US rules for models, data and AI in 2026: what SR 26-2 is, and what surrounds it2026 年美国对模型、数据和 AI 的规则：SR 26-2 是什么，周围还有什么
A reading of the newest US model-risk guidance, and a map of the federal, financial-sector, standards, state and data documents around it, with a clear line between what I read at the source and what I only saw summarised.
解读最新的美国模型风险指引，并梳理它周围的联邦、金融行业、标准、州和数据文件；清楚区分哪些是我在原文读到的，哪些只是看到别人的摘要。
- Date · 日期
- 2026-10-04
- Status · 状态
- Draft — SR 26-2, EO 14365 and the March 2026 framework read in full; many other items from summaries (flagged)草稿 — SR 26-2、EO 14365 和 2026 年 3 月框架已读全文；其他不少条目来自摘要（已标注）
- Author · 作者
- 🌱
- Tags · 标签
- regulation, US, model risk, SR 26-2, AI policy, NIST, state law, data
- Question · 研究问题
- Is SR 26-2 the newest US model and data rule, what else governs data, ML and AI in the US, and what does each document actually require?SR 26-2 是不是美国最新的模型与数据监管文件？美国还有哪些关于数据、ML 和 AI 的文件？每份实际要求什么？
- Addresses · 对应
- Q8, overlaps Q7, Q3 · H1, H5
Short answer先说结论
Is SR 26-2 the newest? It is the newest US banking-supervisor guidance on model risk (issued 2026-04-17 by the Fed, FDIC and OCC; it supersedes SR 11-7 and SR 21-8) [1][2]. It is not a general US rule for models, data or AI. It is guidance, not law: it says it sets no enforceable standards and non-compliance will not by itself draw supervisory criticism [1]. It is aimed mainly at banks above $30 billion in assets, and it excludes generative and agentic AI [1]. The agencies have said they plan a request for information on model risk and AI, including generative and agentic AI [3][4]; I found no source showing it has been issued yet.
SR 26-2 是最新的吗？它是美国银行监管机构关于模型风险的最新指引（2026-04-17 由美联储、FDIC、OCC 发布，取代 SR 11-7 和 SR 21-8）[1][2]。但它不是针对模型、数据或 AI 的通用美国规则。它是指引，不是法律：文件自己说不设可执行的标准，不遵守本身不会招致监管批评 [1]。它主要面向资产超过 300 亿美元的银行，并把生成式与智能体 AI 排除在外 [1]。监管机构表示计划就模型风险与 AI（含生成式和智能体）发布征求意见稿 [3][4]；我没有找到来源表明它已经发布。
What else is there? There is no comprehensive US federal AI or model law. The picture is layers: a federal executive policy that favours light regulation and pushes against state laws [5][6]; sector regulators and exam priorities; voluntary NIST standards [13][14]; a growing set of state laws (Colorado, California, Texas and others); and older sector and data statutes that still apply to ML models. Several of the newest items are proposals, voluntary frameworks or litigation, not binding rules.
还有什么？美国没有综合性的联邦 AI 或模型法律。现状是分层的：偏向轻监管并推动压过州法的联邦行政政策 [5][6]；行业监管机构及其检查重点；自愿性的 NIST 标准 [13][14]；越来越多的州法（科罗拉多、加州、德州等）；以及仍然适用于 ML 模型的旧行业法规和数据法规。最新的几项多是提案、自愿框架或诉讼，不是有约束力的规则。
Map of the documents文件地图
| Layer层 | Document文件 | Binding?约束力 | What I read我读到的程度 |
|---|---|---|---|
| Bank supervisors银行监管机构 | SR 26-2 / OCC Bulletin 2026-13 (2026-04-17) [1][2][3] | Guidance; no enforceable standards指引；无可执行标准 | Attachment: full text. OCC page: summariser附件：全文。OCC 页面：摘要工具 |
| Planned agencies' request for information on model risk and AI监管机构计划发布的模型风险与 AI 征求意见稿 | Not a rule; not found issued不是规则；未找到已发布 | Only the statement that it is planned [3][4]仅见“计划发布”的表述 [3][4] | |
| Older issuances that SR 26-2 / OCC 2026-13 replace (SR 11-7, SR 21-8, OCC 2011-12, 1997-24, 2021-19, the handbook booklet)被 SR 26-2 / OCC 2026-13 取代的旧文件（SR 11-7、SR 21-8、OCC 2011-12、1997-24、2021-19、手册册子） | Rescinded已废止 | Via OCC page summary [3]来自 OCC 页面摘要 [3] | |
| Federal executive policy联邦行政政策 | EO 14365, 2025-12-11 [5] | Directs agencies; creates no private rights指令各机构；不创设私人权利 | Full text全文 |
| National Policy Framework, legislative recommendations, 2026-03-20 [6]国家政策框架立法建议，2026-03-20 [6] | Non-binding recommendations to Congress对国会的非约束性建议 | Full text全文 | |
| EO "Promoting Advanced AI Innovation and Security", 2026-06-02 [8]行政令“促进先进 AI 创新与安全”，2026-06-02 [8] | Voluntary frontier-model framework自愿性前沿模型框架 | Fact sheet via summariser (low confidence)经摘要工具读取的情况说明（低可信度） | |
| EO replacing the term "AI" with "Super Intelligence" in the executive branch, 2026-09-29 [7]行政令：在行政部门以“Super Intelligence”取代“AI”一词，2026-09-29 [7] | Terminology only, per the fact sheet据情况说明，仅涉及术语 | Fact sheet read; order text not read读了情况说明；未读行政令文本 | |
