跳到正文
小南AI实验室·· 14 天前AI 评分68

AIforPharma 对比 12 个监管机构公开文件,FDA 警告信中 AI 相关 CGMP 违规仅 1 例

从 FDA 到全球监管机构视角看AI的GMP合规问题

AI 导读

AIforPharma.org 于 2026 年 11 月 7 日发布对 12 个监管机构公开文件的对比分析,结论是公开证据过于碎片化,无法计算 AI 相关 GMP 缺陷的可靠全球总数。

正文

FDA documents referring to AI, 2022-2026

图 1 2022–2026 年提及 AI 的 FDA 监管文件时间线:横线上方为指南、政策与内部工具,下方为警告信、483 表观察项与检查报告;文中的 Purolea 警告信位于时间线右端 2026 年处。(图源:Kwiecinski & Nandy,Discover Pharmaceutical Sciences,2026-10-31,开放获取,DOI 10.1007/s44395-026-00061-5)

制药 AI 现场观察 | 主体材料:AIforPharma.org(2026-11-07)

来源:AIforPharma.org,Insights 栏目,2026 年 11 月 7 日发布

原文标题:How often do pharmaceutical regulators actually identify artificial intelligence as a GMP compliance problem? From FDA enforcement to the global regulatory perspective.

原文链接:https://www.aiforpharma.org/2026/11/07/how-often-do-pharmaceutical-regulators-actually-identify-artificial-intelligence-as-a-gmp-compliance-problem-from-fda-enforcement-to-the-global-regulatory-perspective

作者:aiforpharma(站方署名)

性质:对 12 个监管机构公开文件的对比分析;其 FDA 定量数据引自 2026-10-31 发表于 Discover Pharmaceutical Sciences 的同行评审研究(Kwiecinski & Nandy,DOI 10.1007/s44395-026-00061-5)

一、引言

How often do pharmaceutical regulators actually identify artificial intelligence as a GMP compliance problem? A recent analysis of 3,011 FDA Warning Letters found only one direct CGMP violation involving AI. But evidence from the UK, Switzerland, Japan, Europe and other regulatory authorities reveals a more complex picture: inspectors are already encountering AI-related failures, while regulatory expectations are developing faster than publicly reported enforcement cases.

监管者究竟多常将人工智能认定为 GMP 合规问题?一项对 3,011 封 FDA 警告信的最新分析发现,与 AI 直接相关的 CGMP 违规只有一例。但来自英国、瑞士、日本、欧洲及其他监管机构的证据揭示出更复杂的图景:检查员已经在实际遭遇与 AI 相关的失败,而监管期望的成型速度快于公开报告的执法案例。

On 31 October 2026, a peer-reviewed study published in Discover Pharmaceutical Sciences examined references to artificial intelligence in FDA regulatory correspondence. The researchers reviewed 3,011 Warning Letters issued between 1 January 2022 and 28 September 2026. Only seven letters mentioned AI, and just one directly identified a company’s use of AI as a CGMP violation.

2026 年 10 月 31 日,一篇发表于《Discover Pharmaceutical Sciences》的同行评审研究考察了 FDA 监管函件中对人工智能的提及。研究者审查了 2022 年 1 月 1 日至 2026 年 9 月 28 日之间发出的 3,011 封警告信。仅 7 封提及 AI,其中只有 1 封直接将企业使用 AI 认定为 CGMP 违规。

AIforPharma.org reviewed publicly available regulatory publications, inspection-related communications, enforcement examples and GMP policy documents from 12 regulatory authorities and networks. The results show that regulators are addressing AI in substantially different ways. Some have already documented actual compliance problems. Others are developing explicit expectations for AI in manufacturing, requiring notification of significant changes, or introducing AI into their own regulatory oversight systems. However, published information does not yet support a reliable worldwide estimate of AI-related GMP deficiencies.

