Univ. of Michigan Builds Model-Agnostic AI to Curb Vendor Dependency

From Policy to Practice: A Blueprint for Responsible AI Leadership With University of Michigan — Ai4

Ai4

Decoupling AI orchestration from commercial foundation models enables institutions to preserve data sovereignty and avoid proprietary vendor lock-in while augmenting human critical judgment.

AI-assisted summary of From Policy to Practice: A Blueprint for Responsible AI Leadership With University of Michigan — original source. Check the original for full context.

Source published: 2026-08-15

Detailed notes

## Background Published on YouTube by Ai4 on August 15, 2026, this closing keynote features Ravi Pendse, Vice President for Information Technology and Chief Information Officer at the University of Michigan. The presentation focuses on translating policy discussions into an actionable, concrete framework for responsible AI leadership and enterprise governance.

## TL;DR Ravi Pendse outlines the University of Michigan's journey as the first educational institution to roll out enterprise generative AI to 140,000 campus members, anchoring adoption in human-centered values and data sovereignty. Highlighting clinical breakthroughs like the FastGlioma surgical tool and internal platforms like Maizey, Pendse argues that AI must augment human ingenuity rather than replace it. He introduces the concept of 'AIQ'—the critical literacy of knowing both how and when not to use AI—and supplies an executive checklist for navigating hype, preserving institutional data ownership, and building vendor-agnostic architectures.

## Atmosphere & style The address feels warm, pragmatic, and collegial, balancing cutting-edge technological ambition with deliberate humility and cultural grounding. Pendse engages the audience with a relaxed, conversational delivery, weaving sports maxims, personal family philosophy, and audience pledges alongside high-stakes clinical data. The overall vibe is reassuring and thoughtful, cutting through industry panic and commercial hype to deliver an empowering, human-first vision for institutional leadership.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=wEymOOayGPI&t=4s) Responsible AI adoption requires profound institutional humility, positioning leadership and community members as continuous students of technology rather than absolute experts. - [00:04](https://www.youtube.com/watch?v=wEymOOayGPI&t=4s) In August 2023, the University of Michigan became the first educational institution to deploy enterprise-wide generative AI across its entire campus of 140,000 students, faculty, and staff. - [02:24](https://www.youtube.com/watch?v=wEymOOayGPI&t=144s) Principled AI governance requires intentionally surfacing skepticism; U-M hosted a dedicated university summit titled 'Everything That's Wrong About AI' to address anxieties over workforce displacement and eroding critical thinking. - [05:05](https://www.youtube.com/watch?v=wEymOOayGPI&t=305s) Broad organizational transformation succeeds through targeted, small-scale experiments, such as enabling instructors to launch custom course-specific AI tutors in four minutes. - [05:05](https://www.youtube.com/watch?v=wEymOOayGPI&t=305s) The ultimate return on investment for AI is the preservation of human life, exemplified by FastGlioma cutting residual malignant brain tumor margins from 25% down to 4% within a 10-second intraoperative analysis. - [05:05](https://www.youtube.com/watch?v=wEymOOayGPI&t=305s) AI functions as an amplifier rather than a substitute for human intellect, expressed in the formula AI + Human > AI because human uniqueness and domain intuition remain irreplaceable. - [09:01](https://www.youtube.com/watch?v=wEymOOayGPI&t=541s) Cultivating a high 'AIQ' (Artificial Intelligence Quotient) requires understanding algorithmic bias and environmental costs, along with practicing the intellectual discipline of knowing when not to use AI. - [11:24](https://www.youtube.com/watch?v=wEymOOayGPI&t=684s) Data sovereignty is a vital governance requirement, ensuring that sensitive healthcare records and proprietary institutional research are never leaked to external vendors or used for public training. - [11:24](https://www.youtube.com/watch?v=wEymOOayGPI&t=684s) Decoupling orchestration from underlying foundation models via vendor-agnostic architecture (such as U-M's Maizey) shields institutions from model obsolescence and market bubbles. - [11:24](https://www.youtube.com/watch?v=wEymOOayGPI&t=684s) Empowering citizen developers through codeless platforms enables tens of thousands of staff members, from procurement teams to nurses, to build tailored agents for their specific workflows. - [15:38](https://www.youtube.com/watch?v=wEymOOayGPI&t=938s) Effective leadership requires filtering substantive signal from sensational media noise and apocalyptic doom-mongering, which alienate the public and spark community backlash. - [18:27](https://www.youtube.com/watch?v=wEymOOayGPI&t=1107s) Everyday operational AI delivers tangible value across diverse domains, including digital instructor twins for 24/7 student guidance, automated administrative inquiry triage, and interactive virtual patients for clinical training. - [20:41](https://www.youtube.com/watch?v=wEymOOayGPI&t=1241s) Because artificial intelligence is built by humans for humanity, organizations must move beyond performative compliance to anchor deployment in core human-centered values and continuous observability.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=wEymOOayGPI&t=4s) Humility, Heritage, and Human-Centered AI

