in Andersen v.Stability AI, the visual-artists' class action, teeing up what would be the first U.S. verdict testing the theory that a diffusion model is itself an infringing copy of its training images. A verdict here — on infringement, fair use, and any output-similarity findings — will shape the leverage in every other pending generative-AI case. Treat all characterizations as preliminary until the docket and any verdict are confirmed against primary court records.
California — Legislature ends 2026 session with ~30 AI bills; Governor's deadline Sept. 30
California's Legislature adjourned its 2026 session having passed roughly thirty AI-related measures spanning chatbot safety, children's online safety, mental-health and medical uses, consumer rights, and frontier-model oversight.Among the most watched is Adam's Law (SB 1119), which would strengthen pre-release testing and safety protocols for AI chatbots that interact with teenagers. Governor Newsom has until September 30, 2026 to sign or veto; none of the new session bills is law yet. This builds on the state's existing frontier-AI regime (SB 53 / TFAIA) and companion-chatbot law (SB 243), both effective January 1, 2026.
China — Supreme People's Court issues first national judicial rules on AI disputes (Sept. 7, 2026)
The Supreme People's Court released its Opinions on the Trial of Cases Involving Artificial Intelligence Disputes — 24 articles across five parts, the first national adjudication rules on AI from China's highest court.The Opinions address personality-rights infringement from AI training and generation, expressly covering AI face-swapping, voice cloning, and “AI resurrection” of the deceased (with standing for close relatives), alongside AI-generated content liability, algorithmic “big-data” discriminatory pricing, autonomous driving, training data, open-source software, and the treatment of AI-generated evidence. Reporting notes the Opinions stop short of resolving the copyrightability of AI-generated works.
PRACTITIONER'S COMMENTARY
For clients operating in or targeting China, this is the practical rulebook courts will now apply.Any service that clones a real person's face or voice, or generates an identifiable likeness, needs documented consent — the SPC ties liability directly to name, portrait, and voice rights, and extends protection to the deceased through their relatives. Pair this with the CAC's anthropomorphic-AI companion rules (effective July 15, 2026): consent, labeling, and anti-addiction safeguards are now reinforced by both regulator and judiciary. Revisit consent flows, dataset provenance, and personalized-pricing algorithms before year-end.
EU — First GPAI systemic-risk evaluations due to the AI Office (Sept. 15, 2026)
Following the AI Act's GPAI enforcement powers taking effect on August 2, 2026, providers of general-purpose models above the systemic-risk compute threshold face a mid-September milestone to submit their first formal systemic-risk evaluations to the European AI Office, which is expected to review red-teaming methodology, energy-consumption disclosures, and use of the standardized copyright training-data summary template.Fines for GPAI non-compliance can reach 3% of global annual turnover or €15 million, whichever is higher.
Summary Table — New and Ongoing
Shaded rows marked “NEW” are from this period. “Ongoing” items were reported earlier and are kept for tracking.
Sept. 1, 2026 · US Fed.
DOJ statement backing OpenAI fair-use defense in AI copyright suits
Status：NEW
What it means：First federal position; persuasive, not binding — check provenance risk
Sept. 8, 2026 · US Fed.
Andersen v. Stability AI trial opens
Status：NEW
What it means：First U.S. jury test of “model-as-copy” theory; verdict pending
Sept. 7, 2026 · China
SPC Opinions on AI disputes (24 articles)
Status：NEW
What it means：Deepfakes, voice/likeness, AI-generated evidence now governed by court rules
By Sept. 30, 2026 · California
~30 AI bills passed; await Governor's signature (incl. SB 1119)
Status：NEW
What it means：Watch signings; new chatbot/child-safety and frontier duties may take effect 2027
Sept. 2, 2026 · US Fed.
CFTC DMO no-action letter to Electron Exchange
Status：NEW
What it means：Narrow relief; monitor for broader digital-asset market-structure signals
Sept. 15, 2026 · EU
First GPAI systemic-risk evaluations due to AI Office
Status：NEW
What it means：Systemic-risk model providers must file; fines up to 3% turnover
Sept. 15, 2026 · US Fed.