AIforPharma.org 审阅了来自 12 个监管机构与网络的公开监管出版物、检查相关沟通、执法案例与 GMP 政策文件。结果显示,各监管机构处理 AI 的方式差异很大。有的已经记录了真实的合规问题;有的正在为生产中的 AI 制定明确期望、要求对重大变更履行通知义务,或将 AI 引入自身的监管体系。但已发布的信息尚不足以支撑对全球 AI 相关 GMP 缺陷数量的可靠估计。

二、美国:3,011 封 FDA 警告信中仅一例直接 CGMP 案例

The retrospective study by Kwiecinski and Nandy provides an unusually useful quantitative baseline. Of 3,011 FDA Warning Letters examined, seven referred to AI, representing approximately 0.23% of the dataset.

Kwiecinski 与 Nandy 的回顾性研究提供了一个少有的有用定量基线。在被审查的 3,011 封 FDA 警告信中,7 封提及 AI,约占数据集的 0.23%。

Importantly, the references to AI did not all represent violations involving the inappropriate use of AI. In the medical-device cases, AI-related product claims were relevant to regulatory classification or authorisation. In the tobacco cases, AI-related features were discussed in connection with unauthorised products. Only one drug Warning Letter directly addressed AI used in CGMP activities.

重要的是,这些对 AI 的提及并不全都涉及不当使用 AI 的违规。在医疗器械案例中,AI 相关的产品声明与监管分类或授权有关;在烟草案例中,AI 相关功能是围绕未经授权的产品被讨论的。只有一封药物警告信直接处理了在 CGMP 活动中使用的 AI。

Purolea Cosmetics Lab 案例(2026 年 4 月 2 日)

FDA identified the company’s use of AI agents to generate drug product specifications, procedures and master production or control records. The agency found that the AI-generated documents had not been subjected to the required Quality Unit review and approval. FDA explicitly cited 21 CFR 211.22(c). The violation was not the use of AI itself, but the failure to ensure that AI-generated GMP documents were appropriately reviewed and approved before use.

FDA 认定该公司使用 AI 智能体生成药品规格、程序与主生产(控制)记录,并发现这些 AI 生成的文件未经质量部门按规定审核批准。FDA 明确引用了 21 CFR 211.22(c)。违规的不是使用 AI 本身,而是未能确保 AI 生成的 GMP 文件在使用前经过适当的审核与批准。

The study also identified three publicly released inspection records mentioning AI, including Form FDA 483 observations and Establishment Inspection Reports. None of these additional inspection records cited a firm’s use of AI as a deficiency.

该研究还识别出 3 份提及 AI 的公开检查记录(含 FDA 483 表观察项与设施检查报告),其中没有任何一份将企业使用 AI 列为缺陷。

FDA 数据说明了什么

The evidence demonstrates that existing CGMP requirements can be applied directly to activities supported by AI. It does not demonstrate that AI-related deficiencies are rare across all pharmaceutical manufacturing operations. Warning Letters represent selected enforcement actions, and the study’s small number of relevant cases does not establish a reliable trend.

这些证据证明既有 CGMP 要求可以直接适用于由 AI 支持的活动;但并不能证明 AI 相关缺陷在全部药品生产运营中都很罕见。警告信只是经过筛选的执法行动,研究中相关案例数量少,不构成可靠趋势。

三、英国:MHRA 报告了真实的 AI 生成检查回复失败案例

The UK’s Medicines and Healthcare products Regulatory Agency (MHRA) has published one of the clearest accounts of real problems associated with AI in regulatory compliance communications. On 29 June 2026, MHRA inspector Peter Brown described cases involving AI-assisted responses submitted to the agency following GxP inspections.

英国药品与保健品监管局(MHRA)就监管合规沟通中与 AI 相关的真实问题发布了最清晰的记述之一。2026 年 6 月 29 日,MHRA 检查员 Peter Brown 描述了企业在 GxP 检查后向该局提交 AI 辅助回复的多个案例。问题包括:

References to MHRA guidance that did not exist. Incorrect or inappropriate regulatory citations. Responses containing significant factual inaccuracies. Extensive documentation that failed to address the actual inspection deficiencies. Insufficient verification and technical oversight. One response exceeded 90 pages without adequately addressing the deficiencies. In another case, an AI-generated response to a serious deficiency affecting patient safety contained significant inaccuracies and non-existent references. MHRA reported that the time required to review and respond increased from approximately four hours to more than 20 hours.

