Boeing AI Head Replaces Vague Ethics Principles With Testable Audits

Presentation by Mae-Beth Magon, Boeing — Ai4

Ai4

Unenforced corporate AI values create legal liabilities unless organizations enforce single-owner accountability and gated continuous risk reviews across all production models.

AI-assisted summary of Presentation by Mae-Beth Magon, Boeing — original source. Check the original for full context.

Source published: 2026-08-17

Detailed notes

## Background This presentation by Mae-Beth Magno, Head of AI Governance Center of Excellence at Boeing, was published on YouTube by Ai4 on August 17, 2026. The session details a five-pillar governance framework and a five-stage assurance lifecycle designed to make AI accountability, security, and compliance testable and auditable rather than aspirational.

## TL;DR The presentation argues that relying on vague corporate values and sentiment for Responsible AI creates liability, as traditional principles lack shared definitions, lifecycle ownership, and consequences. To address this, organizations must implement a five-pillar governance framework encompassing testable standards, single-owner accountability, GRC integration, enhanced security, and model value governance. These standards are maintained through a five-stage continuous assurance lifecycle spanning from initial classification to retirement. Ultimately, executive leadership can assess their true operational readiness by answering three direct questions regarding ownership, scheduled re-reviews, and defensible evidence of fairness.

## Atmosphere & style The presentation carries a deeply personal, pragmatic, and authoritative tone as the speaker bridges her own physical rehabilitation journey with the engineering rigor required for AI governance. Her delivery moves from intimate vulnerability to structured, high-stakes operational advice, framing complex technical concepts with clarity and conviction. Listening to the talk feels disciplined yet inspiring, motivating organizations to move away from passive 'governance theater' toward active, evidence-based accountability.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=48_tzsXyYOk&t=4s) Personal recovery following major spine surgery taught Mae-Beth Magno that building an infrastructure requires measured, gated progress and documented evidence rather than unverified beliefs. - [02:47](https://www.youtube.com/watch?v=48_tzsXyYOk&t=167s) Organizations face a 'platitude problem' where Responsible AI values live in slide decks without testable math or enforcement mechanisms. - [02:47](https://www.youtube.com/watch?v=48_tzsXyYOk&t=167s) The sentiment-based approach fails due to a lack of shared definitions for concepts like fairness, a lack of continuous lifecycle ownership, and an absence of real consequences or enforcement. - [02:47](https://www.youtube.com/watch?v=48_tzsXyYOk&t=167s) Responsible AI must be treated as an assurance discipline with testable standards and continuous verification, similar to cybersecurity or financial controls. - [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) Pillar 1 (Testable Standards) requires defined metrics with clear thresholds for fairness and documented workflow intervention points for human oversight. - [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) Pillar 2 (Accountability) dictates that every model must have a single named human owner across its entire lifecycle to prevent diffusion of responsibility. - [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) Pillar 3 (Governance, Risk, and Compliance) involves proactively embedding external regulatory frameworks into platform infrastructure as internal controls. - [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) Pillar 4 (Security) highlights that traditional cybersecurity fails to address AI-specific attack vectors, emphasizing that internal tools behind a firewall can still expose regulated data. - [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) Pillar 5 (Model Performance and Value Governance) requires proactive evaluation of whether a model remains worth its cost and continues to serve its approved purpose over time. - [10:41](https://www.youtube.com/watch?v=48_tzsXyYOk&t=641s) The five-stage assurance lifecycle spans Discover & Classify, Assess & Decide, Build & Validate, Deploy & Monitor, and Reassess or Retire. - [13:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=828s) Anthropic's Responsible Scaling Policy and Google's proposed Frontier Risk Framework demonstrate how industry leaders translate governance principles into operational protocols and independent audits. - [13:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=828s) AI governance introduces a shift from physical data exfiltration to 'purpose exfiltration,' where data leaves its authorized purpose even within an internal perimeter. - [13:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=828s) Data governance verifies if data is correct, whereas AI governance determines whether a model's decision is fair, explainable, accountable, compliant, and defensible. - [18:37](https://www.youtube.com/watch?v=48_tzsXyYOk&t=1117s) Executives can evaluate organizational readiness by asking three key questions about named ownership, scheduled re-review dates, and defensible proof of fairness.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=48_tzsXyYOk&t=4s) A Personal Journey and the Reality of Building Infrastructure

Mae-Beth Magno introduces herself as the Head of AI Governance Center of Excellence and AI Lifecycle Assurance at Boeing. Expressing her delight at speaking to the audience, she sets the stage for a discussion on moving past platitudes to operationalize Responsible AI. She shares a deeply personal story from exactly one year prior, when she was lying in a hospital bed following major spine surgery. Having lost mobility in her legs and being unable to walk for several months, she was given a clear goal by her care team: "The goal is to walk again. Not run, walk."

Over the following 12 months, building back her mobility taught her what it truly means to construct an infrastructure from the ground up when the foundation has failed. Approaching her recovery like an engineer, she experienced physical therapy, repetition, tested controls, progressive difficulty, and daily accountability with one named accountable person showing up every day. She emphasizes that at no point was she allowed to merely say, "I believe in walking" and count that as progress. Her recovery was measured and gated, relying on documented evidence of her capabilities before she was allowed to advance to the next goal. Walking onto the stage today represents a milestone for her, and she notes that the principles required to rebuild her own physical movement are identical to those needed to build robust AI governance programs.

- **Speaker Role:** Head of AI Governance Center of Excellence and AI Lifecycle Assurance at Boeing. - **Personal Lesson:** Physical recovery required measured, gated progress and documented evidence rather than unverified beliefs.

---

[02:47](https://www.youtube.com/watch?v=48_tzsXyYOk&t=167s) The "Platitude Problem" and the Failure of Principles

Organizations face a widespread "platitude problem" where commitments to fairness, transparency, accountability, and safety are frequently declared but rarely operationalized. While these core values guide how AI should be developed, treating "Responsible AI" as merely a value statement turns it into a corporate liability. It lives inside slide decks, website paragraphs, or value statements signed off by legal, flourishing primarily because the regulatory and legal landscape has not kept pace with technological advancements—meaning "no one is checking the math behind the sentiment."

This sentiment-based approach fails in three specific structural ways: 1. **No Shared Definition:** Concepts like "fairness" lack a universal baseline. A legal team, a data scientist, and a business owner will each define fairness differently. Without arbitrating a testable standard, fairness becomes a "Rorschach test." 2. **No Lifecycle Ownership:** Principles are often evaluated once during the build phase and forgotten as models drift and regulations shift. 3. **No Consequence:** A principle without enforcement is not a control; it is merely a preference, and "preferences don't survive an audit."

To solve this, organizations must reframe Responsible AI as an assurance discipline similar to cybersecurity or financial controls. Just as engineering organizations reject "we believe this supplier is safe" without a certification program, AI initiatives require testable standards (pillars) and a continuous lifecycle checking mechanism.