Ravi Pendse opens the keynote by invoking a legendary quote from University of Michigan football coach Bo Schembechler in 1969: "Those who stay will be champions." Commending the audience members who remained for the closing presentation as "AI champions," he playfully dispels any morning keynote banter about professional hierarchies or egos, remarking that Michigan approaches everything with profound humility. Despite being a professor of computer science alongside his executive duties, Pendse insists he does not consider himself an absolute expert, arguing instead that given the breathtaking pace of technological evolution, everyone in the room should consider themselves a student of AI.

```mermaid flowchart TD accTitle: FastGlioma Real-Time Intraoperative AI Workflow subgraph OR["Operating Room (Intraoperative)"] A["Patient Brain Surgery in Progress"] --> B["Continuous Intraoperative Imaging"] F["Surgeon Receives Tumor Margin Report\n(Under 10 Seconds)"] --> G["Targeted Tumor Resection\n(Residual Tumor: 25% down to 4%)"] end

subgraph Compute["Data Center & HPC"] B --> C["Secure Data Transmission"] C --> D["Machine Learning / AI Analysis\n(FastGlioma)"] D --> E["Residual Malignant Tissue Quantified"] E --> F end ```

Grounding his philosophy in family heritage, Pendse shares a foundational Sanskrit principle taught to him by his mother: *"Acharya Devo Bhava"*, meaning the teacher is made in the image of God and should be revered accordingly. While modern large language models could easily generate a literal translation, he insists on keeping human lived experience at the core of technology. Echoing Dr. Fei-Fei Li's emphasis on human-centered AI, Pendse explains that the University of Michigan became the first organization and educational institution in the country to roll out generative AI services across its entire enterprise of 140,000 students, faculty, and staff in August 2023, marking more than three years of institutional groundwork and ongoing innovation.

- **Foundational Quote:** "Those who stay will be champions." — Bo Schembechler (1969) - **Scale of Enterprise Deployment:** 140,000 campus members served simultaneously since August 2023. - **Guiding Cultural Ethos:** *"Acharya Devo Bhava"* (reverence for the educator/human guide at the center of the learning loop).

---

[02:24](https://www.youtube.com/watch?v=wEymOOayGPI&t=144s) Listening to All Voices: The "Everything That's Wrong About AI" Summit

To gauge the real sentiment surrounding responsible deployment, Pendse relates how he informally surveyed attendees across hotel gyms, elevators, and corridors. The recurring anxieties he collected centered around unconstrained AI agents, missing guardrails, imminent job displacement, generalized tech fears, and the erosion of human critical thinking skills. He emphasizes that a truly responsible AI strategy cannot operate in an echo chamber; leadership must build deliberate mechanisms to surface concerns across every tier of an organization, particularly giving voice to the silent majority caught between enthusiastic evangelists and staunch opponents.

Demonstrating this commitment to open discourse, Michigan's IT leadership sponsored a full-day summit in April titled "Everything That's Wrong About AI," specifically inviting ardent critics and skeptics to voice their objections. Pendse explains that while the goal was never to force a change of mind, creating space for mutual understanding proved so successful that a follow-up summit was scheduled for the autumn term. He challenges executive leaders to evaluate whether their institutional cultures are transparent and secure enough to host an open critique of their AI roadmaps, warning that without transparency and inclusivity, responsible AI remains empty rhetoric.

- **Community Concerns Identified:** Autonomous agent guardrails, workforce displacement, societal anxiety, and the degradation of independent critical thinking. - **Skeptics Summit:** A university-wide, day-long summit dedicated entirely to AI critiques, repeated biannually to bridge institutional divides. - **Strategic Imperative:** AI adoption requires transparent, inclusive governance where leadership actively listens rather than imposing unilateral mandates.

---

[05:05](https://www.youtube.com/watch?v=wEymOOayGPI&t=305s) Small Experiments and Saving Lives: FastGlioma in the Operating Room

Reflecting on Geoffrey Hinton’s morning remarks regarding AI tutors accelerating children's learning speeds through curiosity-driven inquiry, Pendse highlights that Michigan enabled personalized AI tutors for its faculty in August 2023—allowing instructors to configure a class-specific bot in roughly four minutes. While broad transformation can seem daunting, he advocates for launching targeted, small-scale experiments that demonstrate tangible value.

He illustrates this through the groundbreaking clinical work of Michigan neurosurgeon Dr. Todd Hollon, whose team developed FastGlioma, an AI-powered diagnostic system published in *Nature*. During delicate brain surgeries, surgeons historically faced a agonizing tradeoff: aggressive tissue removal risked crippling vital motor or cognitive functions, forcing conservative procedures that left an average of 25% of malignant tumor tissue inside patients worldwide. FastGlioma captures intraoperative brain images, transmits them to high-performance compute clusters, and returns precise margin analyses in under 10 seconds. This capability slashed residual malignant tumor rates from 25% down to just 4% while maintaining strict patient data privacy within Michigan's infrastructure.

Pendse leverages this clinical breakthrough to reframe enterprise discussions on return on investment: when financial officers question the capital expenditure of AI, the preservation of human life is the definitive ROI. He articulates the fundamental human-AI relationship through the formulation $\text{AI} + \text{You} > \text{You}$ and $\text{AI} + \text{You} > \text{AI}$. AI functions as an amplifier of human ingenuity rather than a substitute, because human uniqueness and domain intuition remain fundamentally irreplaceable.

- **Faculty AI Tutor Deployment:** Configurable in four minutes per course across the university. - **FastGlioma Surgical Impact:** Reduces residual malignant brain tumor tissue from 25% to 4% within a 10-second intraoperative processing loop. - **The Human-Centric Formula:** $\text{AI} + \text{Human} > \text{AI}$, emphasizing augmentation and irreplaceable human ingenuity over labor replacement.

---

[09:01](https://www.youtube.com/watch?v=wEymOOayGPI&t=541s) Raising AIQ and Guarding Human Critical Thinking

Pendse introduces the concept of "AIQ" (Artificial Intelligence Quotient), clarifying that it extends far beyond mechanical prompt engineering. High AIQ encompasses an awareness of the technological landscape, an understanding of algorithmic bias amplification, an accounting of environmental and energy footprints, and most crucially, the deliberate discipline of knowing *when not to use AI*. He cautions against reflexively querying large language models for everyday questions, warning that cognitive dependency diminishes analytical sharpness and models poor intellectual habits for younger generations.

He posits that higher education and enterprise organizations share a duty to ensure their students, employees, and stakeholders cultivate the highest possible AIQ. Organizations that proactively build critical literacy retain the authority to shape technology rather than being passively shaped by commercial vendors. In an interactive moment with the audience, Pendse leads the room in a unison pledge: "Responsible AI has to be human-centered."

- **Core Pillars of AIQ:** Broad technical awareness, conscious avoidance of over-reliance, environmental stewardship, bias mitigation, and preserving cognitive independence. - **Active Shaping vs. Passive Adoption:** Cultivating internal literacy prevents organizations from being held captive by proprietary vendor products. - **Audience Affirmation:** Collective consensus that human judgment must remain the absolute anchor of AI governance.

---

[11:24](https://www.youtube.com/watch?v=wEymOOayGPI&t=684s) Sovereign Data and the Maizey Ecosystem: A Vendor-Agnostic Blueprint

Organizations face a binary strategic choice: allow commercial vendors to dictate their operational destiny, or run sovereign experiments that establish internal guardrails to steer the future. Michigan implemented this philosophy by engineering *Maizey*, a proprietary, no-code AI orchestration platform built on top of modular large language models. Pendse highlights the absolute necessity of maintaining a vendor-agnostic architecture. By decoupling orchestration from underlying frontier models—whether OpenAI, open-source weights, or Llama variants—institutions protect themselves from model obsolescence and the systemic risks of potential market bubbles or circular tech valuations.

Data sovereignty represents another non-negotiable pillar of Michigan's blueprint. Sitting on vast repositories of sensitive healthcare and institutional data, the university enforces strict architectural boundaries ensuring that no proprietary or patient data is shared with external vendors or used for public model training. Looking ahead, Pendse reveals that Michigan will launch its custom "Maize & Blue App Store," empowering roughly 63,000 citizen developers across campus—from procurement specialists automating workflows to nurses crafting clinical support assistants—to build domain-specific agents without writing code.

- **Maizey Architecture:** U-M's codeless orchestration layer abstracting foundation models to ensure instantaneous, multi-vendor interchangeability. - **Data Sovereignty Mandate:** Zero external leakage of health system records or proprietary institutional research. - **Maize & Blue App Store (October Rollout):** Platform empowering 63,000 campus citizen developers to create domain-specific AI agents.

---

[15:38](https://www.youtube.com/watch?v=wEymOOayGPI&t=938s) Separating Signal from Noise: A Leadership Checklist

Pendse critiques the sensationalist headlines and apocalyptic rhetoric that dominate the public discourse—ranging from premature declarations of "the singularity" to hyperbolic historical claims regarding models too dangerous to release (such as GPT-2 in 2019) or rogue hacking agents. He argues that sweeping, uncontextualized pronouncements alienate the public, fueling community backlash against local data centers and broader AI adoption. Practicing responsible leadership requires cutting through panic and separating substantive signal from media noise.

As a practical takeaway for executive leadership, Pendse outlines a strategic checklist to guide institutional priorities:

- **Centrality of Human Judgment:** Ensure human verification and decision-making remain at the core of all operational workflows. - **Sovereign Data Governance:** Retain complete ownership and perimeter defense over institutional data assets. - **Vendor-Agnostic Orchestration:** Construct internal orchestration pipelines so that foundation model providers can be swapped without operational disruption. - **Signal Filtering:** Cultivate an institutional culture capable of evaluating breakthrough claims calmly without falling prey to hype cycles. - **Embedded Core Values:** Anchor ethical frameworks directly in established organizational values rather than treating ethics as an afterthought. - **Full Observability & Security:** Implement continuous telemetry and observability across all deployed AI systems to guarantee security.

Concluding his address as the countdown clock winds down, Pendse stresses that infrastructure designed today must be architecturally resilient for the demands of tomorrow, turning to the audience to decide whether to use his remaining minute to screen a real-world video demonstration.

[18:27](https://www.youtube.com/watch?v=wEymOOayGPI&t=1107s) Practical AI in Action Across Campus: From Digital Twins to Virtual Patients

Ravi Pendse shares a video showcase highlighting how artificial intelligence is actively woven into daily academic, clinical, and operational workflows across the University of Michigan. Faculty, staff, and students describe their personal reasons for adopting AI, centered around the imperative to engage directly with the technologies shaping the future. In teaching, one professor explains how they leverage U-M's custom AI tools to build a digital twin of themselves. Trained on their specific curriculum and course materials, this virtual persona is accessible to students 24/7, fielding questions, working through practice assignments, and driving unprecedented levels of classroom interactivity. In administrative and student support roles, AI streamlines institutional operations, from informing and expediting the university procurement process to acting as a Google Drive assistant that intercepts routine student inquiries—such as asking for office hours schedules or assignment extensions—significantly cutting down inbox clutter.

Students and campus engineering teams detail how custom tools like U-M Maizey serve as essential academic companions, compiling lecture notes into structured study guides for midterms and finals, or embedding directly into engineering resource hubs to support on-demand technical learning. In healthcare and clinical education, the impact ranges from simulated training—allowing students to hone their diagnostic skills against interactive virtual patients—to global health initiatives. Medical researchers and clinicians utilize AI models to convert dense, complex health datasets into actionable clinical insights, promoting equitable, efficient, and life-saving care in resource-constrained communities worldwide.