引用并不存在的 MHRA 指南;错误或不当的法规引用;回复中含重大事实错误;大量文件却未回应真正的检查缺陷;核查与技术监督不足。一份回复超过 90 页仍未充分处理缺陷;另一案例中,针对影响患者安全的严重缺陷的 AI 生成回复包含重大不准确与不存在的参考文献。MHRA 报告称,审核与回复所需时间从约 4 小时增至 20 小时以上。

These are particularly important examples because they document actual failures encountered by the regulator, rather than hypothetical AI risks.

这些案例尤其重要,因为它们记录的是监管者实际遭遇的失败,而不是假设性的 AI 风险。

MHRA 的监管立场

MHRA does not prohibit AI-assisted preparation of inspection responses. Instead, it expects submissions to be accurate, verifiable, technically reviewed, appropriately approved and supported by evidence. The agency also offers voluntary disclosure of AI use in inspection responses.

MHRA 不禁止使用 AI 辅助准备检查回复,但要求提交材料准确、可验证、经技术审查、获得适当批准并有证据支持。该局还提供 AI 使用的自愿披露机制:企业可以标明哪些章节由 AI 协助撰写,并确认经过人工核查与批准。

MHRA warns that inadequate responses may be rejected, returned for correction, or contribute to higher future inspection priority and possible regulatory escalation.

MHRA 警告:不充分的回复可能被拒收、退回更正,或导致未来检查优先级提高乃至监管升级。

四、瑞士:引入 AI 可能触发正式的通知义务

Rather than reporting an individual AI-related GMP violation, the Swiss medicines inspectorates have explicitly addressed the introduction of AI into pharmaceutical operations. A technical interpretation document, effective from 21 May 2025, describes major changes requiring notification to Swissmedic under Article 41 of the Medicinal Products Licensing Ordinance.

瑞士药品检查机构没有报告单起 AI 相关 GMP 违规,而是明确处理了在药品运营中引入 AI 的问题。一份自 2025 年 5 月 21 日生效的技术解释文件(I-SMI.TI.22e,3.0 版)描述了依《药品许可条例》第 41 条须向 Swissmedic 通知的重大变更,其示例包括在生产或供应活动中引入 AI 或机器学习应用替代计算机化软件。

Notably, the document explicitly mentions: LLMs supporting complaint management. AI applications for Product Quality Reviews. AI-supported batch-record review. Machine learning applied to manufacturing processes. Automated visual inspection. New AI applications in quality control.

值得注意的是,文件明确列举了:支持投诉管理的 LLM、用于产品质量回顾(PQR)的 AI 应用、AI 辅助批记录审核、应用于生产工艺的机器学习、自动目视检查、质量控制中的新 AI 应用。

This is particularly relevant to GMP because the scope is not restricted to AI controlling manufacturing equipment. It can also encompass AI applications supporting quality-system activities.

这对 GMP 尤其重要,因为其范围不限于控制生产设备的 AI,还涵盖支持质量体系活动的 AI 应用。

However, it establishes that significant AI-related changes may have external regulatory implications beyond internal validation and change control. In January 2026, Swissmedic also introduced guidance for Scientific GMDP Meetings, explicitly identifying AI-enabled computer-controlled systems as potential subjects for early regulatory discussion.

该文件确立了:重大 AI 相关变更可能具有内部验证与变更控制之外的外部监管影响。2026 年 1 月,Swissmedic 还发布了《科学 GMDP 会议》指南,明确将 AI 赋能的计算机控制系统列为可提前监管讨论的议题。

五、欧盟:Annex 22 在大规模执法历史形成之前处理 AI 风险

In July 2025, the European Commission launched consultation on a proposed Annex 22 concerning artificial intelligence in pharmaceutical manufacturing. The consultation draft focused on static AI models and stated that generative, probabilistic, dynamic and adaptive models should not be used in critical GMP applications.