CLARITY Act Senate cloture vote scheduled
Status：Ongoing
What it means：Market-structure bill; passage odds uncertain given tight calendar
Aug. 18, 2026 · US Fed.
SEC proposes Regulation Crypto Assets
Status：Ongoing
What it means：60-day comment window open; tailored crypto offering/safe-harbor regime
Aug. 2, 2026 · EU
AI Act GPAI enforcement powers in force
Status：Ongoing
What it means：AI Office may demand docs, evaluate models, impose fines
July 15, 2026 · China
CAC anthropomorphic-AI companion rules effective
Status：Ongoing
What it means：Filing, security assessment, disclosure, anti-addiction, minors protection
Apr. 10, 2026 · Hong Kong
HKMA grants first stablecoin issuer licences (HSBC, Anchorpoint)
Status：Ongoing
What it means：HKD stablecoins launching 2H 2026 under Stablecoins Ordinance
Feb. 2026 · China
Crypto/stablecoin/RWA ban reaffirmed (incl. cross-border)
Status：Ongoing
What it means：No RMB-linked stablecoin issuance abroad without approval
Jan. 1, 2027 · New York
RAISE Act takes effect
Status：Ongoing
What it means：Frontier-model safety protocols; 72-hour incident reporting to DFS
Jan. 1, 2026 · California
SB 53 (TFAIA) & SB 243 in effect
Status：Ongoing
What it means：Frontier transparency/incident reporting; companion-chatbot duties
Call for Compliance
For AI Companies
Document training-data provenance now.The DOJ's fair-use posture helps on the training question but does nothing for how the corpus was acquired. Maintain records showing datasets were lawfully sourced and avoid pirated collections — that distinction, not transformation, drove the largest AI copyright payout to date.
Re-paper consent for likeness, voice, and persona features.China's SPC Opinions and the CAC companion rules both hinge on consent. Any face-swap, voice-clone, digital-human, or “resurrection” feature offered into China needs documented authorization, clear AI-generated labeling, and safeguards for minors and against over-reliance.
Map exposure to California's incoming session bills.Track which of the ~30 measures the Governor signs by September 30, especially teen-chatbot testing (SB 1119) and any frontier-model amendments, and diarize likely 2027 effective dates so product and safety-testing changes are scoped in advance.
Stand up EU GPAI compliance evidence.If you provide a general-purpose model at or above the systemic-risk threshold, ensure your systemic-risk evaluation, red-teaming records, energy disclosures, and copyright training-data summary are ready for the AI Office; non-compliance now carries fines up to 3% of global turnover.
For Web3 / Crypto Companies
Engage on the SEC's Regulation Crypto Assets proposal.The comment window is open. Issuers and platforms should assess whether the proposed offering exemptions and investment-contract safe harbor fit their token and, if so, prepare comments before the deadline — the final rule will shape U.S. token-distribution structures.
Track the CLARITY Act cloture vote (Sept.15). Market-structure allocation between the SEC and CFTC remains unsettled; the CFTC has signaled it will move on its own digital-asset rules if the bill stalls. Build compliance plans that work under either outcome.
Steer clear of RMB-linked stablecoin and RWA activity touching China.The reaffirmed ban reaches cross-border conduct; offshore issuance of yuan-linked stablecoins without approval is prohibited. Screen products and counterparties with China nexus accordingly.
Position for Hong Kong's licensed stablecoin market.With the first HKMA issuer licences granted and HKD stablecoins launching in 2H 2026, firms seeking Hong Kong distribution should align reserve, redemption, and AML arrangements with the Stablecoins Ordinance.
This briefing is prepared for informational purposes only and does not constitute legal advice. If you are seeking AI legal governance or Web3 & digital-assets compliance, please reach out to our lawyers.