```mermaid flowchart TD accTitle: Practical Applications of AI Across University of Michigan subgraph Academic["Education & Learning"] A["Digital Twin Instructor"] --> B["24/7 Q&A & Practice"] C["Virtual Patients"] --> D["Clinical Simulation"] E["U-M Maizey"] --> F["Exam Prep & Note Compilation"] end subgraph Operational["Operations & Research"] G["Google Drive Bot"] --> H["Routine Query Automation"] I["Procurement AI"] --> J["Informed Purchasing"] K["Health Data Analytics"] --> L["Equitable Global Care"] end ```

- **24/7 Interactive Teaching:** Instructors deploy digital twins trained on course content to conduct interactive practice sessions and answer student questions around the clock. - **Study & Team Resource Hubs:** Students integrate U-M Maizey to digest coursework for midterm and final exam preparation and to provide real-time guidance across student engineering projects. - **Healthcare & Global Impact:** AI analyzes complex health data to deliver actionable insights and foster equitable care in low-resource environments globally, alongside clinical simulations using interactive virtual patients. - **Administrative Relief:** AI-driven Google Drive assistants triage frequent student logistical queries, such as office hours and extension requests, dramatically reducing email overhead.

---

[20:41](https://www.youtube.com/watch?v=wEymOOayGPI&t=1241s) Closing Reflections: Human-Centered AI for Humanity

Wrapping up his address, Pendse expresses his gratitude to the audience and leaves them with a foundational principle to guide their institutional efforts. He urges leaders to remember that artificial intelligence is fundamentally "created by humans for humanity," holding tremendous power to transform and save lives on a daily basis. Achieving that potential requires collective dedication, proactive leadership, and shared accountability across organizations to move beyond mere compliance and genuinely implement responsible AI.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=wEymOOayGPI&t=4s) Those who stay will be champions. — Bo Schembechler

> [00:04](https://www.youtube.com/watch?v=wEymOOayGPI&t=4s) Acharya Devo Bhava — Ravi Pendse

> [02:24](https://www.youtube.com/watch?v=wEymOOayGPI&t=144s) Everything That's Wrong About AI — Ravi Pendse

> [05:05](https://www.youtube.com/watch?v=wEymOOayGPI&t=305s) AI + You > You — Ravi Pendse

> [05:05](https://www.youtube.com/watch?v=wEymOOayGPI&t=305s) AI + You > AI — Ravi Pendse

> [09:01](https://www.youtube.com/watch?v=wEymOOayGPI&t=541s) Responsible AI has to be human-centered. — Ravi Pendse

> [11:24](https://www.youtube.com/watch?v=wEymOOayGPI&t=684s) Maize & Blue App Store — Ravi Pendse

> [15:38](https://www.youtube.com/watch?v=wEymOOayGPI&t=938s) the singularity — Ravi Pendse

> [20:41](https://www.youtube.com/watch?v=wEymOOayGPI&t=1241s) created by humans for humanity — Ravi Pendse

## Mentioned Ravi Pendse, University of Michigan, Bo Schembechler, Dr. Fei-Fei Li, Geoffrey Hinton, Dr. Todd Hollon, Nature, FastGlioma, OpenAI, Llama, Maizey, Maize & Blue App Store, GPT-2, Google Drive

## Video description Ravi Pendse - Vice President for Information Technology and Chief Information Officer, University of Michigan

The closing keynote shifts from debate to action — translating the day's policy conversations into a concrete framework for responsible AI leadership. Against a backdrop of federal deregulation and international regulatory divergence, this session asks what governance structures, accountability mechanisms, and leadership commitments truly separate principled AI adoption from performative compliance — and closes the summit with a practical blueprint.