2025 年 7 月,欧盟委员会就拟议的 GMP 附录 22(AI 专题)启动公众咨询。咨询草案聚焦静态 AI 模型,并规定生成式、概率性、动态与自适应模型不应用于关键 GMP 应用。

Following stakeholder feedback, EMA held an expert workshop in June and July 2026. The official report, published on 2 October 2026, documents discussion of a potentially broader, risk-based approach that could accommodate such technologies under a robust control strategy.

在利益相关方反馈之后,EMA 于 2026 年 6-7 月召开专家研讨会。2026 年 10 月 2 日发布的官方报告记录了关于更广泛、基于风险的方法的讨论——在稳健控制策略下或可容纳此类技术。讨论议题包括:概率与生成式 AI 的验证、错误 AI 输出的检测与遏制、人类监督与问责、模型监控与漂移、数据完整性与可审计性、供应商监督与网络安全。

Regulatory status: Annex 22 remains a proposal under development. The consultation wording is not an enforceable GMP prohibition, and the workshop report does not establish a final regulatory position.

监管状态:Annex 22 仍是制定中的提案。咨询稿措辞并非可执行的 GMP 禁令,研讨会报告也未确立最终监管立场。

EU and EEA inspectors also publish statements of GMP non-compliance through EudraGMDP. However, the public database does not provide a dedicated AI-related deficiency classification. The available sources reviewed did not establish a confirmed AI-specific GMP non-compliance statement.

欧盟与欧洲经济区检查员通过 EudraGMDP 发布 GMP 不合规声明,但该公开数据库没有 AI 专属的缺陷分类。本次审阅的公开来源中未确认任何 AI 特定的 GMP 不合规声明。

六、日本:明确处理 GMP 下运行的 AI 模型

PMDA 原则(2026 年 11 月 2 日)

The Pharmaceuticals and Medical Devices Agency published its Basic Principles for the Use of AI in Regulatory Activities. The document covers AI-supported regulatory judgments throughout the pharmaceutical lifecycle, including manufacturing. Its central themes include: Human accountability. Risk-based evaluation. Clearly defined context of use. Data governance and traceability. Model evaluation. Lifecycle management and monitoring. The document explicitly states that it does not establish new regulatory requirements.

日本医药品医疗器械综合机构(PMDA)发布《监管活动中使用 AI 的基本原则》,覆盖包括生产在内的药品全生命周期中由 AI 支持的监管判断。核心主题包括:人类问责、基于风险的评估、明确的使用情境、数据治理与可追溯性、模型评估、生命周期管理与监控。文件明确声明不构成新的监管要求。

MHLW CMC 模型考量(2026 年 11 月 5 日)

A more directly GMP-relevant document was circulated by Japan’s Ministry of Health, Labour and Welfare (MHLW). It addresses the use of models, including AI and machine learning, in pharmaceutical manufacturing and quality control. The document provides a framework for evaluating model credibility according to its intended use, including verification, validation and uncertainty assessment.

一份与 GMP 更直接相关的文件由日本厚生劳动省(MHLW)流转。该文件处理包括 AI 与机器学习在内的模型在药品生产与质量控制中的使用,提供了按预期用途评估模型可信度的框架,包括验证、确认与不确定性评估。

One particularly interesting distinction concerns models operating under GMP without directly supporting registered manufacturing or control-strategy information. Examples include deviation investigation support and internal quality-management tools. The document indicates that such models may not generally need to be included in the Common Technical Document, but the relevant information should be available for presentation to regulators on request.

一个特别值得注意的区分涉及在 GMP 下运行、但不直接支持注册生产或控制策略信息的模型,例如偏差调查支持与内部质量管理工具。文件指出,此类模型一般无需纳入 CTD(通用技术文档),但相关信息应随时可供监管机构调阅。

It also addresses monitoring of model performance deterioration, updating or retraining models, fallback arrangements and integration with GMP change and deviation management. The framework is limited to models whose parameters remain fixed during operation. Automatically updating models and models whose behaviour cannot be sufficiently specified or verified are outside its scope. This exclusion is not equivalent to a general prohibition of such technology.