| FTC proposed policy statement on AI accuracy, 2026-07-01 [9] | Proposed; not final (as of my sources)拟议稿；尚未定稿（据我的来源） | Opening sections read (about a quarter)读了开头部分（约四分之一） | |
| Standards标准 | NIST AI RMF 1.0, AI 600-1 (GenAI profile) [13] | Voluntary自愿 | NIST page via summariserNIST 页面经摘要工具 |
| NIST CAISI AI Agent Standards Initiative, 2026-02 [14] | Voluntary自愿 | NIST page via summariserNIST 页面经摘要工具 | |
| States各州 | Colorado SB 26-189 (signed 2026-05-14) [11][12] | Law; key duties from 2027-01-01法律；主要义务自 2027-01-01 | Legislature page via summariser立法机构页面经摘要工具 |
| California CCPA regulations on ADMT, risk assessments, audits [15] | Regulation; in force 2026-01-01法规；2026-01-01 生效 | CPPA page via summariser (thin)CPPA 页面经摘要工具（信息很少） | |
| Texas, California statutes, others德州、加州的其他法律，及其他州 | Laws法律 | Search summaries only (see leads table)仅见搜索摘要（见线索表） |
Reading SR 26-2 in detail详细解读 SR 26-2
This section is based on the full text of the attachment [1]. The stage labels (Require … Monitor) are the repository's own vocabulary, not the document's.
本节依据附件全文 [1]。阶段标签（Require … Monitor）是本仓库自己的词汇，不是文件里的。
| # | Finding · 发现 | Cite |
|---|---|---|
| S1 | Legal force. The text says it "does not set forth enforceable standards or prescriptive requirements" and non-compliance will not result in supervisory criticism. A footnote adds that supervisory action may still result from violations of law or unsafe or unsound practices stemming from insufficient management of model risk. So the guidance is soft, but the underlying duty to manage model risk is not.法律效力。文本说它“不设可执行的标准或规定性要求”，不遵守不会招致监管批评。脚注补充：若因模型风险管理不足而违法或构成不安全不稳健的做法，仍可能引发监管行动。所以指引本身是“软”的，但背后管理模型风险的义务并不软。 | [1] |
| S2 | Who it is for. Most relevant to banking organisations with over $30 billion in total assets; at or below that size, models are typically covered by internal practices, though the guidance may be relevant where model exposure is significant or activities go beyond traditional community banking.适用对象。主要针对总资产超过 300 亿美元的银行机构；规模在此以下的，模型一般由内部实践管理，但若模型风险敞口大，或业务超出传统社区银行范围，指引可能仍相关。 | [1] |
| S3 | What counts as a model. A "complex quantitative method, system, or approach that applies statistical, economic, or financial theories to process input data into quantitative estimates". Simple spreadsheet arithmetic and deterministic rule-based processes with no such theory behind them are excluded. A classical ML model (for example a credit or fraud scorer) fits this definition on my reading; that is my interpretation.什么算模型。“运用统计、经济或金融理论，把输入数据处理成量化估计的复杂量化方法、系统或做法”。简单的电子表格算术，以及背后没有这类理论的确定性规则流程被排除在外。按我的理解，经典 ML 模型（例如信用或欺诈评分）符合这个定义；这是我的解读。 | [1] |
| S4 | AI scope. Footnote 3: generative AI and agentic AI models are "novel and rapidly evolving" and are not within the scope; the principles "apply to traditional statistical and quantitative models and non-generative, non-agentic AI models". The footnote adds that a bank's own risk management and governance should guide controls for anything not covered. So non-generative ML is in scope; LLMs and agents are not, and are left to general governance.AI 范围。脚注 3：生成式 AI 和智能体 AI 模型“新颖且快速演变”，不在范围内；这些原则“适用于传统统计与量化模型，以及非生成式、非智能体的 AI 模型”。脚注还说，未覆盖的内容应由银行自己的风险管理和治理来确定控制措施。所以非生成式 ML 在范围内；大模型和智能体不在，留给一般治理。 | [1] |
| S5 | Risk logic. Model risk depends on inherent risk (assumptions, complexity, input quality, data constraints), exposure (importance of the output to decisions, for example portfolio size) and purpose (models used for regulatory requirements or financial risk exposure are generally riskier). Exposure plus purpose gives materiality, which sets the rigour. Immaterial models still need to be identified and monitored in case they become material.风险逻辑。模型风险取决于固有风险（假设、复杂度、输入质量、数据限制）、敞口（输出对决策的重要性，例如组合规模）和用途（用于满足监管要求或管理金融风险敞口的模型通常风险更高）。敞口加用途决定重要性，重要性决定严格程度。不重要的模型也要被识别并监控，以防日后变得重要。 | [1] |
| S6 | Aggregate risk. Risk should be assessed individually and in aggregate, where aggregate risk reflects "interactions and dependencies among models; reliance on common assumptions, data, or methodologies". This is the sentence in the document closest to a data-dependency or lineage expectation, though the text never uses the word lineage.聚合风险。风险应当既按单个模型、也按整体来评估，其中整体风险反映“模型之间的相互作用与依赖；对共同假设、数据或方法的依赖”。这是文件中最接近“数据依赖或血缘”要求的一句，不过全文从未使用“血缘”一词。 | [1] |