---

[06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) The Five Pillars of Responsible AI Governance

A complete governance program rests on five foundational pillars designed to make AI accountability and security testable and auditable:

1. **Testable Standards, Not Sentiments:** Fairness must be a defined metric with a clear threshold that is tested pre-deployment and continuously monitored. Human oversight requires a documented workflow intervention point where a human can step in, along with auditable evidence of that intervention. 2. **Accountability:** Every model must have one single named human owner across all stages of its life cycle—not a committee, department, or AI agent. If an organization cannot identify the accountable owner in under 10 seconds, it has a "diffusion of responsibility with better branding." 3. **Governance, Risk, and Compliance (GRC):** External frameworks (such as NIST or the EU AI Act) must be translated into internal control language and embedded directly into platform infrastructure. Organizations should build controls proactively and map them backward rather than waiting for regulators. 4. **Security:** Traditional cybersecurity does not fully account for new attack vectors like model theft, data poisoning, prompt injection, or reputational harm. Organizations must remember that "internal does not mean safe"—a tool running inside a firewall can still expose regulated data based on how the model processes information. 5. **Model Performance and Value Governance:** Governance is proactive and must evaluate whether a model remains worth its cost and continues to serve its approved purpose. Model drift and cost creep can degrade value over time, meaning governance functions must ask both "is this safe?" and "is this worth it?"

---

[10:41](https://www.youtube.com/watch?v=48_tzsXyYOk&t=641s) The Five-Stage Assurance Lifecycle

To prevent governance from becoming static theater, standards must be verified across a five-stage continuous assurance lifecycle:

- **Stage 1: Discover and Classify:** Before writing code, classify the use case by risk level. Treating a resume-screening model with the same scrutiny as an internal meeting summarizer buries teams in bureaucracy or causes them to miss critical risks. - **Stage 2: Assess and Decide:** Perform formal risk assessments with real decision gates (Approve, Approve with Conditions, or Reject). A process that approves every request is a rubber stamp, not governance. - **Stage 3: Build and Validate:** Validate standards like bias, robustness, and security through independent testing rather than self-attestation. - **Stage 4: Deploy and Monitor:** Implement continuous monitoring in production against approval standards. Post-mortem analyses reveal that most AI incidents occur not because of missed approval checks, but because monitoring stopped post-deployment. - **Stage 5: Reassess or Retire:** Establish automated, scheduled re-reviews. Unmonitored legacy models silently running in production serve as a leading root cause of AI incidents.

---

[13:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=828s) Industry Examples and Core Technical Principles

Real-world mechanisms illustrate how organizations translate governance into practice: - **Anthropic's Responsible Scaling Policy (RSP):** Designates a single Responsible Scaling Officer who reports quarterly to the Board of Directors. Their evaluation protocol requires capability testing for every four-fold increase in compute, mandating a pause in training if risk thresholds are triggered. - **Google's Proposed Frontier Risk Framework (FRO):** Proposes annual procedural audits conducted by independent auditors and filed confidentially with an oversight committee, reflecting regulatory structures seen in finance (FINRA) or aviation (FAA).

Two technical distinctions separate operationalized governance from abstract philosophy: 1. **Data Exfiltration vs. Purpose Exfiltration:** AI shifts the primary risk from data physically leaving the perimeter to data leaving its *authorized purpose*. An internal tool summarizing regulated content can constitute an unauthorized disclosure even if no data crosses the firewall. 2. **Data Governance vs. AI Governance:** Data governance verifies if data is correct, whereas AI governance determines if a decision is fair, explainable, accountable, compliant, and defensible. Good data governance alone solves only a fraction of the problem.

---

[18:37](https://www.youtube.com/watch?v=48_tzsXyYOk&t=1117s) Translating AI Governance for Executive Leadership

Bridging the gap between governance teams and executive leadership is a translation challenge rather than a technical one. To evaluate an organization's actual readiness, directors and board members can ask three straightforward questions:

1. *"Can I name the accountable owner of every AI system running in production right now?"* 2. *"Does every one of those systems have a scheduled re-review date, or is approval assumed to be permanent?"* 3. *"If asked to prove, not describe, prove that a specific AI decision was fair and authorized, can you produce that defensible evidence today?"*

Answering "no" to any of these questions reveals a governance gap and a reliance on "governance theater," which inevitably fails during regulatory scrutiny. With the EU AI Act in force, US state-level regulations expanding, and sector regulators shifting toward active enforcement, building testable infrastructure now is essential for long-term compliance.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=48_tzsXyYOk&t=4s) The goal is to walk again. Not run, walk. — Mae-Beth Magno

> [00:04](https://www.youtube.com/watch?v=48_tzsXyYOk&t=4s) I believe in walking — Mae-Beth Magno

> [02:47](https://www.youtube.com/watch?v=48_tzsXyYOk&t=167s) no one is checking the math behind the sentiment. — Mae-Beth Magno

> [02:47](https://www.youtube.com/watch?v=48_tzsXyYOk&t=167s) preferences don't survive an audit. — Mae-Beth Magno

> [02:47](https://www.youtube.com/watch?v=48_tzsXyYOk&t=167s) we believe this supplier is safe — Mae-Beth Magno

> [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) diffusion of responsibility with better branding. — Mae-Beth Magno

> [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) internal does not mean safe — Mae-Beth Magno

> [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) is this safe? — Mae-Beth Magno

> [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) is this worth it? — Mae-Beth Magno

> [18:37](https://www.youtube.com/watch?v=48_tzsXyYOk&t=1117s) Can I name the accountable owner of every AI system running in production right now? — Mae-Beth Magno

> [18:37](https://www.youtube.com/watch?v=48_tzsXyYOk&t=1117s) Does every one of those systems have a scheduled re-review date, or is approval assumed to be permanent? — Mae-Beth Magno

> [18:37](https://www.youtube.com/watch?v=48_tzsXyYOk&t=1117s) If asked to prove, not describe, prove that a specific AI decision was fair and authorized, can you produce that defensible evidence today? — Mae-Beth Magno

## Mentioned Mae-Beth Magno, Boeing, Ai4, NIST, EU AI Act, Anthropic, Responsible Scaling Policy, Responsible Scaling Officer, Board of Directors, Google, Frontier Risk Framework, FINRA, FAA

## Video description Mae-Beth Magno - Head of AI Governance Center of Excellence, Boeing

Every organization claims a commitment to Responsible AI but in a regulatory vacuum, that commitment rarely gets tested. This session moves past principles into practice: a five-pillar governance framework and five-stage assurance lifecycle that make AI accountability, security, and compliance testable and auditable, not aspirational. Walk away with three concrete questions to bring back to your own organization on Monday morning.