文件还涵盖模型性能退化的监控、模型的更新或再训练、回退安排,以及与 GMP 变更和偏差管理的整合。该框架限于运行期间参数固定的模型;自动更新的模型、行为无法充分规定或验证的模型不在其范围内——但这种排除并不等同于对此类技术的一般性禁止。

Japan is developing an increasingly explicit approach to the evidence expected for AI models in pharmaceutical manufacturing and quality-control activities. However, these publications are regulatory principles and technical considerations, not published inspection observations concerning individual manufacturers.

日本正在就药品生产与质控活动中 AI 模型所需的证据形成愈发明确的路径。但这些出版物属于监管原则与技术考量,并非针对个别企业的检查观察记录。PMDA 年度 GMP 检查报告及主要缺陷匿名清单中,亦未确认 AI 特定案例。

七、加拿大:AI 支持合规工作,但公开检查细节有限

Health Canada has described AI initiatives relevant to regulatory inspection and compliance activities. Its 2026-27 Departmental Plan refers to using image recognition and document analysis to support inspectors assessing regulated products, including drugs.

加拿大卫生部描述了与监管检查和合规活动相关的 AI 计划。其 2026-27 部门计划提到使用图像识别与文档分析支持检查员评估包括药品在内的受监管产品。

On 26 October 2026, Health Canada announced that public posting of Initial Inspection Deficiencies in its Drug and Health Products Inspections Database had been discontinued effective 21 October 2026. This applies to drug GMP inspections, among other inspection programmes.

2026 年 10 月 26 日,加拿大卫生部宣布自 10 月 21 日起停止在药品与健康产品检查数据库中公开初始检查缺陷,药品 GMP 检查亦在适用之列。最终合规结果仍可公开报告。该变化限制了研究者获取初步发现,也令历史检查观察与未来数据的对比更加困难。

八、澳大利亚:TGA 支持 AI 专属 GMP 指南的制定

On 26 August 2025, TGA published a communication concerning proposed revisions to the PIC/S GMP Guide. The proposed changes include a new Annex 22 dedicated to artificial intelligence.

2025 年 8 月 26 日,TGA 就 PIC/S GMP 指南的拟议修订发布沟通文件,拟议变更包括专用于人工智能的新附录 22。TGA 强调随着 AI 等数字技术进入药品生产,更新 GMP 指南十分重要。但处于征求意见阶段的 Annex 22 并不自动成为有效的澳大利亚 GMP 要求。审阅的 TGA 公开出版物中未发现因不当 AI 使用导致的 GMP 缺陷。

九、中国:NMPA 规划 AI 支持的检查与生产风险监控

Its Implementation Opinions on Artificial Intelligence + Drug Regulation, issued in 2026 and made available in English in August, describe plans to expand AI capabilities across regulatory activities.

中国 NMPA 发布的《人工智能+药品监管实施意见》(2026 年发布,2026 年 8 月提供英文版)描述了在监管活动中扩展 AI 能力的计划,包括:AI 支持的监管风险评估、更智能的检查规划、潜在合规信号的自动识别、生产相关信息的分析、对高风险产品的 AI 支持监控,以及监管数据与智能体的更大规模使用。文件还涉及模型可靠性、验证、透明度与信息安全,其愿景延伸至 2030 年及以后。

However, this is a regulatory development and implementation strategy. It does not demonstrate how many GMP inspections currently use AI, nor does it report the number of manufacturers cited for inappropriate AI applications.

不过这是一份监管发展实施战略,并未说明当前有多少 GMP 检查使用了 AI,也未报告因不当 AI 应用被查处的企业数量。

十、巴西:AI 模型结果可参与 GMP 证书风险评估

Brazil’s Anvisa introduced an explicit reference to artificial intelligence in RDC 982/2025, dated 28 July 2025. The resolution addresses risk-management criteria relevant to issuing or renewing Good Manufacturing Practices Certificates. Article 3, paragraph 3, allows results produced by AI models to be used as risk-management criteria when applicable technical requirements are met.