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

中文简报

密歇根大学构建模型无关AI以遏制供应商依赖将AI编排与商业基础模型解耦，使机构能够在增强人类批判性判断的同时，维护数据主权并避免被专有供应商锁定。 Ravi Pendse概述了密歇根大学作为首个向全校14万名成员推广企业级生成式AI的教育机构的历程，强调技术落地应扎根于以人为本的价值观和数据主权。结合FastGlioma手术工具等临床突破以及Maizey等内部平台，Pendse指出AI必须增强而非取代人类的独创力。他提出了“AIQ”概念——即既懂得如何使用AI、也清楚何时不应使用AI的批判性素养——并为管理者提供了一份行动清单，旨在应对技术炒作、维护机构数据所有权以及构建独立于特定供应商的架构。密歇根大学于 2023 年 8 月成为首个在全校 14 万名师生及员工中全面部署生成式 AI 的教育机构，始终坚持以人为本与机构谦逊原则，将领导层和组织成员定位为技术的持续学习者。 (0:04) 为应对技术恐慌及对批判性思维削弱的担忧，该校举办了名为“关于 AI 的所有问题”的专题峰会，主动吸纳批评与质疑声音，构建开放透明的治理机制。 (2:24) AI 的核心价值在于赋能人类与拯救生命；密歇根大学研发的 FastGlioma 诊断系统可在 10 秒术中分析内将残留恶性脑肿瘤边缘比例从 25% 降至 4%，印证了“AI + 人 > AI”的人类能力放大理念。 (5:05) 该校坚持严格的数据主权与供应商中立架构，打造自主无代码编排平台 Maizey，并计划向 6.3 万名平民开发者开放应用商店，确保敏感医疗与机构科研数据绝不外泄。 (11:24)

Reading summary

University of Michigan CIO Ravi Pendse argues that responsible enterprise AI adoption requires human-centered governance, sovereign and vendor-agnostic infrastructure, and deliberate institutional literacy. Rather than treating AI as labor replacement or submitting to commercial vendor lock-in, organizations must design tools that augment human capabilities and actively incorporate skeptical voices. Life-Saving Augmentation and Tangible Clinical ROI (5:05) Pendse frames the foundational purpose of AI around human augmentation—formulated as AI + Human > AI—rather than labor replacement, asserting that domain expertise and human uniqueness remain irreplaceable. To demonstrate return on investment beyond narrow financial calculations, he points to FastGlioma, an intraoperative diagnostic tool developed by Michigan neurosurgeon Dr. Todd Hollon and published in Nature. During brain surgery, where surgeons traditionally left an average of 25% of malignant tumor tissue behind to avoid damaging critical motor or cognitive functions, FastGlioma delivers margin analyses within 10 seconds, slashing residual malignant tissue to 4%. This augmentation philosophy also underpins broader academic workflows. Following Michigan's August 2023 rollout across its 140,000 students, faculty, and staff, instructors deployed personalized course-level bots configurable in four minutes, digital twins to provide 24/7 student guidance, and interactive virtual patients for clinical training. Institutional Governance Through Open Skepticism (2:24) Responsible deployment cannot function within an administrative echo chamber; leadership must deliberately capture the anxieties of the broader community. Informal surveys conducted by Pendse revealed recurring campus concerns regarding unconstrained autonomous agents, missing guardrails, imminent job displacement, and the gradual erosion of human critical thinking skills. Rather than ignoring these anxieties, Michigan IT sponsored a full-day summit titled "Everything That's Wrong About AI," specifically inviting vocal critics and skeptics to articulate their objections. Pendse explains that the objective was not to force consensus but to establish mutual understanding, leading to a follow-up summit the following term. He challenges executive leaders to assess whether their workplace cultures possess sufficient transparency to openly debate and critique internal AI initiatives. Sovereign Architecture and Vendor Independence (11:24) To avoid having operational priorities dictated by commercial technology vendors, Michigan built Maizey, a proprietary, no-code AI orchestration platform. Maizey decouples campus applications from underlying foundation models, allowing the university to interchange models—including OpenAI, open-source weights, and Llama variants—without operational friction. Pendse argues this vendor-agnostic architecture shields institutions from rapid model obsolescence and market volatility. Simultaneously, the blueprint mandates strict data sovereignty. Because Michigan manages sensitive healthcare and institutional research records, the platform enforces architectural perimeters ensuring zero internal data is shared externally or absorbed into public model training. This infrastructure is set to support approximately 63,000 campus citizen developers through a custom app store, allowing faculty, nurses, and administrative staff to build domain-specific automations without writing code. Cultivating AIQ and Resisting Apocalyptic Hype (9:01) Pendse proposes the "Artificial Intelligence Quotient" (AIQ) as an organizational imperative extending beyond mechanical prompt engineering. High AIQ entails recognizing algorithmic bias, accounting for the energy and environmental footprint of computing, and exercising the discipline of knowing when not to use AI. He cautions that reflexively outsourcing daily intellectual tasks to large language models degrades analytical independence and fosters cognitive dependency. At the executive level, responsible leadership requires separating substantive signal from sensationalist media narratives, such as hyperbolic declarations of the singularity or panic over rogue hacking agents that alienate the public. Pendse outlines a leadership checklist emphasizing full system observability, continuous human verification in operational loops, and alignment with foundational institutional values to ensure technology serves human needs. While Pendse presents Michigan's sovereign orchestration and enterprise rollout as a viable blueprint, the model relies heavily on significant internal technical capacity and high-performance computing resources. How smaller institutions with limited engineering budgets can replicate this level of vendor agnosticism and clinical-grade data isolation remains an open question.