[Source](https://www.youtube.com/watch?v=48_tzsXyYOk)

中文简报

波音AI主管用可测试审计取代模糊伦理原则除非企业在所有生产模型中严格落实单一责任人问责制与门控式持续风险审查，否则未付诸执行的AI价值观将带来法律责任。演讲指出，在负责任AI方面依赖模糊的企业价值观和情感倾向会带来法律隐患，因为传统原则缺乏统一共识的定义、生命周期责任归属以及违规后果追究。为解决此问题，组织必须实施包含可测试标准、单一责任人问责、GRC整合、增强安全以及模型价值治理在内的五大支柱治理框架。这些标准通过涵盖从初始分类到模型退役的五阶段持续保障生命周期来维持。最终，高管团队可以通过回答关于责任归属、定期复审以及具有抗辩力的公平性证据这三个直接问题，来评估自身的真实运营就绪度。波音公司 AI 治理卓越中心负责人 Mae-Beth Magno 结合自身重大脊柱手术后的康复经历指出，构建 AI 治理体系需要可衡量的阶段性进展与确凿证据，而非未经检验的空泛信念。 (0:04) 许多组织在负责任 AI 方面面临“陈词滥调”困境，因缺乏统一标准定义、生命周期归属权及实际问责机制，导致治理流于 PPT 形式而无法通过审计。 (2:47) 负责任 AI 治理框架包含五大支柱：可测试的标准、明确到个人的单一责任人、融入基础设施的 GRC、应对新型攻击的安全性，以及涵盖成本与漂移的模型价值治理。 (6:48) 治理必须贯穿五阶段保障生命周期（发现分类、评估决策、构建验证、部署监控、重新评估或淘汰），多数 AI 事故并非源于初始审批疏漏，而是由于部署后缺乏持续监控。 (10:41) Magno 建议企业高管通过三道核心问题自检合规防线：能否指出生产中每个 AI 系统的唯一责任人、是否有定期复审安排，以及能否在今天立即拿出证明特定决策公平合规的防御性证据。 (18:37)

Reading summary

Mae-Beth Magno, Head of AI Governance Center of Excellence and AI Lifecycle Assurance at Boeing, argues that organizations must move past abstract ethical platitudes by treating Responsible AI as an auditable, testable assurance discipline. The Failure of Value Statements Without Controls (2:47) Magno asserts that organizations face a widespread "platitude problem" where commitments to fairness, transparency, and safety remain confined to corporate decks and legal statements because regulations have not kept pace with technology. This sentiment-based approach breaks down due to three structural weaknesses: the absence of a shared, testable definition of concepts like fairness among legal, data science, and business stakeholders; a lack of ongoing lifecycle ownership after the initial build; and an absence of enforcement, reducing principles to mere preferences that cannot withstand an audit. Drawing from her personal experience recovering from major spine surgery—which required measurable, gated milestones and daily accountability rather than mere belief—she contends that AI initiatives similarly require verifiable certification programs rather than unverified good intentions. Five Foundational Pillars for Operational Governance (6:48) To turn governance into an operational assurance function analogous to cybersecurity or financial controls, Magno outlines five core pillars. First, organizations must establish testable standards with defined numerical metrics, pre-deployment thresholds, and documented intervention points. Second, accountability must belong to a single named human owner for each model rather than a committee or AI agent, enabling immediate identification. Third, Governance, Risk, and Compliance (GRC) functions must proactively translate external regulations like NIST or the EU AI Act into internal platform controls. Fourth, security must address AI-specific risks such as model theft, data poisoning, and prompt injection, acknowledging that tools behind a firewall can still expose regulated information. Fifth, performance and value governance must continuously monitor whether models suffer from drift or cost creep, answering whether a system remains safe and worth its expense. Executing Continuous Lifecycle Assurance (10:41) To avoid static "governance theater," standards must be verified across a continuous five-stage lifecycle: Discover and Classify, Assess and Decide, Build and Validate, Deploy and Monitor, and Reassess or Retire. Magno emphasizes that stage-one risk classification prevents low-risk applications, such as internal meeting summarizers, from receiving the same bureaucratic burden as high-risk use cases like resume screening. In stage two, risk assessments must include genuine decision gates to reject or conditionally approve models. During the build phase, independent testing must replace self-attestation. Crucially, post-mortem reviews show that most AI incidents arise because monitoring halts after deployment. Organizations must therefore maintain continuous production oversight and scheduled, automated re-evaluations to decommission unmonitored legacy models, which serve as a leading driver of incidents. Technical Distinctions and Industry Frameworks (13:48) Operationalizing governance requires recognizing technical boundaries that separate AI oversight from traditional practices. Magno distinguishes data governance—which verifies if data is correct—from AI governance, which examines whether algorithmic decisions are fair, explainable, and compliant. Furthermore, the risk landscape has shifted from physical data exfiltration to "purpose exfiltration," where an internal tool processes data outside its authorized scope despite remaining inside a perimeter. In industry practice, Anthropic addresses accountability through a Responsible Scaling Officer reporting to its board with mandatory pauses if thresholds are met at four-fold compute increases, while Google has proposed an external Frontier Risk Framework relying on independent procedural audits. Magno concludes that bridging the gap with executive leadership requires translating governance into three direct accountability questions: identifying single human owners for every production model, verifying scheduled re-review dates, and producing immediate, defensible proof of fair and authorized decisions. With expanding US state rules and the EU AI Act shifting sector regulators toward active enforcement, failure to provide documented evidence leaves organizations exposed.

精炼阅读摘要

波音公司负责AI治理卓越中心与生命周期保障的负责人梅-贝思·马格诺（Mae-Beth Magno）指出，负责任人工智能（Responsible AI）不能仅停留在空洞的价值口号上，而必须转化为具备可测试标准、单一责任人和全生命周期审计机制的工程保障学科。摆脱口号困境与责任归属明确化 (2:47) 马格诺结合自己进行脊柱手术后通过渐进式控制和严格文档记录重新学习行走的工程化康复经历指出，当前的负责任人工智能普遍陷入了“陈词滥调问题”。诸如公平、透明和安全等原则常被写在幻灯片或网页声明中，但缺乏量化标准，使得不同部门之间缺乏统一定义，进而沦为企业负债。如果一项原则缺乏强制执行手段，它就只是一种无法在审计中存活的偏好。针对这种结构性缺陷，她提出了AI治理的五大支柱：首先是建立可测试的标准，将公平性量化为部署前后的具体门槛，并要求人类干预具有可审计的证据；其次是绝对的问责制，每个模型在整个生命周期中必须拥有唯一的实名人类所有者，如果在10秒内无法确认负责人，即属于责任扩散；其余支柱还包括将外部合规要求直接转化为平台基础设施内部控制、防范模型投毒与提示词注入等新型安全威胁，以及持续评估模型的成本与性能漂移以维护业务价值。五阶段全生命周期保障流程 (10:41) 为了避免AI治理沦为静态的形式主义，组织必须实施涵盖模型全生命周期的五阶段持续保障流程。第一阶段是“发现与分类”，在编写代码之前必须根据风险等级划分用例，避免简历筛选模型与会议纪要工具混用同一审核标准导致行政冗余或遗漏风险。第二阶段是“评估与决策”，通过包含“批准、有条件批准或拒绝”的真实决策关卡进行严格风险评估，杜绝沦为盖橡皮图章。后序阶段要求建立独立验证与长期监督机制。在第三阶段“构建与验证”中，对偏见、鲁棒性和安全性的验证必须依靠独立测试而非开发者的自我证明；第四阶段“部署与监控”重点防范部署后的监控缺失，马格诺强调事后分析显示多数AI事故并非源于初始审批疏漏，而是因为投产后失去了监控；最后，第五阶段“重新评估或退役”需设定自动化的定期复审计划，以消除因遗留模型无监控静默运行而引发重大事故的安全隐患。技术防线重构与前沿治理范式 (13:48) 马格诺在技术维度明确区分了传统数据治理与AI治理的差异：传统数据治理仅验证数据的正确性与物理边界流动，而AI治理则关注决策的公平性、可解释性、可合规性与可辩护性。特别是在风险边界上，AI将主要威胁从“数据外泄”转变为“目的外泄”，即便数据未离开防火墙，若内部模型用于非授权目的或处理受规管内容，依然构成未经授权的信息披露。在行业前沿实践中，企业正在将这些原则落实为具体控制措施。例如，Anthropic在其负责任扩展政策（RSP）中任命了一名直接向董事会季度汇报的负责任扩展官，并规定算力每提升四倍就必须进行能力测试，触及风险阈值时强制暂停训练；谷歌提出的前沿风险框架（FRO）则倡导由独立审计员进行年度程序审计并向监督委员会保密备案，其运作模式借鉴了金融和航空领域的监管架构。管理层评估与外部监管合规 (18:37) 弥合治理团队与高管层之间的认知鸿沟是一项翻译挑战。马格诺建议董事会和高管通过三个关键问题评估组织是否真正具备AI合规能力：第一，能否立即指认当前在生产环境中运行的每一个AI系统的实名责任人；第二，每一个运行中的系统是否都有明确的定期复审日期，而非被默认获得永久批准；第三，如果被要求当场证明某项具体的AI决策是公平且经过授权的，组织能否在今天拿出具备法律抗辩力的证据。若对这三个问题中任意一个给出否定答案，即表明组织存在严重的治理缺陷，正处于无法应对外部监管审计的“治理表演”状态。随着欧盟《人工智能法案》生效、美国各州法规扩容以及各行业监管机构转向主动执法，企业唯有尽早建立可测试的防御性治理基础设施，才能在长期监管收紧中维持业务安全与合规。马格诺总结强调，随着全球AI法规由原则宣示走向积极执法，企业依靠未经核实的理念已无法应对合规审查；只有将治理转化为贯穿全生命周期的可审计工程控制，并建立清晰的个人问责，才能彻底解决合规赤字。