| S7 | Development and use (Require, Build). Start from a clear statement of purpose; testing may include "a critical assessment of data quality, relevance, and inputs"; its rigour is proportional to complexity and materiality. Using a model beyond its intended purpose calls for additional analysis and review of controls.开发与使用（Require、Build）。从清晰的用途说明开始；测试可包括“对数据质量、相关性和输入的批判性评估”，严格程度与复杂度和重要性相称。把模型用于原定用途之外，需要额外分析并复查控制措施。 | [1] |
| S8 | Validation (Validate, Approve). Validation generally happens before first use; if an urgent need forces earlier use, sound practice is to pay more attention to limitations, tell stakeholders, and add controls such as limits or closer monitoring. Components: conceptual soundness (design, assumptions, data selection; interpretability or benchmarking where theory is hard), outcomes analysis, and ongoing monitoring. The text says validation quality depends on the rigour of the review, not on organisational structure.验证（Validate、Approve）。验证一般在首次使用前完成；若因紧急需要提前使用，良好做法是更关注局限、告知相关方，并增加限额或加强监控等控制。组成部分：概念合理性（设计、假设、数据选择；理论难以评估时用可解释性或基准比较）、结果分析和持续监控。文本称验证质量取决于审查的严谨程度，而不取决于组织架构。 | [1] |
| S9 | Monitoring (Monitor). Check whether the model still performs as expected given changes in products, exposures, activities, clients, data relevance or market conditions; a model that no longer performs may need overlays, adjustment or redevelopment under the bank's policy.监控（Monitor）。检查模型在产品、敞口、业务、客户、数据相关性或市场条件变化后是否仍按预期运行；不再达标的模型，可能需要按银行政策做叠加调整、修正或重新开发。 | [1] |
| S10 | Governance, inventory, documentation. Clear roles and accountability, with attention to conflicts of interest between development and validation lines. Internal audit evaluates whether model risk management is rigorous and effective and does not duplicate it. An effective inventory holds enough information to understand model risk "at the individual and aggregate levels". Documentation supports continuity and the tracking of recommendations, responses, exceptions and remediation. The text is short on detail about both; it gives no required fields.治理、清单、文档。角色与责任清晰，关注开发线与验证线之间的利益冲突。内部审计评估模型风险管理是否严谨有效，不重复其工作。有效的模型清单应含足够信息，以便在“单个和整体层面”理解模型风险。文档支持业务连续，并用于跟踪建议、回应、例外和整改。这两部分文本都比较简短，没有规定必填字段。 | [1] |
| S11 | Vendor and third-party models. The principles still apply even where code, data or method are proprietary. Sound practice includes validating vendor products, understanding their conceptual soundness, design, development data and performance, ongoing monitoring and outcomes analysis, and documenting and evaluating any customisation.供应商及第三方模型。即使代码、数据或方法是专有的，这些原则仍然适用。良好做法包括验证供应商产品、理解其概念合理性、设计、开发数据和表现，做持续监控与结果分析，并记录和评估任何定制。 | [1] |
| S12 | OCC companion. OCC Bulletin 2026-13 (same date) rescinds the OCC's 2011 model risk management bulletin, the Comptroller's Handbook booklet, Bulletin 1997-24 on credit scoring and the 2021 BSA/AML statement; it repeats the $30 billion focus and says community banks with significant model exposure may find it relevant. The summaries I read mention no effective date or transition.OCC 配套文件。OCC 公告 2026-13（同日）废止 OCC 2011 年模型风险管理公告、《监理官手册》相关册子、关于信用评分的 1997-24 号公告和 2021 年 BSA/AML 声明；同样以 300 亿美元为重点，并称有重大模型敞口的社区银行也可能相关。我读到的摘要里没有提到生效日期或过渡安排。 | [3] [4] |
Reading it for the MDLC. SR 26-2 is a lifecycle document that does not mention tooling. Mapped to the repository's seven stages: Require = statement of purpose; Build = data selection and testing including data quality; Validate/Approve = pre-use validation with a documented exception path; Deploy/Monitor = monitoring for changed data relevance and use beyond purpose. Two points matter for the control-layer question (Q7): the aggregate-risk sentence (S6) makes shared data and assumptions across models something the bank should assess, and the inventory (S10) is where that evidence naturally lives. The text asks for the outcome, not for a catalog or lineage tool. For generative and agentic systems the guidance gives no lifecycle at all, which is consistent with H5.