精炼阅读摘要

密歇根大学副校长兼计算机科学教授 Ravi Pendse 提出，负责任的企业级人工智能治理必须坚持“以人为本”，通过中立于供应商的技术架构和主权数据防护，使 AI 成为人类能力的放大器而非替代品。人本理念与全员包容性治理 (0:04) Pendse 指出，技术演进节奏极快，即便资深从业者也应将自己视作学习者。密歇根大学早在 2023 年 8 月即向全校 14 万名师生员工同步开放生成式 AI 服务，其核心文化秉持敬重导师与人类引领者的伦理基准。在推行过程中，校方收集并直面了来自各阶层对未受约束的自主代理、岗位替代担忧以及批判性思维退化的真实焦虑。为了打破自说自话的技术狂热，密歇根大学专门组织了名为“AI 的一切缺陷”全天峰会，邀请尖锐批评者深入对话，并将该讨论机制确立为半年度例会，要求管理层建立透明且具包容性的倾听机制，而非强制推行单边指令。敏捷实验与临床价值：以挽救生命重塑 ROI (5:05) 针对宏观变革无从下手的顾虑，演讲建议从具备切实成效的小规模实验切入。密歇根大学在 2023 年秋季支持授课教师以大约 4 分钟的速度为具体课程配置定制 AI 导师；而在临床医学领域，神经外科医生 Todd Hollon 团队研发的 FastGlioma 诊断系统则直接证明了 AI 对专业人员的增强作用。在传统脑外科手术中，受制于损伤运动或认知功能的风险，全球平均有 25% 的恶性肿瘤残余留在患者体内；FastGlioma 在严格保护患者隐私的前提下，利用术中成像与高性能计算在 10 秒内返回边界分析，将残余恶性组织比例大幅削减至 4%。Pendse 据此重塑了投资回报率（ROI）的考量，提出挽救生命即是决定性回报，并总结出“AI + 人类 > 人类”且“AI + 人类 > AI”的合作公式。培育 AIQ 与警惕思维外包 (9:01) 演讲提出了“AI 商数”（AIQ）的概念，强调其不仅局限于机械的提示词工程（Prompt Engineering），更涵盖对技术版图的认知、对算法偏见放大的警惕、对能源消耗与环境足迹的关照，以及最关键的“知晓何时不该使用 AI”。Pendse 警告称，如果对日常琐事无节制地依赖大语言模型，不仅会削弱个人的独立分析敏锐度，还会给年轻一代树立不良认知习惯。企业与高校应当通过主动培养内部全员的 AIQ，保持对技术的主导权，避免被商业软件开发商牵着鼻子走，确保人类常识与判断力始终作为决策锚点。数据主权、Maizey 架构与去厂商依赖蓝图 (11:24) 为摆脱被商业巨头绑架业务逻辑的被动局面，密歇根大学构建了名为 Maizey 的无代码编排层，将其与底层前沿模型解耦，从而在 OpenAI、开源权重或各类衍生模型之间实现无缝热切换。这种中立于供应商的设计不仅能够防范模型过时风险与科技市场估值泡沫，还建立了严密的数据主权边界，严禁将敏锐的医疗健康记录和学校专属科研资产用于公用模型训练。此外，校方计划推出包含 6.3 万名校内公民开发者的应用商店，让采购专家与护士等业务人员均可在无代码环境下构建符合特定领域需求的智能体。校园全场景落地与管理层甄别清单 (15:38) 在批判了“技术奇点即将来临”或“某些模型过度危险而无法发布”等制造恐慌的炒作后，Pendse 为机构管理者梳理了一份清单：必须将人类复核置于工作流核心、捍卫数据资产所有权、搭建去厂商绑定的中间层、建立平息舆论噪音的理性文化、将既有价值观内嵌于技术准则，并在所有部署系统中落实全链路可观测性与安全监控。随后展示的实践场景证实了其综合价值：授课教师借助基于课程资料构建的“数字分身”实现全天候答疑与互动；学生利用 Maizey 提炼复习材料；AI 驱动的驱动器助手分担了如请假和办公时间答疑等常规行政邮件负载；医学教育则通过交互式虚拟病人展开模拟问诊，并借助复杂健康数据分析推进全球资源匮乏地区的公平诊疗。 Pendse 总结指出，AI 本质上是“由人类创造并服务于人类”的技术，若要将其潜力转化为挽救生命和赋能社会的现实生产力，组织必须超越纯粹的技术合规思维，将弹性架构设计与长期责任治理落到实处。