Detailed notes — zh

## Background 该内容由 YouTube 频道 Ai4 于 2026 年 8 月 17 日发布，主讲人为波音公司（Boeing）AI 治理卓越中心负责人 Mae-Beth Magno。视频围绕如何在监管真空期摆脱负责任 AI 的空洞口号展开，详细介绍了包含五大支柱与五阶段生命周期的工程化治理框架，旨在将 AI 问责制、安全与合规转化为可测试、可审计的落地方案。 ## TL;DR Mae-Beth Magno 结合自身脊柱手术康复的严密工程控制经验，直指当前企业负责任 AI 沦为“空洞口号”的三大结构性危机。她提出治理应从价值主张转型为严谨的质量保证学科，并详细拆解了包含可测试标准、单一责任人、GRC 映射、专属安全防护以及价值治理在内的五大支柱。结合覆盖发现、评估、验证、监控到退役的全生命周期五阶段，演讲强调了“内部不等于安全”和“AI 治理非数据治理”两大核心原则，并为企业高管与董事会提供了三个穿透性的检验问题，以应对日益收紧的外部监管风暴。 ## Atmosphere & style 整场演讲开篇以极具冲击力的瘫痪康复亲身经历切入，迅速营造出一种严谨、克制且高度务实的工程学质感。讲者用词犀利干脆，毫不留情地刺破企业界“合规表演”与“责任扩散”的浮夸修辞，将原本枯燥抽象的 GRC 议题转化为节奏紧凑的实战推演。观众在倾听过程中能清晰感受到航空工业级严密思维与紧迫合规现实带来的沉浸感与压迫力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=48_tzsXyYOk&t=4s) 将工程控制思维引入 AI 治理：缺乏可检验控制机制的信念只是空话，AI 体系建设需要依赖量化测量、阶段性门禁（Gated controls）与实证数据支撑。 - [02:47](https://www.youtube.com/watch?v=48_tzsXyYOk&t=167s) 负责任 AI 的“空洞口号危机”：当承诺仅停留在宣传文案与演示 PPT 中而缺乏工程化落地时，“负责任 AI”在监管真空期极易沦为企业的潜在负债。 - [02:47](https://www.youtube.com/watch?v=48_tzsXyYOk&t=167s) 传统原则性治理的三大结构性失效：缺乏跨部门可测试的统一定义、缺乏覆盖全生命周期的所有权致使模型漂移、缺乏具有强制约束力的违规后果。 - [05:37](https://www.youtube.com/watch?v=48_tzsXyYOk&t=337s) 治理体系范式重构：负责任 AI 绝非企业道德声明，而是一门与网络安全（Cybersecurity）和财务内控同等严谨的“质量保证学科（Assurance discipline）”。 - [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) 治理支柱一（可测试标准）：将公平性与人类监督等抽象原则转化为具有具体阈值的量化指标，并保留人类实际干预操作的可审计证据。 - [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) 治理支柱二（明确问责制）：生产环境中的每个模型在各阶段必须绑定唯一实名的自然人负责人，拒绝以委员会名义掩盖的责任扩散。 - [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) 治理支柱三（GRC 映射）：将 NIST、欧盟《人工智能法案》等外部监管合规要求前置翻译为企业内部控制语言，并直接嵌入治理平台底层架构。 - [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) 治理支柱四（专用安全防护）：“内部并不等同于安全”，AI 场景的核心风险在于数据脱离了授权使用目的，需针对投毒与提示词注入等建立专用防护。 - [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) 治理支柱五（性能与价值治理）：治理不仅要审查安全性，更要持续验证模型是否持续创造预期业务价值，防止模型漂移与成本蔓延（Cost creep）。 - [10:41](https://www.youtube.com/watch?v=48_tzsXyYOk&t=641s) 五阶段质量保证生命周期：包含基于风险等级的发现分类、非“橡皮图章”的决策门禁、独立第三方构建验证、部署后持续监控以及自动化定期重审或退役。 - [13:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=828s) AI 治理与数据治理截然不同：数据治理关注“数据本身是否正确”，而 AI 治理解决的是“决策系统是否正确、公平、合规，且能否在审计面前拿出证明”。 - [18:37](https://www.youtube.com/watch?v=48_tzsXyYOk&t=1117s) 识破“治理表演”的三大穿透性问题：能否立即指名模型负责人、是否有定期重审日期而非默认永久获批、能否立刻拿出证明决策合规的实质证据。 - [18:37](https://www.youtube.com/watch?v=48_tzsXyYOk&t=1117s) 合规环境加速收紧：全球 AI 监管正快速从指导建议转向实质执法，企业必须在强制监管全面到来之前主动构筑工程化治理底座。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=48_tzsXyYOk&t=4s) 从脊柱手术康复谈起：将工程控制思维引入 AI 治理波音公司（Boeing）AI 治理与 AI 生命周期保证负责人 Mae-Beth Magno 以一段深刻的个人经历拉开演讲序幕。整整一年前，她刚经历了一场重大的脊柱手术，双腿完全瘫痪并在病床上躺了数月。术后康复团队给她设定的目标非常朴素却极其严苛——“目标是重新站立行走，不是奔跑，而是走路”。在随后的十二个月里，当身体的基础架构彻底崩溃时，她切身体会到了如何从零开始重建一整套支持系统。这一康复过程完全遵循了严谨的工程师思维：通过物理治疗、高频重复、经受检验的控制措施、渐进式难度提升，以及指定唯一的责任人每天到场指导。在康复过程中，绝不允许仅仅口头宣称“我相信自己能走路”就将其算作进展；每一步都必须经过量化测量与阶段性门禁（gated controls），团队严格根据可记录的证据评估能力，只有达标前置指标后才允许进入下一阶段。正是这套严密的工程化方法，让她今天得以亲自走上演讲台。她在养伤期间同步推进波音的 AI 项目体系建设，并意识到两者的底层逻辑完全相通：缺乏可检验控制机制的信念只是空话，唯有扎实的基础设施才能支撑长远发展。 - 康复工程学核心要素：重复性训练、经过验证的控制措施（tested controls）、渐进式难度梯次、唯一指定的责任主体。 - 门禁机制：以实证数据为依据，未通过前置能力评估绝不放行进入下一阶段。 --- ### [02:47](https://www.youtube.com/watch?v=48_tzsXyYOk&t=167s) 负责任 AI 的“空洞口号危机”与三大结构性失败当前企业界普遍存在严重的“空洞口号问题”（platitude problem）。几乎所有企业都在高调宣称自身致力于“负责任 AI”（Responsible AI），罗列着公平性、透明度、问责制、安全性及以人为本等宏大价值。然而，Mae-Beth Magno 提出了一个令人不安的论点：“负责任 AI”在很多时候已经变成了一种负债。这并非因为其背后的价值观有误，而是因为它沦为企业口头宣传的修辞，而非实际构建的工程体系。它往往只存在于演示文稿、官网宣传语或经法务审核的原则声明中，并在监管尚未跟上的真空期内大行其道，因为眼下根本无人核验这些情怀背后的实际算力与算法逻辑。