巴西 Anvisa 在 2025 年 7 月 28 日的 RDC 982/2025 决议中明确提及人工智能。该决议处理与 GMP 证书颁发或更新相关的风险管理标准,其第 3 条第 3 款允许在满足技术一致性、数据保护、可追溯性与验证等技术要求时,将 AI 模型产生的结果用作风险管理标准。

十一、能否测量 AI 相关 GMP 缺陷的全球趋势?

Not reliably from the currently reviewed public sources. The available evidence is highly uneven. The FDA study provides a defined document population, an explicit search methodology and reproducible counts. MHRA provides real examples of AI-related compliance failures but no total number of affected inspections or submitted responses.

从目前审阅的公开来源无法可靠测量。现有证据高度不均衡:FDA 研究提供了明确的文件总体、清晰的检索方法与可复现的计数;MHRA 提供了 AI 相关合规失败的真实案例,但没有受影响检查或提交回复的总数;其他机构发布的检查统计、不合规声明、政策文件或一般缺陷分类均未将 AI 作为独立条目持续报告。

A Warning Letter is not the same as a routine inspection observation. An incorrect response to a GxP inspection is not necessarily a manufacturing GMP deficiency. A GMP non-compliance statement may summarise several underlying deficiencies. A draft AI guideline is not evidence that a manufacturer violated a requirement. A regulator’s internal use of AI is not evidence of industry non-compliance. A missing publicly accessible case is not proof that no case exists.

警告信不等同于常规检查观察;GxP 检查回复有误未必构成生产环节的 GMP 缺陷;GMP 不合规声明可能汇总了多项底层缺陷;AI 指南草案不是企业违规的证据;监管者内部使用 AI 不是行业违规的证据;公开渠道查不到案例也不等于案例不存在。

The central finding is not that the world has recorded only one AI-related GMP violation. It is that the public evidence is currently too fragmented to calculate a defensible global total.

核心结论不是“全球只记录了一例 AI 相关 GMP 违规”,而是公开证据目前过于碎片化,无法计算一个站得住脚的全球总数。

十二、制药企业能从这份对比中学到什么

1. Existing GMP responsibilities remain applicable — The FDA case demonstrates that AI-generated documentation remains subject to established Quality Unit responsibilities.

一、既有 GMP 责任继续适用——FDA 案例证明 AI 生成的文件仍受既有的质量部门责任约束。

2. AI-generated regulatory communications must be verified — The MHRA examples show that plausible but incorrect AI-generated information can create immediate compliance consequences.

二、AI 生成的监管沟通必须经过核验——MHRA 案例显示,看似合理但不正确的 AI 生成信息会立即带来合规后果。

3. AI implementation may require external regulatory assessment — The Swiss framework demonstrates why manufacturers should evaluate notification obligations when introducing significant AI-enabled systems or modifying established processes.

三、AI 实施可能需要外部监管评估——瑞士框架说明为何企业在引入重大 AI 系统或修改既有流程时应评估通知义务。

4. Validation should reflect the intended use and model risk — The Japanese and European publications emphasise context-specific verification, validation, uncertainty, human oversight and lifecycle controls.

四、验证应反映预期用途与模型风险——日本与欧盟的出版物强调情境化的验证、确认、不确定性、人类监督与生命周期控制。

5. Regulators are developing their own AI capabilities — Developments in China, Brazil, Canada and the United States indicate increasing interest in AI-supported regulatory analysis and inspection prioritisation.

五、监管机构正在发展自身的 AI 能力——中国、巴西、加拿大与美国的动态表明对 AI 支持的监管分析与检查优先级排序的兴趣日增。企业不应假定这些举措已提供经过验证的全面自动化 GMP 监控,但这一监管发展方向值得关注。

十三、AIforPharma 评估:监管跑在公开执法证据前面

The comparison reveals an interesting asymmetry. Pharmaceutical regulators are developing increasingly detailed expectations for AI used in manufacturing, quality control and regulatory documentation. At the same time, publicly available enforcement material contains relatively few cases explicitly identifying AI as the source of a GMP failure.