Detailed notes — zh

## Background 本视频由 Ai4 于 2026 年 8 月 15 日在 YouTube 平台发布，记录了密歇根大学副总裁兼首席信息官拉维·潘德塞（Ravi Pendse）的主题演讲。演讲聚焦于将政策层面的讨论转化为切实的负责任 AI 领导力蓝图，探讨在去监管化与国际规则分歧的背景下，如何通过健全的治理架构与领导力承诺实现原则性的 AI 规模化落地。 ## TL;DR 密歇根大学副总裁兼 CIO 潘德塞系统分享了该校面向 14 万师生全量推行生成式 AI 的实践经验与以人为本的治理框架。他提出组织应通过倾听反对声音（如举办直面 AI 弊端的专项峰会）与培育“AIQ”来消除焦虑，并以神经外科 AI 工具 FastGlioma 将术后肿瘤残留率从 25% 降至 4% 的突破，重新定义了算力挽救生命的终极 ROI。技术层面，学校坚持“数据不出校”的数据主权铁律，打造了供应商中立的 Maizey 无代码编排平台，并拟通过应用商店赋能 6.3 万名公民开发者构建垂直应用。演讲最后给出了包含模型解耦、噪音过滤与可观测性在内的管理者自查清单，重申所有技术创新皆由人类创造且为人类福祉服务。 ## Atmosphere & style 现场演讲充满感染力与人文温度，潘德塞巧妙融合了密歇根橄榄球文化金句与古老梵文谚语，以幽默自谦的姿态迅速拉近与听众的距离。他不仅带领全场与会者起立齐诵“负责任的 AI 必须以人为中心”，更在技术架构的严密推演与脑胶质瘤手术挽救生命的动人纪实间自如切换。整场分享节奏明快且极富号召力，既有对技术狂热与泡沫噪音的清醒反思，又通过鲜活的师生实践案例传递出充满希望的务实信念。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=wEymOOayGPI&t=4s) 密歇根大学早在 2023 年 8 月即向全校 14 万名师生员工全量上线生成式 AI，坚持以李飞飞倡导的“以人为本（Human-Centered AI）”为核心哲学，强调人类在决策回路中的中心地位。 - [02:24](https://www.youtube.com/watch?v=wEymOOayGPI&t=144s) 推行负责任 AI 的首要策略是“为沉默者发声”并包容质疑；密歇根大学通过主办“AI 的所有弊端（Everything That's Wrong About AI）”反向专题峰会，直面技术伦理危机与失业焦虑。 - [05:05](https://www.youtube.com/watch?v=wEymOOayGPI&t=305s) 通过敏捷微型实验驱动高价值突破：托德·霍隆团队开发的 FastGlioma 术中实时成像 AI 系统可在 10 秒内识别肿瘤残留，将术后脑胶质瘤残留率从 25% 骤降至 4%。 - [05:05](https://www.youtube.com/watch?v=wEymOOayGPI&t=305s) 确立协同演进公式：`AI + You > You`（AI 赋能人类）与 `AI + You > AI`（人类反馈反向塑造 AI），并在保护患者隐私的前提下将“挽救生命”定义为技术的最高投资回报率（ROI）。 - [09:01](https://www.youtube.com/watch?v=wEymOOayGPI&t=541s) 提升组织与个人的“AI 商（AIQ）”超越了单纯的提示词技巧，核心在于清晰掌握前沿工具的技术边界、深刻明确何时不该使用 AI，并坚决防范独立批判性思维的退化。 - [09:01](https://www.youtube.com/watch?v=wEymOOayGPI&t=541s) 高水平的 AIQ 还涵盖对模型偏见放大效应、算法训练能耗及算力扩张对生态环境影响的清醒反思与节制意识。 - [11:54](https://www.youtube.com/watch?v=wEymOOayGPI&t=714s) 技术架构必须奉行“供应商中立（Vendor-Agnostic）”路线：密歇根大学自研 Maizey 无代码编排平台，实现底层商业 API 与开源本地模型的自由插拔替换，避免被单一厂商绑架。 - [11:54](https://www.youtube.com/watch?v=wEymOOayGPI&t=714s) 坚守“数据主权（Data Sovereignty）”铁律，严格执行数据私有化与合规隔离，确保临床医疗等海量敏感数据绝对不出校区。 - [11:54](https://www.youtube.com/watch?v=wEymOOayGPI&t=714s) 赋能全员“公民开发者”：依托即将上线的 Maizey App Store，支持全校 6.3 万名非技术背景师生员工自主构建采购合规、临床护理等垂直场景的专属 AI Agent。 - [15:38](https://www.youtube.com/watch?v=wEymOOayGPI&t=938s) 负责任 AI 领导力需要“从炒作与恐慌噪音中提炼真实信号（Separating signal from noise）”，不被过度恐慌或虚假承诺分散注意力。 - [15:38](https://www.youtube.com/watch?v=wEymOOayGPI&t=938s) 管理者需对照六项核心机制自查：人类判断核心化、数据主权治理、供应商中立架构、噪音过滤机制、核心价值观绑定以及端到端可观测性（Observability）。 - [18:27](https://www.youtube.com/watch?v=wEymOOayGPI&t=1107s) 在教学、医疗与运营中广泛落地场景创新，包括 24/7 答疑的教授数字分身、虚拟助教、临床虚拟病人演练及前瞻性智能采购决策。 - [20:41](https://www.youtube.com/watch?v=wEymOOayGPI&t=1241s) 负责任 AI 的终极准则是“由人类创造并服务于全人类（created by humans for humanity）”，技术的根本价值在于切实改善并挽救人类生命。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=wEymOOayGPI&t=4s) 密歇根精神与以人为本的 AI 起点密歇根大学副总裁兼首席信息官拉维·潘德塞（Ravi Pendse）以传奇橄榄球教练博·辛巴克勒（Bo Schembechler）在 1969 年的名言开场：“坚持到底的人终将成为冠军（Those who stay will be champions）。”他借此向所有坚持听到闭幕主题演讲的与会者致敬，称大家都是他心目中的“AI 冠军”。潘德塞风趣地调侃了当天上午主旨演讲中的一些花絮，并强调密歇根大学没有膨胀的自我意识，而是始终以谦逊的态度行事。作为一名计算机科学教授，他自称不是什么“AI 专家”，而是“AI 的终身学习者”，并认为鉴于该领域的飞速发展，每个人都应当保持学生的求知心态。潘德塞分享了他母亲教给他的一句梵文谚语：“Acharya Devo Bhava”（意为“尊师如神”，教师应当受到如同敬神般的尊重）。他幽默地表示，自己完全可以用大语言模型（LLM）来翻译这句话，但他选择亲自讲述，因为他坚信“以人为中心的 AI（Human-centered AI）”，人必须始终处于回路的核心。密歇根大学在 2023 年 8 月就成为全美第一家（不仅是高校，而是全美首个组织机构）向全校 14 万名师生员工大规模推出生成式 AI 服务的机构，这项工作至今已持续推进了三年有余。 ```mermaid flowchart LR accTitle: 以人为本的 AI 协同演进模型 subgraph Core ["核心理念"] direction TB Human["人类智慧与独特性 (You)"] AI["人工智能技术 (AI)"] end subgraph Synergies ["协同放大"] Augment["赋能增强：AI + You > You"] Refine["反向塑造：AI + You > AI"] end Human --> Augment AI --> Augment Augment --> Refine Refine --> Human ``` - **全校级落地规模**：密歇根大学于 2023 年 8 月面向全校 140,000 名用户全量上线生成式 AI 基础服务。 - **核心指导哲学**：秉承李飞飞（Dr. Fei-Fei Li）倡导的 Human-Centered AI 理念，坚持以人为本的技术发展路径。 --- ### [02:24](https://www.youtube.com/watch?v=wEymOOayGPI&t=144s) 负责任 AI 的倾听策略与“为沉默发声” 在探讨“什么是负责任的 AI”时，潘德塞没有空谈理论，而是通过在健身房、电梯和会场过道与参会者直接交流，收集大家最真实的关切：AI 智能体缺乏护栏、潜在的裁员失业危机、对技术失控的恐惧，以及 AI 对人类批判性思维能力的负面侵蚀。潘德塞指出，一个组织要推行负责任的 AI，其战略首先必须建立在“倾听组织内部所有声音”的基础之上，尤其是“为沉默者发声”。为了践行这种开放与包容，密歇根大学在今年 4 月由 CIO 办公室牵头主办了一场全天峰会，主题直截了当地定为“AI 的所有弊端（Everything That's Wrong About AI）”。校方特意邀请了所有坚决反对 AI 的师生员工前来阐述观点。潘德塞坦言，主办方并没指望通过一场会议改变谁的想法，但核心目标是增进彼此的理解。这场峰会反响极为热烈，以至于学校决定在秋季学期再次举办第二届。 - **倾听沉默群体**：在激进支持者与坚决反对者之间，重视并挖掘占据中立、沉默的多数派的声音。 - **反向主题峰会**：通过举办“Everything That's Wrong About AI”专题大会，直面技术伦理危机、失业焦虑与思维退化风险。 - **领导力行动法则**：负责任的 AI 不能停留在纸面口号，必须转化为透明、包容且经得起质疑的组织行动。 --- ### [05:05](https://www.youtube.com/watch?v=wEymOOayGPI&t=305s) 小步快跑的实践与 FastGlioma 救命案例面对纷繁复杂的 AI 技术浪潮，密歇根大学选择通过敏捷的“微型实验”快速破局。例如，早在 2023 年 8 月，学校就响应了杰弗里·辛顿（Geoffrey Hinton）关于“AI 家教能让充满好奇心的学生学习速度翻倍”的理念，为全校教师推出了可在 4 分钟内自主配置完成的课程专属 AI 助教工具。而在众多应用案例中，最具代表性的是神经外科专家托德·霍隆（Dr. Todd Hollon）团队开发的脑胶质瘤实时分析工具 FastGlioma。在传统的恶性脑肿瘤切除手术中，由于外科医生担心切除过多正常脑组织会导致患者丧失关键身体机能，操作往往偏向保守，全球临床上患者术后脑内残留恶性肿瘤组织的比例平均高达 25%。托德·霍隆博士团队开发的 FastGlioma 系统颠覆了这一现状：手术进行时，系统对手术区域进行持续成像，数据实时传输至密歇根大学的高性能计算数据中心，通过机器学习模型在 10 秒内完成恶性肿瘤残留区域的精准识别并反馈给主刀医生。 ```mermaid flowchart TD accTitle: FastGlioma 术中实时 AI 辅助流程 subgraph OR ["手术室 (Intraoperative)"] Surg["脑肿瘤切除手术进行中"] Img["高分辨率连续脑组织成像"] end subgraph HPC ["密歇根大学高性能计算中心"] ML["机器学习模型实时推理分析"] end subgraph Feedback ["临床决策与收益"] Decision["< 10 秒极速反馈肿瘤残留分布"] Outcome["恶性肿瘤残留率：25% 骤降至 4%"] end Surg --> Img Img --> ML ML --> Decision Decision --> Outcome ``` - **突破性临床成果**：该成果已在《自然》（Nature）等顶尖期刊发表，成功将术后恶性肿瘤残留率从 25% 骤降至 4%。 - **终极 ROI（投资回报率）**：面对财务官对技术投入回报的质疑，潘德塞直言，在患者数据受到绝对保护的前提下，通过算力与算法挽救生命，就是无可辩驳的最大回报。 - **公式化人机关系**：确立了 `AI + You > You`（AI 增强人类）与 `AI + You > AI`（人类反馈让 AI 更优）的协同逻辑，强调人类独创性永远不可被算法替代。 --- ### [09:01](https://www.youtube.com/watch?v=wEymOOayGPI&t=541s) 提升“AIQ”与保持人类批判性思维潘德塞提出了衡量组织与个体素养的关键指标——“AI 商（AIQ, AI Quotient）”。他澄清，AIQ 绝不仅仅是指“提示词工程（Prompt Engineering）”或会用几款热门软件，它包含更立体的认知体系：第一，了解前沿 AI 工具的技术边界；第二，深刻明确“何时不该使用 AI”。潘德塞告诫与会者，当今人们面临任何问题时往往习惯性地直接求助于大模型，这种依赖会严重退化人类的独立思考能力，尤其是对有孩子的家庭而言，家长的行为模式会直接影响下一代。高水平的 AIQ 还涵盖对 AI 训练能耗、环境影响以及社会偏见放大效应的清醒认知。