这种停留于表层的“负责任 AI”在结构上存在三大根本性失效： 1. **缺乏统一定义（No shared definition）**：若在会议室或企业内部询问十个人对“AI 系统公平性”的定义，法务部门、数据科学家与业务主管会给出截然不同的答案。如果无法将概念裁定为可测试的技术标准，公平性就不是原则，而是一场“罗夏墨迹测验”（Rorschach test）。 2. **缺乏全生命周期所有权（No lifecycle ownership）**：大多数原则仅在模型构建阶段被勾选核对一次，随后便被彻底遗忘，任由模型发生漂移（model drift），对外部监管环境的变动毫无感知。 3. **缺乏违规后果（No consequence）**：没有强制执行力的原则根本不是控制机制，充其量只是一种“偏好”（preference），而“偏好在合规审计面前不堪一击”。 --- ### [05:37](https://www.youtube.com/watch?v=48_tzsXyYOk&t=337s) 重新定义治理：从价值主张转向质量保证学科针对上述困境，治理体系必须进行根本性重构：负责任 AI 绝不是一份企业价值观声明，而是一门保证学科（assurance discipline），其运作逻辑应与网络安全（Cybersecurity）或财务内部控制（Financial Controls）保持高度一致。在严谨的工程组织中，绝不会有人接受“我们相信这家供应商是安全的”来替代严苛的供应商认证程序；同理，面对新兴的 AI 技术，也不能降低对其工程化验证的标准。这一转型依赖于两大核心支柱的支撑：首先是一组清晰、数量精炼且可被测试的具体标准，明确界定“合规达标”的技术基准；其次是明确这些标准何时被检查的生命周期流程——而答案绝非仅在上线时审查一次，而是“贯穿始终的持续检查”。 --- ### [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) 负责任 AI 治理的五大核心支柱为将抽象概念转化为可落地的控制架构，Mae-Beth Magno 提出了负责任 AI 治理的五大核心支柱： ```mermaid flowchart TD accTitle: 负责任 AI 治理的五大核心支柱 subgraph P1["支柱 1: 可测试标准"] A["指标与阈值量化"] B["记录人类介入证据"] end subgraph P2["支柱 2: 明确问责制"] C["单一实名人所有者"] D["10 秒内确认归属"] end subgraph P3["支柱 3: GRC 控制映射"] E["外部框架转内部语言"] F["内嵌至平台底层架构"] end subgraph P4["支柱 4: 专用安全防护"] G["防御投毒与注入攻击"] H["内部不等于安全原则"] end subgraph P5["支柱 5: 性能与价值治理"] I["持续监控模型漂移"] J["ROI 与成本蔓延审查"] end ``` - **第一支柱：可测试标准化的负责任 AI（Responsible AI as Testable Standards）**：摒弃模糊的情感表述。例如“公平性”必须体现为带有具体阈值的量化指标，在部署前完成测试并保持持续监控；“人类监督”必须是在工作流中明确记录的介入点，并能提供人类实际执行干预的审计证据。 - **第二支柱：明确问责制（Accountability）**：每个模型在全生命周期的每个阶段，必须绑定唯一指定的实名自然人负责人（Named human owner）。责任主体不能是委员会、部门或智能代理，而必须是一个具体的人。如果不能在 10 秒钟内回答出“谁在为当前生产环境中的这个模型负责”，那么企业并未建立问责制，只是拥有“换了更好包装的责任扩散”（diffusion of responsibility with better branding）。 - **第三支柱：GRC（治理、风险与合规）映射**：将外部监管框架（如 NIST 风险管理框架、欧盟《人工智能法案》EU AI Act 以及行业特定法规）深度翻译为企业内部的控制语言，并直接嵌入到 AI 治理平台的基础设施中，做到前置设计控制措施并逆向映射法规要求，而非被动等待监管机构发号施令。 - **第四支柱：AI 专属安全（Security）**：AI 引入了全新的攻击面，传统网络安全体系无法有效应对模型窃取、数据投毒（data poisoning）、提示词注入（prompt injection）及声誉损害等新型威胁。在此必须树立一条核心认知：“内部并不等同于安全（Internal does not mean safe）”。即使工具完全运行在企业防火墙内部，依然可能发生受规管或敏感数据的泄露，因为风险本质不在于工具运行在哪里，而在于模型获取数据后如何进行处理。 - **第五支柱：模型性能与价值治理（Model Performance and Value Governance）**：这是目前极少有企业真正落地的板块。治理不仅是前置拦截坏事，更是主动验证模型是否依然物有所值、是否持续产生批准时预期的业务价值。面对模型漂移和成本悄然攀升（cost creep），一个 18 个月前获批的模型可能早已失去投资回报率（ROI），但若无人负责价值治理，系统就会沦为盲区。“如果治理功能只问‘这是否安全’而从不问‘这是否值得’，那么你只完成了一半的治理工作。” --- ### [10:41](https://www.youtube.com/watch?v=48_tzsXyYOk&t=641s) 覆盖全生命周期的五阶段保证体系五大支柱定义了治理的“内容”（what），而五阶段保证生命周期则定义了治理的“时机”（when）。若缺乏全周期的动态追踪，任何标准都只是一场静态的“合规表演”。 ```mermaid flowchart LR accTitle: AI 质量保证生命周期五阶段 S1["1. 发现与分类 (按风险分级)"] --> S2["2. 评估与决策 (严格准入门禁)"] S2 --> S3["3. 构建与验证 (独立第三方测试)"] S3 --> S4["4. 部署与监控 (生产环境持续核查)"] S4 --> S5["5. 重评或退役 (自动化定期审查)"] ``` - **第一阶段：发现与分类（Discover and Classify）**：在编写模型的第一行代码之前，必须根据风险等级对业务用例进行分类。这一步对下游治理起到了决定性作用，它直接界定了后续审查的深度。若将简历筛选模型与内部会议摘要工具施加完全相同的审查流程，治理体系不仅会被官僚主义拖垮，更会错失对真正高风险事项的把控。 - **第二阶段：评估与决策（Assess and Decide）**：在系统工具内部嵌入正式的风险评估流程，形成真实的决策门禁。决策结果必须具备分流能力——批准、附条件批准或直接驳回。“如果每一项申请都能顺理成章通过审批，那不叫治理，叫橡皮图章（rubber stamp）。” - **第三阶段：构建与验证（Build and Validate）**：技术团队构建模型的同时，各项预设标准必须针对真实模型进行偏见测试、鲁棒性检验与安全攻防。此阶段必须引入独立验证，严禁开发团队进行自我背书式声明（self-attested）。 - **第四阶段：部署与监控（Deploy and Monitor）**：绝大多数企业在这一阶段投入严重不足。上线后的生产模型必须按照最初获批时的同一套标准接受持续监控。大量 AI 事故复盘显示，故障的根源往往不是审批阶段漏掉了什么，而是在部署之后陷入了无人看管的盲区。 - **第五阶段：重新评估或退役（Reassess or Retire）**：每个上线模型都必须设定自动化的定期重审机制。许多行业 AI 事故的事后复盘（post-mortem）一致表明，模型在无人监控的状态下静默运行数年、甚至审批者早已离职遗忘，是引发灾难性后果最常见的根本原因之一。 --- ### [13:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=828s) 行业落地机制与两项底层治理原则为了说明该体系已具备工程可行性，Mae-Beth Magno 列举了头部科技企业已付诸实践的具体机制（而非口头声明）： - **Anthropic 的负责任扩展策略（RSP）**：设立了法定的“负责任扩展官”（Responsible Scaling Officer）而非多头委员会，由该独立专人直接向董事会按季度汇报。其评估协议明确规定：算力规模每增长 4 倍就必须强制执行能力测试，一旦触发风险阈值，模型训练“必须暂停（pause）”而不是“考虑暂停”，项目可能直接在该门禁前被否决。 - **谷歌提议的独立审查机制（Front-tier / Foundation Model Oversight）**：主张在模型发布前由独立审计员进行年度程序化审计，并将审计报告机密归档至监督委员会。这正如金融行业的 FINRA 或航空领域的 FAA 一样，AI 领域亟需建立同等级别的外部刚性监管构造。在技术落地层面，还需厘清两项关键认知差异： 1. **“内部不等于安全”的本质机制**：传统数据安全主要基于网络边界（Perimeter），关注“数据是否流出公司”；AI 彻底打破了这一假设，AI 场景下的核心风险是“数据脱离了其授权使用目的（data leaving its authorized purpose）”。例如，一名工程师使用内部合规部署的 AI 工具摘要一份受规管文档，尽管没有任何字节流出公司防火墙，但该摘要行为本身已构成了非授权披露。若数据防泄漏（DLP）项目未能与 AI 治理体系深度打通，将形成巨大的安全漏洞。 2. **AI 治理与数据治理截然不同**：数据治理的核心问题是“数据本身是否准确正确（Is the data right?）”，而 AI 治理的核心问题是“系统做出的决策是否正确、公平、可解释、负责任且合规（Is our decision right?），以及我们能否给出证明”。拥有治理完备的高质量数据，依然可能驱动一个完全失控、未经治理的决策模型。如果试图用“我们具备完善的数据治理”来代指 AI 治理，企业仅仅解决了不到三分之一的问题却在盲目宣称胜利。 --- ### [18:37](https://www.youtube.com/watch?v=48_tzsXyYOk&t=1117s) 董事会层面的语言翻译与检验治理的三个核心问题治理团队向董事会与高管层汇报工作时，最大的沟通障碍从来不是技术深度，而是“翻译断层”（translation gap）。为了让非技术背景的董事会成员在没有任何技术简报的情况下快速摸清企业真实的治理水位，可直接抛出三个穿透性问题： 1. **“我们能否立即指名道姓，说出当前在生产环境中运行的每一个 AI 系统的具体负责人？”** 2. **“生产环境中的每一个 AI 系统是否都设定了（且最好是自动化的）定期重审日期，还是说所有系统都被默认为永久获批？”** 3. **“如果现在被要求拿出证据——是证明（prove）而非描述（describe）——证明某项具体的 AI 决策是公平且经过授权的，我们今天能否立刻拿出具有法律抗辩力的实质证据？”** 如果上述任何一个问题的答案是“否”，说明企业内部存在严重的治理缺口，正在上演“治理表演”（governance theater）。这种表演在平静时期尚可掩人耳目，但在遭遇真正的外部审查或危机降临的最坏时刻必然彻底崩溃。当前外部合规环境正以超乎预期的速度收紧：欧盟《人工智能法案》（EU AI Act）已经生效，美国各州立法的密度不断加大，各垂直行业的监管部门正迅速从发布“指导意见”转向“实质执法”。在监管尚未全面强制之前主动构筑工程化治理底座的企业，才不会在未来的合规风暴中陷入被动狼狈。负责任 AI 的理念宣导并不困难，将其彻底工程化与机制化落地，才是真正的决胜关键。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=48_tzsXyYOk&t=4s) 目标是重新站立行走，不是奔跑，而是走路 — Mae-Beth Magno > [02:47](https://www.youtube.com/watch?v=48_tzsXyYOk&t=167s) 偏好在合规审计面前不堪一击 — Mae-Beth Magno > [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) 换了更好包装的责任扩散 — Mae-Beth Magno > [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) 内部并不等同于安全 — Mae-Beth Magno > [06:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=408s) 如果治理功能只问‘这是否安全’而从不问‘这是否值得’，那么你只完成了一半的治理工作 — Mae-Beth Magno > [10:41](https://www.youtube.com/watch?v=48_tzsXyYOk&t=641s) 如果每一项申请都能顺理成章通过审批，那不叫治理，叫橡皮图章 — Mae-Beth Magno > [13:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=828s) 必须暂停 — Mae-Beth Magno > [13:48](https://www.youtube.com/watch?v=48_tzsXyYOk&t=828s) 数据脱离了其授权使用目的 — Mae-Beth Magno > [18:37](https://www.youtube.com/watch?v=48_tzsXyYOk&t=1117s) 我们能否立即指名道姓，说出当前在生产环境中运行的每一个 AI 系统的具体负责人？ — Mae-Beth Magno > [18:37](https://www.youtube.com/watch?v=48_tzsXyYOk&t=1117s) 生产环境中的每一个 AI 系统是否都设定了（且最好是自动化的）定期重审日期，还是说所有系统都被默认为永久获批？ — Mae-Beth Magno > [18:37](https://www.youtube.com/watch?v=48_tzsXyYOk&t=1117s) 如果现在被要求拿出证据——是证明而非描述——证明某项具体的 AI 决策是公平且经过授权的，我们今天能否立刻拿出具有法律抗辩力的实质证据？ — Mae-Beth Magno ## Mentioned Mae-Beth Magno, 波音公司（Boeing）, Anthropic, 谷歌（Google）, NIST 风险管理框架, 欧盟《人工智能法案》（EU AI Act）, FINRA, FAA ## Video description Mae-Beth Magno - Head of AI Governance Center of Excellence, Boeing Every organization claims a commitment to Responsible AI but in a regulatory vacuum, that commitment rarely gets tested. This session moves past principles into practice: a five-pillar governance framework and five-stage assurance lifecycle that make AI accountability, security, and compliance testable and auditable, not aspirational. Walk away with three concrete questions to bring back to your own organization on Monday morning. [Source](https://www.youtube.com/watch?v=48_tzsXyYOk)