放到 MDLC 里读。SR 26-2 是一份生命周期文件，没有提到任何工具。对应本仓库的七个阶段：Require = 用途说明；Build = 数据选择与包括数据质量在内的测试；Validate/Approve = 使用前验证并有文档化的例外路径；Deploy/Monitor = 监控数据相关性变化和超出原定用途的使用。有两点与控制层问题（Q7）有关：聚合风险这句话（S6）使“跨模型共享的数据与假设”成为银行应当评估的对象，而模型清单（S10）正是这类证据自然存放的地方。文本要求的是结果，并不要求数据目录或血缘工具。对于生成式和智能体系统，指引完全没有给出生命周期，这与 H5 一致。
Federal executive policy: the direction of travel联邦行政政策：大方向
| # | Finding · 发现 | Cite |
|---|---|---|
| E1 | EO 14365 (2025-12-11) states a policy of a "minimally burdensome national policy framework for AI". It tells the Attorney General to set up, within 30 days, an AI Litigation Task Force to challenge state AI laws (on interstate-commerce, preemption or other grounds); the Commerce Secretary to publish within 90 days an evaluation of "onerous" state laws, including ones that require models to alter truthful outputs or compel disclosures that may violate the First Amendment; Commerce to condition BEAD non-deployment funds on that evaluation; the FCC to consider a federal reporting and disclosure standard that preempts conflicting state laws; and the FTC to issue a policy statement on state laws that require altering truthful outputs. It names Colorado's algorithmic-discrimination law as an example. It says legislation should not preempt state laws on child safety, AI compute and data-center infrastructure, or state procurement. It creates no enforceable rights.EO 14365（2025-12-11）宣示“负担最小的国家 AI 政策框架”。它要求：司法部长在 30 天内设立 AI 诉讼工作组，挑战各州 AI 法律（理由包括州际商业、联邦优先或其他）；商务部长在 90 天内发布对“苛刻”州法的评估，包括要求模型改变真实输出的法律，或可能因强制披露而违反第一修正案的法律；商务部据此对 BEAD 的非部署资金设置条件；FCC 考虑制定能优先于冲突州法的联邦报告与披露标准；FTC 就要求改变真实输出的州法发布政策声明。它以科罗拉多州的算法歧视法为例。它说立法建议不应优先于儿童安全、AI 算力与数据中心基础设施、州政府采购方面的州法。它不创设可执行的权利。 | [5] |
| E2 | The March 2026 legislative recommendations (to Congress, not binding) say Congress should preempt state AI laws that impose "undue burdens", should not create any new federal AI rulemaking body but rely on existing sector regulators and industry-led standards, and should add regulatory sandboxes and AI-ready federal datasets. They list what preemption should not cover: generally applicable state laws on child protection, fraud and consumer protection, zoning, and a state's own use of AI. They say states should not regulate AI development, should not unduly burden lawful use of AI, and should not penalise developers for a third party's unlawful conduct. They contain no model-documentation or testing mandate that I found.2026 年 3 月的立法建议（给国会的，无约束力）说：国会应优先于那些施加“不当负担”的州 AI 法律；不应设立新的联邦 AI 规则制定机构，而应依靠现有行业监管机构和行业主导的标准；应增设监管沙盒和 AI 可用的联邦数据集。它们列出优先权不应涵盖的内容：适用于一般情形的州儿童保护、反欺诈和消费者保护法律，分区规划，以及州自身对 AI 的使用。它们还说州不应监管 AI 开发，不应对合法使用 AI 施加不当负担，也不应因第三方的违法行为惩罚开发者。我没有在其中发现模型文档或测试方面的强制要求。 | [6] |
| E3 | Implementation, as far as my sources show: the FTC published a proposed policy statement dated 2026-07-01 that says steering AI outputs away from what consumers reasonably expect can be deceptive under Section 5 of the FTC Act, even if done to comply with a state law, and names Colorado's revised Act. It is a proposal for comment. DOJ moved on 2026-04-24 to join xAI's lawsuit against Colorado's original AI Act. A law-firm article says a magistrate judge blocked enforcement of that Act on 2026-04-27. I did not find a published Commerce evaluation; one law-firm source said it had not been released by March, and I have no later information.据我的来源，落实情况是：FTC 发布了日期为 2026-07-01 的政策声明拟议稿，称把 AI 输出引向消费者合理预期之外可能构成《FTC 法》第 5 条下的欺骗，即使是为了遵守州法，并点名了科罗拉多州修订后的法案。这只是征求意见稿。司法部于 2026-04-24 申请加入 xAI 起诉科罗拉多州原 AI 法的诉讼。一篇律所文章称，一名治安法官在 2026-04-27 阻止了该法的执行。我没有找到已发布的商务部评估；一篇律所资料称到 3 月尚未发布，之后的情况我没有信息。 | [9] [10] [12] |
| E4 | Other 2026 federal acts that I saw only in fact sheets: a 2026-06-02 order promoting a voluntary framework for "covered frontier models" with a statement that nothing authorises mandatory licensing or pre-clearance [8]; and a 2026-09-29 order that replaces the term "Artificial Intelligence" with "Super Intelligence" (SI) in executive-branch documents and asks the science adviser to propose a federal definition [7]. The September fact sheet announces no regulatory change. Practical effect: federal documents from now on may say "SI", so searches should use both terms.我只在情况说明中看到的其他 2026 年联邦行动：2026-06-02 的一项行政令，推动针对“受覆盖前沿模型”的自愿框架，并声明不授权任何强制许可或预先批准 [8]；以及 2026-09-29 的一项行政令，在行政部门文件中以“Super Intelligence”（SI）取代“Artificial Intelligence”，并要求科技顾问提出联邦定义 [7]。九月的情况说明没有宣布任何监管变化。实际影响：今后的联邦文件可能写 “SI”，所以检索应同时使用两个词。 | [7] [8] |
Method note: my summariser tool produced a confident but wrong summary of the March 2026 recommendations (it said they preserve state authority over employment law, that a new coordinating agency was possible, and that developers should document training data). The PDF text says the opposite on the agency point and contains no such documentation mandate. I used only the text I extracted from the PDF. Summaries from that tool elsewhere in this note should be treated as leads.