他呼吁现场所有观众起立齐声朗诵：“负责任的 AI 必须以人为中心（Responsible AI has to be human-centered）”，强调组织必须在被动接受供应商摆布与主动开展实验建立护栏之间做出明确选择。 - **AIQ 核心维度**： 1. 工具认知：全面掌握主流与新兴模型工具的能力画像。 2. 节制意识：在需要深度思辨与主观判断的关键场景克制使用 AI。 3. 伦理与社会责任：警惕模型偏见放大效应，关注算力扩张对生态环境的影响。 - **塑造者而非消费者**：盲目采购外部套件无法塑造技术走向；唯有将人类判断置于核心，才能主导 AI 的发展轨迹。 --- ### [11:54](https://www.youtube.com/watch?v=wEymOOayGPI&t=714s) 架构解耦：Maizey 平台、数据主权与应用商店在技术架构层面，密歇根大学打造了名为 Maizey 的无代码 AI 编排平台。为了应对 AI 头部企业相互注资可能引发的资本泡沫风险（潘德塞将其类比为“安然式泡沫危机”），密歇根大学确立了严格的“供应商中立（Vendor-Agnostic）”战略。无论是 OpenAI 还是开源模型（如 Llama 系列），底层模型均可随时插拔替换，确保学校的未来不被任何单一商业公司绑架。 ```mermaid flowchart TD accTitle: 密歇根大学 Maizey 平台解耦架构 subgraph UserLayer ["用户与应用生态"] Store["Maizey App Store (10月上线)"] Users["6.3 万全员公民开发者 (师生/护士/采购等)"] end subgraph Orchestration ["核心中枢平台"] Maizey["Maizey 无代码编排平台 (自主知识产权)"] Sec["数据主权与隔离护栏 (数据永不出校)"] end subgraph ModelLayer ["底层模型基础设施 (随时热插拔)"] OpenAI["OpenAI 商业模型"] Llama["Llama 等开源本地模型"] Cloud["多云异构 AI 算力"] end Users --> Store Store --> Maizey Maizey --> Sec Sec --> OpenAI Sec --> Llama Sec --> Cloud ``` 更核心的支柱是数据主权（Data Sovereignty）。密歇根大学拥有庞大的医疗医疗系统，坐拥海量敏感临床数据。学校坚持“所有数据资产绝对不出校区”，筑牢数据防护墙。潘德塞宣布，密歇根大学将于 10 月正式推出专属的“Maizey App Store”，赋能全校 6.3 万名“公民开发者”（涵盖护士、采购人员、行政与科研师生），让业务人员无需深厚编程背景即可针对采购合规、护理流程等特定领域构建专属 AI Agent。 - **模型解耦**：自研 Maizey 编排层，实现对闭源商业 API 与开源模型的无缝切换。 - **数据主权铁律**：“谁掌控你的数据，谁就掌控你的未来”，严格执行数据私有化与合规隔离。 - **全员公民开发者**：通过 10 月上线的应用商店，让 63,000 名非技术背景的师生员工自主开发垂直领域智能体。 --- ### [15:38](https://www.youtube.com/watch?v=wEymOOayGPI&t=938s) 剥离噪音的行动蓝图与核心提问潘德塞回顾了科技史上层出不穷的恐慌言论与过度炒作——从 2019 年 GPT-2 发布时声称“过于危险而不敢开源”，到近年来甚嚣尘上的“技术奇点降临”与“智能体失控黑客攻击”。他指出，缺乏语境的宏大论调不仅引发了公众对数据中心和 AI 的盲目抵制，也分散了产业界的注意力。负责任的 AI 领导力核心在于“从海量噪音中提炼真实信号（Separating signal from noise）”。在演讲尾声，潘德塞给所有准备在周一返回工作岗位的管理者们留下了六项核心自查清单，呼吁各组织在为当下构建基础设施的同时，必须做好面向未来的技术就绪准备： - **人类判断核心化**：人类审校与决策是否真正置于所有 AI 业务流的核心？ - **数据主权战略**：组织是否拥有坚不可摧、不被第三方越权利用的私有数据治理框架？ - **供应商中立架构**：是否自建了可自由解耦、随时切换底层模型的编排系统？ - **噪音过滤机制**：团队是否具备辨别噱头炒作与真实技术趋势的理性文化？ - **核心价值观绑定**：AI 伦理审查是否真正融入产品设计，而非浮于表面的事后补救？ - **端到端可观测性（Observability）**：系统是否具备完善的实时监控与安全护栏机制？ ### [18:27](https://www.youtube.com/watch?v=wEymOOayGPI&t=1107s) 密歇根大学师生的多样化 AI 实践案例展示环节通过一段生动的视频集锦，呈现了密歇根大学（University of Michigan）师生与教职员工在日常教学、学习、科研和管理中深度融入 AI 的广泛实践。师生们纷纷表达了拥抱前沿技术的初衷——因为这些工具正在深刻塑造人类的未来。在教学创新方面，有教授利用学校提供的丰富 AI 工具构建了自己的“数字分身”（Digital Twin），该模型基于其讲授的课程内容进行训练，能够实现全天候 24/7 为学生答疑解惑、辅导课后练习，为课堂教学注入了前所未有的高互动性。行政与运营管理方面，AI 也被深度应用于采购流程，实现了更具前瞻性和高效率的决策与物资采购。学生和医疗教育工作者也分享了他们切实的获益体验。学生们利用密歇根大学自研的 AI 平台 Maizey 整理复习笔记，高效备战期中与期末考试；工程团队学生将其整合至工程团队资源中心，实现了按需自主学习。在医学与健康领域，AI 负责将复杂的医疗健康数据转化为具有可操作性的洞察，在全球资源匮乏的背景下推动高效、公平且切实的医疗护理，医学院教师甚至在课堂中引入了“虚拟病人”（Virtual Patient）供学生模拟实操互动。此外，还有教师利用 AI 搭建了连接 Google Drive 的虚拟助教，专门自动回复“今天有答疑时间（Office Hours）吗？”或“可以申请延期交作业吗？”等高频行政咨询，大幅减轻了邮件负担；更有学者与学生借助 AI 激发创新灵感、攻关技术难题，快速推进学术科研进程。 ```mermaid flowchart TD accTitle: 密歇根大学全场景 AI 应用架构图 subgraph EDU["教学与学习赋能"] A["教授数字分身 (24/7 答疑与互动)"] B["Maizey 工具 (笔记整理/备考/工程团队学习)"] C["虚拟助教 (自动处理日常答疑与申请)"] end subgraph HEALTH_RESEARCH["医疗健康与科研探索"] D["虚拟病人模拟 (临床实操教学)"] E["复杂健康数据分析 (推动公平医疗与洞察)"] F["科研攻坚与创意激发 (加速技术突破)"] end subgraph OPS["校务与行政运营"] G["智能采购系统 (高效决策与物资流程)"] end ``` - **全天候个性化教学**：教授通过数字分身与 Google Drive 虚拟助教，接管了繁琐的例行答疑与排班咨询，极大释放了师生沟通的时间成本。 - **自主学习与工程协作**：学生借助自研平台 Maizey 提炼核心知识点、准备考试，并将 AI 工具作为团队知识库的即时检索中枢。 - **医疗与科研实效**：涵盖虚拟病人问诊演练、全球稀缺资源下的医疗健康数据挖掘，以及面向尖端科研课题的技术排障与灵感激发。 --- ### [20:41](https://www.youtube.com/watch?v=wEymOOayGPI&t=1241s) 负责任 AI 的核心使命与结语寄语在全场演讲的尾声，Ravi Pendse 总结了密歇根大学的实践探索，并向全体参会者致以谢意。他着重呼吁大家在离开会场时牢记一个核心准则："Remember, this is all created by humans for humanity"（这一切都是由人类创造，并为了全人类的福祉而存在）。技术的终极价值在于每天切实地改变甚至挽救生命。他强调，迈向负责任的人工智能之路并非一蹴而就，每个人肩上都有明确的责任，唯有通力合作，各界才能真正将负责任 AI 的治理框架与道德蓝图化为现实。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=wEymOOayGPI&t=4s) 坚持到底的人终将成为冠军（Those who stay will be champions）。 — 博·辛巴克勒（Bo Schembechler / 由拉维·潘德塞引用） > [00:04](https://www.youtube.com/watch?v=wEymOOayGPI&t=4s) Acharya Devo Bhava（尊师如神）。 — 拉维·潘德塞（Ravi Pendse） > [00:04](https://www.youtube.com/watch?v=wEymOOayGPI&t=4s) 以人为中心的 AI（Human-centered AI），人必须始终处于回路的核心。 — 拉维·潘德塞（Ravi Pendse） > [02:24](https://www.youtube.com/watch?v=wEymOOayGPI&t=144s) 为沉默者发声。 — 拉维·潘德塞（Ravi Pendse） > [02:24](https://www.youtube.com/watch?v=wEymOOayGPI&t=144s) AI 的所有弊端（Everything That's Wrong About AI）。 — 拉维·潘德塞（Ravi Pendse） > [05:05](https://www.youtube.com/watch?v=wEymOOayGPI&t=305s) AI + You > You — 拉维·潘德塞（Ravi Pendse） > [05:05](https://www.youtube.com/watch?v=wEymOOayGPI&t=305s) AI + You > AI — 拉维·潘德塞（Ravi Pendse） > [09:01](https://www.youtube.com/watch?v=wEymOOayGPI&t=541s) 负责任的 AI 必须以人为中心（Responsible AI has to be human-centered）。 — 拉维·潘德塞（Ravi Pendse） > [11:54](https://www.youtube.com/watch?v=wEymOOayGPI&t=714s) 谁掌控你的数据，谁就掌控你的未来。 — 拉维·潘德塞（Ravi Pendse） > [15:38](https://www.youtube.com/watch?v=wEymOOayGPI&t=938s) 从海量噪音中提炼真实信号（Separating signal from noise）。 — 拉维·潘德塞（Ravi Pendse） > [20:41](https://www.youtube.com/watch?v=wEymOOayGPI&t=1241s) Remember, this is all created by humans for humanity（这一切都是由人类创造，并为了全人类的福祉而存在）。 — 拉维·潘德塞（Ravi Pendse） ## Mentioned 拉维·潘德塞 (Ravi Pendse), 密歇根大学 (University of Michigan), 博·辛巴克勒 (Bo Schembechler), 李飞飞 (Dr. Fei-Fei Li), 杰弗里·辛顿 (Geoffrey Hinton), 托德·霍隆 (Dr. Todd Hollon), FastGlioma, Maizey, Maizey App Store, OpenAI, Llama, GPT-2, Google Drive, 《自然》期刊 (Nature) ## Video description Ravi Pendse - Vice President for Information Technology and Chief Information Officer, University of Michigan The closing keynote shifts from debate to action — translating the day's policy conversations into a concrete framework for responsible AI leadership. Against a backdrop of federal deregulation and international regulatory divergence, this session asks what governance structures, accountability mechanisms, and leadership commitments truly separate principled AI adoption from performative compliance — and closes the summit with a practical blueprint. [Source](https://www.youtube.com/watch?v=wEymOOayGPI)