Source quotes

[0:04] Mae-Beth Magno: 目标是重新站立行走，不是奔跑，而是走路 [2:47] Mae-Beth Magno: 偏好在合规审计面前不堪一击 [6:48] Mae-Beth Magno: 换了更好包装的责任扩散 [6:48] Mae-Beth Magno: 内部并不等同于安全 [6:48] Mae-Beth Magno: 如果治理功能只问‘这是否安全’而从不问‘这是否值得’，那么你只完成了一半的治理工作 [10:41] Mae-Beth Magno: 如果每一项申请都能顺理成章通过审批，那不叫治理，叫橡皮图章 [13:48] Mae-Beth Magno: 必须暂停 [13:48] Mae-Beth Magno: 数据脱离了其授权使用目的 [18:37] Mae-Beth Magno: 我们能否立即指名道姓，说出当前在生产环境中运行的每一个 AI 系统的具体负责人？ [18:37] Mae-Beth Magno: 生产环境中的每一个 AI 系统是否都设定了（且最好是自动化的）定期重审日期，还是说所有系统都被默认为永久获批？ [18:37] Mae-Beth Magno: 如果现在被要求拿出证据——是证明而非描述——证明某项具体的 AI 决策是公平且经过授权的，我们今天能否立刻拿出具有法律抗辩力的实质证据？

Source quotes

[0:04] Mae-Beth Magno: The goal is to walk again. Not run, walk. [0:04] Mae-Beth Magno: I believe in walking [2:47] Mae-Beth Magno: no one is checking the math behind the sentiment. [2:47] Mae-Beth Magno: preferences don't survive an audit. [2:47] Mae-Beth Magno: we believe this supplier is safe [6:48] Mae-Beth Magno: diffusion of responsibility with better branding. [6:48] Mae-Beth Magno: internal does not mean safe [6:48] Mae-Beth Magno: is this safe? [6:48] Mae-Beth Magno: is this worth it? [18:37] Mae-Beth Magno: Can I name the accountable owner of every AI system running in production right now? [18:37] Mae-Beth Magno: Does every one of those systems have a scheduled re-review date, or is approval assumed to be permanent? [18:37] Mae-Beth Magno: If asked to prove, not describe, prove that a specific AI decision was fair and authorized, can you produce that defensible evidence today?