方法说明：我的摘要工具对 2026 年 3 月立法建议给出了自信但错误的摘要（称其保留州的就业法权限、可能设立协调性新机构、开发者应记录训练数据）。PDF 原文在机构问题上说的恰恰相反，也没有这类文档要求。我只采用了从 PDF 提取的文本。本笔记中其他来自该工具的摘要，都应视为线索。
Other financial and sector regulators (leads, not read at source)其他金融与行业监管机构（线索，未读原文）
The items below come from search-result summaries and law-firm or vendor commentary. I have not read the primary documents, so none is stated as fact here. The right-hand column names what to read next.
以下条目来自搜索结果摘要以及律所或厂商评论。我没有读原始文件，所以这里都不作为事实陈述。右栏写明接下来该读什么。
| Lead (as reported)线索（据报道） | Read next下一步该读 |
|---|---|
| CFPB withdrew 67 guidance documents on 2025-05-12, including Circulars 2022-03 and 2023-03 on adverse-action notices for credit decisions made with complex algorithms. Commentary says the underlying duty in Regulation B (specific reasons for adverse action) and fair-lending disparate-impact exposure remain.据报道，CFPB 于 2025-05-12 撤回了 67 份指引，其中包括关于使用复杂算法做信贷决定时不利行动通知的 2022-03 和 2023-03 号通告。评论称，Regulation B 中的底层义务（说明不利行动的具体原因）和公平借贷中的差别影响风险仍然存在。 | 12 CFR 1002.9; CFPB withdrawal notice |
| FINRA's 2026 Annual Regulatory Oversight Report (2025-12-09) reportedly adds a GenAI section and, for the first time, discusses AI agents, tying supervision to Rule 3110 and business continuity to Rule 4370, and warning about overstated AI claims.据报道，FINRA《2026 年度监管监督报告》（2025-12-09）新增生成式 AI 一节，并首次讨论 AI 智能体，把监督与第 3110 号规则、业务连续性与第 4370 号规则联系起来，并提醒不要夸大 AI 宣传。 | FINRA 2026 Annual Regulatory Oversight Report (PDF on finra.org) |
| SEC: reportedly no AI-specific rule; "AI washing" cases and 2026 exam priorities; Regulation S-P amendments with compliance dates 2025-12-03 (larger firms) and 2026-06-03 (smaller).据报道，SEC：没有专门针对 AI 的规则；有“AI 洗白”案件和 2026 年检查重点；Regulation S-P 修订的合规日期为 2025-12-03（大型机构）和 2026-06-03（小型机构）。 | SEC FY2026 examination priorities; enforcement releases |
| NAIC Model Bulletin on insurers' use of AI (adopted 2023-12): one vendor blog says 25 states had formally adopted it by July 2026. It reportedly expects an AI governance programme, bias testing and third-party vendor oversight.NAIC 保险公司使用 AI 示范公告（2023-12 通过）：一篇厂商博客称到 2026 年 7 月已有 25 个州正式采纳。据报道，它要求建立 AI 治理计划、做偏差测试并监督第三方供应商。 | NAIC adoption map (naic.org) |
| FDA: reportedly a final guidance on predetermined change control plans for AI-enabled device software, and a draft lifecycle-management guidance dated 2025-01-07; later status unknown.据报道，FDA：已有关于 AI 医疗器械软件预设变更控制计划的最终指南，以及 2025-01-07 的生命周期管理指南草案；之后的状态未知。 | FDA guidance pages |
| DOJ Data Security Program, 28 CFR Part 202: reportedly effective 2025-04-08, restricting bulk transfers of sensitive US personal data to six countries of concern, with due-diligence and audit duties from 2025-10-06. It matters for training data and vendor chains.据报道，司法部数据安全计划（28 CFR Part 202）：2025-04-08 生效，限制向六个受关注国家大量传输美国敏感个人数据，尽职调查和审计义务自 2025-10-06 起。它与训练数据和供应商链有关。 | 28 CFR Part 202; DOJ NSD page |
| GLBA, FCRA, ECOA, HIPAA and state privacy laws: older statutes that apply to the data used in ML models and to model-driven decisions. I did not read them for this note.GLBA、FCRA、ECOA、HIPAA 和各州隐私法：适用于 ML 模型所用数据和模型驱动决策的旧法规。本笔记没有读它们。 | Statute and regulator pages法条与监管机构页面 |
Standards: NIST标准：NIST
| # | Finding · 发现 | Cite |
|---|---|---|
| N1 | NIST's AI RMF page says AI RMF 1.0 (2023-01-26) is voluntary and "being revised as part of the White House AI Action Plan". The Generative AI Profile (NIST AI 600-1) was published 2024-07-26. A concept note for a critical-infrastructure profile appeared 2026-04-07.NIST 的 AI RMF 页面说，AI RMF 1.0（2023-01-26）是自愿性的，并“正在作为白宫 AI 行动计划的一部分修订”。生成式 AI 画像（NIST AI 600-1）发布于 2024-07-26。关键基础设施画像的概念说明出现在 2026-04-07。 | [13] |
| N2 | NIST's Center for AI Standards and Innovation launched an AI Agent Standards Initiative in February 2026 with three pillars: industry-led standards, community-led protocols, and research on agent security and identity. Items include a request for information on agent security (due March 9), and a concept paper on agent identity and authorisation (due April 2). All outputs are described as voluntary.NIST 的 AI 标准与创新中心于 2026 年 2 月启动 AI 智能体标准倡议，有三大支柱：行业主导的标准、社区主导的协议，以及智能体安全与身份的研究。其中包括关于智能体安全的征求意见（3 月 9 日截止），以及关于智能体身份与授权的概念文件（4 月 2 日截止）。所有成果都被描述为自愿性的。 | [14] |
NIST is where the agent-governance vocabulary (identity, authorisation, logging) is being written. That fits Q2 and Q6, and it fills the gap SR 26-2 leaves for agents, but only as voluntary guidance.