对比揭示出一个有趣的不对称:监管机构正为生产、质控与监管文档中的 AI 制定愈发详细的期望,而公开执法材料中明确将 AI 认定为 GMP 失败根源的案例相对很少。

Several explanations are possible. AI adoption in critical GMP applications may still be limited. Regulators may address AI-related deficiencies using established GMP requirements without specifically mentioning the underlying technology. Inspection reports may also remain confidential or be published only in abbreviated form.

可能的解释有几种:关键 GMP 应用中的 AI 采用可能仍然有限;监管者可能用既有 GMP 要求处理 AI 相关缺陷而不特别点名底层技术;检查报告也可能保持保密或仅摘要发布。

However, three examples already reveal different dimensions of future AI compliance oversight: FDA: Are AI-generated GMP outputs appropriately reviewed and approved? MHRA: Are AI-assisted regulatory submissions accurate, relevant and supported by genuine evidence? Swissmedic: Has the company appropriately assessed and notified significant changes involving AI-enabled systems? Japan and the EU add further questions concerning model credibility, lifecycle management and risk-based controls.

不过,三个案例已经揭示了未来 AI 合规监督的不同维度——FDA:AI 生成的 GMP 输出是否经过适当审核批准?MHRA:AI 辅助的监管提交是否准确、相关并有真实证据支持?Swissmedic:企业是否对涉及 AI 系统的重大变更进行了适当评估与通知?日本与欧盟则进一步补充了模型可信度、生命周期管理与基于风险的控制等问题。

The most important conclusion is that pharmaceutical companies should not wait for a large number of AI-related Warning Letters or inspection findings before establishing appropriate governance. The absence of extensive published enforcement evidence should not be interpreted as regulatory acceptance of uncontrolled AI use. For pharmaceutical quality professionals, the priority remains straightforward: ensure that AI applications are suitable for their intended purpose, supported by appropriate evidence, governed by effective controls and subject to accountable human oversight.

最重要的结论是:制药企业不应等到出现大量 AI 相关警告信或检查发现后再建立适当治理。公开执法证据的缺乏不应被解读为监管层接受不受控的 AI 使用。对制药质量专业人员而言,优先事项依然简单明确:确保 AI 应用适合其预期用途、有适当证据支持、受有效控制治理,并处于可问责的人类监督之下。

资料来源:

Kwiecinski G., Nandy R.:《Artificial Intelligence in FDA Regulatory Correspondence, 2022 to 2026: A Retrospective Document Analysis》,Discover Pharmaceutical Sciences,2026-10-31,开放获取,https://doi.org/10.1007/s44395-026-00061-5

FDA:Purolea Cos Lab 警告信 320-26-58(2026-04-02),fda.gov

MHRA Inspectorate:《Use of AI for GXP Inspection Responses: Setting Standards Without Stifling Innovation》(2026-06-29)

Swissmedic:技术解释 I-SMI.TI.22e 3.0 版(2025-05-21 生效);《Scientific GMDP Meetings》(2026-01 生效)

欧盟委员会:GMP Chapter 4 / Annex 11 / 新 Annex 22 咨询(2025-07 至 2025-10-07);EMA Annex 22 研讨会报告(2026-10-02)

PMDA:《Basic Principles for the Use of AI in Regulatory Activities (Early Consideration)》(2026-11-02)

MHLW:《Considerations for the Use of Models, Including AI and Machine Learning, in Pharmaceutical Manufacturing and Quality Control (CMC)》(2026-11-05,日英文双语)

Health Canada:2026-27 部门计划;Drug Establishment Licensing Bulletin 197(2026-10-26)

TGA:PIC/S GMP 指南修订咨询公告(2025-08-26)

NMPA:《人工智能+药品监管实施意见》(2026 年;英文版 2026-08-20)

Anvisa:RDC 982/2025(2025-07-28)

HSA 年度 GMP 缺陷趋势报告(2023-2025);MFDS 2026 年度工作计划(2025-12-16)

原文:https://www.aiforpharma.org/2026/11/07/how-often-do-pharmaceutical-regulators-actually-identify-artificial-intelligence-as-a-gmp-compliance-problem-from-fda-enforcement-to-the-global-regulatory-perspective

来源:小南AI实验室 · mp.weixin.qq.com