Original transcript

[0:04] My name is May Beth Magno and I lead AI governance and AI life cycle assurance for Boeing. Uh, [0:11] a couple of things you should know. Uh, first of all, the ideas expressed here are representative [0:18] of my own and I am not known for brevity, but I will try to keep this within my allotted 20 [0:27] minutes. Um I'm really really happy to be with this group and talk a little bit about [0:35] um responsible AI and what that means and how we operationalize it and go beyond the platitude. Um [0:45] but first when I was writing my notes last night I realized that tomorrow is a very special day for [0:53] me. Uh, one year ago, I was lying in a hospital bed after having major spine surgery. I lost [1:02] complete mobility in my legs and wasn't able to walk for several months. Um, so after my surgery, [1:10] uh, my team looked at me and said, "The goal is to walk again. Not run, walk." Over the last [1:20] 12 months, I've learned what it truly means to build an infrastructure from the ground up when [1:27] the foundation has failed you. Physical therapy, repetition, tested controls, clearly thinking like [1:36] an engineer here. Um, progressive difficulty, one named accountable person showing up every [1:45] day. And not a single moment was I allowed to say, "I believe in walking and count that as progress." [1:56] It was measured, gated. They gave me documented evidence of what I could and couldn't do, and they [2:05] would not let me move on until I had passed the previous goals that had been set for me. [2:13] Today I am able to walk out on this stage which is a pretty big deal for me. [applause] [2:25] I want to talk to you about the other infrastructure that I was building from [2:30] the same at the same time while I was not only healing um but growing and developing [2:38] you know my my AI programs because the lessons from both are exactly the same. [2:47] Okay, we have a platitude problem. Many companies here online, everywhere you go, [2:59] they will tell you that they are committed to responsible AI, fairness, transparency, [3:06] accountability, safety, human- centered, you know, you've heard it, the values that guide how [3:15] an organization develops and deploys AI. I want to submit something a little less comfortable about [3:23] the phrase responsible AI. It's a liability. Not because the values behind it are wrong, [3:34] but because it's something that companies say instead of something they are building. [3:40] It shows up in a slide deck, a paragraph on a website, a value statement signed off by legal [3:49] and it has flourished specifically because the regulatory and legal landscape haven't caught [3:56] up with the technology which means for the moment no one is checking the math behind the sentiment. [4:09] Oh wait, I think let me check my slides here. Okay, so our platitude problem and [4:18] this idea of responsible AI, it fails in three specific structural ways. First, [4:26] there is no shared definition. Ask 10 people in this room, in this conference, [4:33] in your own organization to define fairness in an AI system. Legal will define it one way, [4:42] a data scientist another, your business owner, a third. If nobody has arbitrated this definition [4:54] into a testable standard, fairness isn't a principle. It's a roar shack test. Second, [5:02] no life cycle ownership. Principles get checked once at build and then forgotten as the model [5:13] drifts and the regulatory ground shifts beneath them. Third, no consequence. A principle without [5:23] enforcement isn't a control, it's a preference. And preferences don't survive an audit. [5:37] So, here's the reframe that I'd like to leave you with. Responsible AI is not a [5:43] value statement. It's an assurance discipline closer in spirit to how we have handled cyber [5:52] security or financial controls rather than just a corporate values page. No engineering organization [6:02] accepts we believe this supplier is safe as a substitute for a robust certification program. [6:12] We don't get to accept the equivalent of this for AI just because the discipline is [6:18] newer. That means two things have to exist. First, [6:25] pillars. A small number of testable standards that define what good looks like. And second, [6:33] a life cycle that defines when these standards get checked. Spoiler alert, continuously. [6:48] So I present five pillars of responsible AI governance. Pillar one of course responsible [6:59] AI operationalized as testable standards not sentiments. Fairness is a defined metric with [7:10] a threshold. Tested pre-eployment monitored continuously. Human oversight is a documented [7:20] point in your workflow where a human can intervene and evidence that they did. Pillar two, [7:30] accountability. One named human owner, every model, every stage of its life, one accountable [7:42] party. Not a committee, not a department, not an agent, a person. If you cannot answer who [7:53] is accountable for this model right now today, in under 10 seconds, you don't have accountability. [8:01] You have diffusion of responsibility with better branding. Pillar three, GRC. This is where you [8:11] translate all of those external frameworks, NIST, the EUAI act, sector specific regulations um into [8:21] internal control language. You are embedding this directly in the infrastructure of the platform you [8:30] are using to govern your AI. You're not waiting for a regulator to tell you what good looks like. [8:37] You are building the control now and mapping it backwards. Pillar four, security. AI introduces [8:49] an attack surface. Traditional cyber security wasn't built for model theft, data poisoning, [8:59] prompt injection, reputational harm. And here's a principle I'd ask you to take home, [9:07] even if you forget everything else. Internal does not mean safe. A tool running entirely [9:16] inside your firewall can still expose regulated or sensitive information because the exposure [9:24] isn't about where the tool runs. It's about what the model is doing with your data once it has [9:31] it. Pillar five and this is the one that almost no one builds model performance and value governance. [9:44] Governance is proactive. It is not only about stopping bad things. It is about verifying the [9:55] model is still worth what it cost and still doing what it was approved to do. Model drift, [10:06] cost creep. A model approved 18 months ago may no longer be delivering the value and no one [10:15] is watching because value governance isn't anyone's job. If your governance function [10:24] only asks is this safe and never asks is this worth it? You're only halfgoverning the system. [10:41] So what is the assurance life cycle look like? Those five pillars are the what and now the when. [10:52] Because a pillar only checked once is theater. Stage one, discover and classify before a single [11:02] line of a model is built. You classify the use case by risk. This one step does more governance [11:12] work than almost anything downstream because it determines how much scrutiny everything after it [11:21] gets. Treating a resume screening model and an internal meeting summarizer with the same [11:30] level of scrutiny is how governance functions drown themsel in bureaucracy or worse miss the [11:42] thing that actually matters. Stage two, assess and decide. A formal risk assessment embedded [11:52] within your tooling if possible. A real decision gate. Approve. Approve with conditions or reject. [12:04] This has to be a decision. If every request gets approved, you don't have governance. [12:14] You have a rubber stamp. Stage three, build and validate. This is where technical teams [12:23] are actually building and where your standards are being tested against the real model. Bias testing, [12:31] robustness, security, independently validated, not selfattested. Stage four, [12:41] deploy and monitor. Again, this is one stage that almost everyone underinvests in. A model [12:51] in production needs continuous monitoring against the same standards that it was approved on. I've [13:00] studied a lot of AI incidents. These didn't happen because governance missed something at approval. [13:07] They happened because no one was watching after that point. Stage five, [13:15] reassess or retire. Every model gets a scheduled re-review. It should be automated. [13:26] An unmonitored model quietly running in production years after anyone remembers even approving it is [13:36] one of the most common root causes in AI incident post-mortems across industries. [13:48] So what this looks like in practice, I want to stop in the abstract for a moment because [13:54] I know some of you are thinking this all sounds right, but like where is it actually being built? [14:04] Oh, I want to talk about three examples from organizations that are already [14:08] building this infrastructure. Not principle statements, mechanisms. [14:16] I'm going to pick on the big guys because I think they can take it a little bit. Okay. Um, [14:21] Anthropic, they have their responsible scaling policy which designates a named [14:28] responsible scaling officer, not a committee. One person with quarterly [14:35] report to their board of directors written into public policy commitment. [14:41] This is accountability and this is what it looks like when it's operational. [14:49] Anthropic also released their RSP evaluation protocol which requires [14:56] formal capability testing at every four time increase in compute. If a threshold triggers, [15:03] training must pause. Not consider pausing pause. It's written into the policy. [15:10] A project can fail at that gate. This is an assess and decide stage and that means [15:18] something. And now Google's proposed FRO which would require annual procedural audits before [15:30] any model release reviewed by independent auditors filed confidentially to an oversight [15:39] committee. Finance has Ferna. Aviation has the FAA. AI needs something of the same structure. [15:52] However, the gap between governance as sentiment and governance as infrastructure is exactly [16:00] this. One of those three examples produces evidence. The other is a poster on a wall. [16:09] I'll let you decide which one that means. So, two principles that make this all real. [16:18] I want to close the technical section with two distinctions that separate organizations [16:23] actually operationalizing governance from organizations that are still writing [16:28] about it. They're still in the philosophical phase of AI. First, as you've already heard, [16:35] internal does not equal safe. Here's the version that changes how you build controls. Most data [16:44] security programs are built around the perimeter. is the data leaving the company. AI breaks that [16:50] assumption because the risk isn't data leaving the company, it's the data leaving its authorized [16:58] purpose. An engineer using an internal AI tool to summarize a document that contains regulated [17:05] content where the summarization itself is the unauthorized disclosure even though nothing [17:14] ever left your firewall. If your DLP program and your AI governance program aren't best friends, [17:23] the blind spot is exactly there. Second, AI governance and data governance are not the same [17:33] function. This is hard for a lot of companies. Data governance answers one question. Is the data [17:45] right? AI governance answers something different. Is our decision right? Is it fair, explainable, [17:56] accountable, compliant? Can we prove it? You can have perfectly governed data feeding a completely [18:08] ungoverned decision. I've seen it happen. The two functions must of course coordinate and partner [18:17] constantly. But if you let we have good data governance stand in the place of AI governance, [18:28] you've only solved a third of the problem. Yet you're declaring victory. [18:37] So to close out, how do we message this? How do we bring it to the board? How do we bring it to [18:43] our leadership? How do we articulate the risk in a way that helps them understand where we are in [18:52] this moment? The gap between governance teams and those leaders or your board is [19:00] never a technical gap. It's a translation gap. So, let me give you that translation. [19:07] Three questions any director can ask in a single sitting without a technical briefing. One, [19:16] can I name the accountable owner of every AI system running in production right now? Two, [19:26] does every one of those systems have a scheduled re-review date? Again, please [19:32] make this automated. or is approval assumed to be permanent? Question three, if asked to prove, [19:43] not describe, prove that a specific AI decision was fair and authorized, can you produce that [19:54] defensible evidence today? If the answer to any of these questions is no, you have a governance gap. [20:04] You have a governance theater problem and the distinction matters because theater [20:10] fails exactly when you need it the most under real scrutiny at the worst possible moment. [20:23] The regulatory landscape is catching up faster than most of us expect. There are [20:30] meetings happening right now. Um the EU AI act is already in force. US state level laws [20:39] are multiplying and sector regulators are moving from guidance to enforcement. The organizations [20:48] building infrastructure now while it's still a choice are the ones that won't [20:55] be scrambling later. Responsible AI that's not the hard part. operationalizing it is. Thanks.