智能体治理的词汇（身份、授权、日志）正在 NIST 被写出来。这与 Q2、Q6 相符，也补上了 SR 26-2 对智能体留下的空白，但只是自愿性的指引。
States各州
| # | Finding · 发现 | Cite |
|---|---|---|
| T1 | Colorado: SB 26-189, "Automated Decision-Making Technology", was signed 2026-05-14 and replaces SB 24-205. Per the legislature's page: developers must give deployers technical documentation (intended uses, categories of training data, known limitations, instructions for oversight); deployers must notify consumers at the point of interaction and, after an adverse decision, give a plain-language explanation within 30 days; consumers can request their data, correction, and human review. Covered decisions include education, employment, housing, financial services, insurance, healthcare and government benefits. The Attorney General enforces as a deceptive trade practice, with a 60-day notice-and-cure period before 2030, and must adopt disclosure rules by 2027-01-01; the key requirements start 2027-01-01. The original SB 24-205 never took effect (a law-firm article says enforcement was blocked 2026-04-27 and the bill was replaced), and, per the same article, the replacement dropped the impact-assessment and algorithmic-discrimination-assessment duties.科罗拉多州：SB 26-189《自动化决策技术》于 2026-05-14 签署，取代 SB 24-205。据立法机构页面：开发者须向部署者提供技术文档（预期用途、训练数据类别、已知局限、监督说明）；部署者须在交互时通知消费者，并在出现不利决定后 30 天内给出通俗解释；消费者可请求查看自己的数据、更正，以及人工复核。涵盖的决定包括教育、就业、住房、金融服务、保险、医疗和政府福利。总检察长按欺骗性贸易行为执法，2030 年前有 60 天的通知与整改期，并须在 2027-01-01 前制定披露规则；主要要求自 2027-01-01 起生效。原 SB 24-205 从未生效（一篇律所文章称其执行在 2026-04-27 被阻止并被替换），同一篇文章称，替换后的法律删去了影响评估和算法歧视评估的义务。 | [11] [12] |
| T2 | California: the CPPA's regulations on automated decision-making technology (ADMT), risk assessments and cybersecurity audits were approved 2025-09-22 and took effect 2026-01-01. The agency page I read says ADMT rules cover pre-use notice, opt-out and access rights, but gives no compliance dates beyond the effective date. Other pages I saw in search results (not read at source) give ADMT compliance from 2027-01-01 and risk-assessment attestations by 2028-04-01; treat those dates as leads.加州：CPPA 关于自动化决策技术（ADMT）、风险评估和网络安全审计的法规于 2025-09-22 获批，2026-01-01 生效。我读到的机构页面称 ADMT 规则涵盖事前通知、选择退出和访问权，但除生效日期外没有给出合规日期。我在搜索结果中看到的其他页面（未读原文）给出 ADMT 自 2027-01-01 起合规、风险评估声明在 2028-04-01 前提交；这些日期仅作线索。 | [15] |
Leads only (search summaries, not read at source): California AB 2013 (training-data transparency, effective 2026-01-01; xAI has sued, with the challenge reported pending at the Ninth Circuit), SB 53 (frontier-model safety transparency, 2026-01-01) and SB 942 (AI content provenance, reported effective 2026-08-02); Texas HB 149, the Responsible AI Governance Act (effective 2026-01-01; Attorney General enforcement, a 36-month sandbox, and reported protection for firms that substantially comply with a recognised framework such as NIST AI RMF); Illinois, New York and Utah measures, which I did not research.
仅为线索（搜索摘要，未读原文）：加州 AB 2013（训练数据透明度，2026-01-01 生效；xAI 已起诉，据报道上诉正在第九巡回法院审理）、SB 53（前沿模型安全透明度，2026-01-01）和 SB 942（AI 内容溯源，据报道 2026-08-02 生效）；德州 HB 149《负责任 AI 治理法》（2026-01-01 生效；由总检察长执法，设 36 个月沙盒，据报道对实质遵守 NIST AI RMF 等公认框架的机构提供保护）；伊利诺伊、纽约、犹他的措施，我没有研究。
What this means for a model, data and AI lifecycle这对模型、数据和 AI 生命周期意味着什么
1. For a bank's classical ML models, SR 26-2 is the reference. The expectation is a documented lifecycle: purpose, data and testing, pre-use validation, monitoring, inventory, vendor oversight, tailored to materiality. It is soft law, but the duty beneath it is not. 2. For LLMs and agents at a bank there is no formal US lifecycle document yet. The agencies point to general governance and have signalled a request for information. In the meantime the working references are the bank's own governance, NIST (voluntary), and exam commentary such as FINRA's. 3. For everyone else, the binding rules are mostly state transparency and notice laws, existing sector and data statutes, and the FTC Act. The federal government is actively trying to narrow state rules, so this layer is unstable: Colorado's law was replaced before it took effect, the FTC proposal is unfinished, and the litigation is open. 4. Evidence you can build once and reuse: an inventory of models and their data dependencies, documented purpose and limitations, validation and monitoring records, vendor documentation, and change records. Almost every document above, from SR 26-2 to Colorado's developer documentation to the EU-style duties noted in the earlier control-layer note, asks for some version of these.
1. 对银行的经典 ML 模型，SR 26-2 是参照。期望是一个有文档记录的生命周期：用途、数据与测试、使用前验证、监控、清单、供应商监督，并按重要性量体裁衣。它是软法，但其下的义务并不软。2. 对银行里的大模型和智能体，美国目前还没有正式的生命周期文件。监管机构指向一般治理，并已表示将发布征求意见稿。在此之前，可用的参照是银行自己的治理、NIST（自愿）以及 FINRA 等检查评论。3. 对其他机构，有约束力的规则主要是州的透明度与通知法、现有的行业与数据法规，以及《FTC 法》。联邦政府正积极试图压缩州规则，所以这一层不稳定：科罗拉多州的法律在生效前就被替换，FTC 的提案尚未定稿，诉讼仍未了结。4. 可以一次建好、反复使用的证据：模型及其数据依赖的清单、有文档的用途与局限、验证与监控记录、供应商文档、变更记录。上面几乎每份文件，从 SR 26-2 到科罗拉多的开发者文档，再到前一篇控制层笔记里提到的欧盟式义务，都以某种形式要求这些。
Hypotheses假设
For the next 12 months, the US will have no binding federal lifecycle rule for generative or agentic AI in banking; what banks do will be set by their own governance plus NIST-style voluntary standards, until the agencies' request for information turns into guidance.
未来 12 个月，美国不会有针对银行生成式或智能体 AI 的、有约束力的联邦生命周期规则；银行的做法将由自身治理加上 NIST 式自愿标准决定，直到监管机构的征求意见稿转化为指引。
Open questions未决问题
Has the agencies' request for information on model risk and AI been issued, and what does it ask? What is the status of the Commerce evaluation of state laws and the FTC statement (finalised or withdrawn)? How does the 2026-09-29 "SI" terminology change show up in later federal documents? What do FINRA, the SEC and the CFPB actually require of ML and GenAI models (primary documents not read)? Which state laws other than Colorado, California and Texas apply to model-driven decisions?
监管机构关于模型风险与 AI 的征求意见稿是否已发布，它问了什么？商务部对州法的评估和 FTC 声明现状如何（已定稿还是撤回）？2026-09-29 的 “SI” 术语变化在之后的联邦文件中如何体现？FINRA、SEC、CFPB 对 ML 和生成式 AI 模型实际要求什么（原始文件未读）？除科罗拉多、加州、德州外，还有哪些州法适用于模型驱动的决策？
What would change my view?什么会改变我的看法？
- The agencies publishing AI-specific model risk guidance or a request for information with concrete lifecycle expectations (would move the generative/agentic gap).
- 监管机构发布针对 AI 的模型风险指引，或带有具体生命周期期望的征求意见稿（会改变生成式/智能体方面的空白）。
- Congress passing a federal preemption statute, or a court ruling against the Executive Order's approach (would change the state-law layer).
- 国会通过联邦优先权法律，或法院裁定不支持该行政令的做法（会改变州法这一层）。
- Reading the primary FINRA, SEC and CFPB documents and finding binding AI-specific duties I have only seen summarised.
- 读到 FINRA、SEC、CFPB 的原始文件，发现我只在摘要里见过的、有约束力的 AI 专项义务。
Link to hypotheses与假设的关系
H5 (formal guidance treats generative and agentic AI separately near term, while organisations extend internal governance): supported again by SR 26-2 footnote 3 and the planned request for information; the Executive Order's light-touch direction points the same way. H1 (evaluation evidence becomes the primary artifact): weak support only, since SR 26-2 makes validation and outcomes analysis central for in-scope models but says nothing about LLM evaluation. Q3: SR 26-2 says validation quality depends on rigour, not organisational structure, which leaves room for non-traditional validators but does not address delegating acceptance to an agent. Q7: the aggregate-risk and inventory language is the regulatory hook for shared-data visibility, without naming lineage. The thesis page is not edited for a single note.
H5（短期内正式指引把生成式和智能体 AI 分开处理，而机构扩展内部治理）：再次得到 SR 26-2 脚注 3 和计划中的征求意见稿的支持；行政令的轻监管方向也指向同一结果。H1（评估证据成为主要产出物）：仅有较弱的支持，因为 SR 26-2 对范围内的模型把验证与结果分析放在核心，但对大模型评估只字未提。Q3：SR 26-2 称验证质量取决于严谨程度而非组织架构，这为非传统验证者留了空间，但没有涉及把接受权委托给智能体。Q7：聚合风险和清单的表述是“共享数据可见性”的监管抓手，但没有点名血缘。论点页不因单篇笔记而修改。
Not read (gaps)未读（缺口）
America's AI Action Plan (2025-07), OMB memoranda M-25-21 and M-25-22 on federal agency AI use (seen only in search summaries), the full text of the 2026-06-02 and 2026-09-29 orders, the Commerce evaluation (status unknown), the rest of the FTC proposed statement, OCC Bulletin 2026-13 at full length (only a page summary), SR 11-7 for a side-by-side comparison, the Texas and California statutes, FINRA/SEC/CFPB/NAIC/FDA primary documents, and the federal privacy and data statutes. The EU and UK are out of scope here (see the earlier convergence note).
《美国 AI 行动计划》（2025-07）、关于联邦机构使用 AI 的 OMB 备忘录 M-25-21 和 M-25-22（只见过搜索摘要）、2026-06-02 和 2026-09-29 两份行政令的全文、商务部评估（状态未知）、FTC 拟议声明的其余部分、OCC 公告 2026-13 的完整文本（只看了页面摘要）、用于逐条对比的 SR 11-7、德州与加州的法律文本、FINRA/SEC/CFPB/NAIC/FDA 的原始文件，以及联邦隐私与数据法规。欧盟和英国不在本文范围内（见前一篇趋同笔记）。
References参考文献
- Federal Reserve, FDIC, OCC, “Supervisory Guidance on Model Risk Management,” SR 26-2 attachment, 2026-04-17; PDF read in full 2026-10-04. https://www.federalreserve.gov/supervisionreg/srletters/SR2602a1.pdf supervisory guidance · S-0041
- Federal Reserve, SR 26-2 letter, 2026-04-17 (supersedes SR 11-7 and SR 21-8). https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm supervisory letter · S-0006
- OCC, Bulletin 2026-13 “Model Risk Management: Revised Guidance,” 2026-04-17; read through a summariser 2026-10-04. https://www.occ.gov/news-issuances/bulletins/2026/bulletin-2026-13.html supervisory bulletin · S-0052
- OCC, News Release 2026-29, “OCC Issues Updated Model Risk Management Guidance”; read through a summariser 2026-10-04. https://www.occ.gov/news-issuances/news-releases/2026/nr-occ-2026-29.html news release · S-0053
- Executive Order 14365, “Ensuring a National Policy Framework for Artificial Intelligence,” 2025-12-11, 90 Fed. Reg. 58499; full text read 2026-10-04. https://www.whitehouse.gov/wp-content/uploads/2025/12/eo-14365.pdf executive order · S-0048
- The White House, “National Policy Framework for Artificial Intelligence: Legislative Recommendations,” March 2026; full text read 2026-10-04. https://www.whitehouse.gov/wp-content/uploads/2026/03/03.20.26-National-Policy-Framework-for-Artificial-Intelligence-Legislative-Recommendations.pdf policy document · S-0049
- The White House, Fact Sheet, “President Donald J. Trump Inaugurates The Era of Super Intelligence,” 2026-09-29; read in a browser 2026-10-04 (order text not read). https://www.whitehouse.gov/fact-sheets/2026/09/fact-sheet-president-donald-j-trump-inaugurates-the-era-of-super-intelligence/ fact sheet · S-0050
- The White House, Fact Sheet, “President Donald J. Trump Promotes Advanced Artificial Intelligence Innovation and Security,” 2026-06; read through a summariser (low confidence). https://www.whitehouse.gov/fact-sheets/2026/06/fact-sheet-president-donald-j-trump-promotes-advanced-artificial-intelligence-innovation-and-security/ fact sheet · S-0051
- Federal Trade Commission, “Proposed Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems,” 2026-07-01; opening sections read 2026-10-04. https://www.ftc.gov/system/files/ftc_gov/pdf/ai-policy-statement_0.pdf proposed policy statement · S-0054
- U.S. Department of Justice, press release on intervening in xAI's lawsuit against Colorado's AI law, 2026-04-24; read through a summariser. https://www.justice.gov/opa/pr/justice-department-intervenes-xai-lawsuit-challenging-colorados-algorithmic-discrimination government press release · S-0055
- Colorado General Assembly, SB26-189 “Automated Decision-Making Technology”; read through a summariser 2026-10-04. https://leg.colorado.gov/bills/sb26-189 state legislature page · S-0056
- McDermott Will & Schulte, “Colorado AI law in flux: Comprehensive replacement bill signed after federal court blocks predecessor's enforcement,” 2026; read through a summariser (law-firm commentary). https://www.mcdermottlaw.com/insights/colorado-ai-law-in-flux-comprehensive-replacement-bill-signed-after-federal-court-blocks-predecessors-enforcement/ law-firm article · S-0057
- NIST, “AI Risk Management Framework” page; read through a summariser 2026-10-04. https://www.nist.gov/itl/ai-risk-management-framework standards body page · S-0058
- NIST CAISI, “AI Agent Standards Initiative” page; read through a summariser 2026-10-04. https://www.nist.gov/caisi/ai-agent-standards-initiative standards body page · S-0059
- California Privacy Protection Agency, “CCPA Updates, Cybersecurity Audits, Risk Assessments, ADMT, and Insurance Regulations” page; read through a summariser 2026-10-04. https://cppa.ca.gov/regulations/ccpa_updates.html state agency page · S-0060
Last updated 2026-10-04.