Source quotes

[0:04] 博·辛巴克勒（Bo Schembechler / 由拉维·潘德塞引用）: 坚持到底的人终将成为冠军（Those who stay will be champions）。 [0:04] 拉维·潘德塞（Ravi Pendse）: Acharya Devo Bhava（尊师如神）。 [0:04] 拉维·潘德塞（Ravi Pendse）: 以人为中心的 AI（Human-centered AI），人必须始终处于回路的核心。 [2:24] 拉维·潘德塞（Ravi Pendse）: 为沉默者发声。 [2:24] 拉维·潘德塞（Ravi Pendse）: AI 的所有弊端（Everything That's Wrong About AI）。 [5:05] 拉维·潘德塞（Ravi Pendse）: AI + You > You [5:05] 拉维·潘德塞（Ravi Pendse）: AI + You > AI [9:01] 拉维·潘德塞（Ravi Pendse）: 负责任的 AI 必须以人为中心（Responsible AI has to be human-centered）。 [11:54] 拉维·潘德塞（Ravi Pendse）: 谁掌控你的数据，谁就掌控你的未来。 [15:38] 拉维·潘德塞（Ravi Pendse）: 从海量噪音中提炼真实信号（Separating signal from noise）。 [20:41] 拉维·潘德塞（Ravi Pendse）: Remember, this is all created by humans for humanity（这一切都是由人类创造，并为了全人类的福祉而存在）。

Source quotes

[0:04] Bo Schembechler: Those who stay will be champions. [0:04] Ravi Pendse: Acharya Devo Bhava [2:24] Ravi Pendse: Everything That's Wrong About AI [5:05] Ravi Pendse: AI + You > You [5:05] Ravi Pendse: AI + You > AI [9:01] Ravi Pendse: Responsible AI has to be human-centered. [11:24] Ravi Pendse: Maize & Blue App Store [15:38] Ravi Pendse: the singularity [20:41] Ravi Pendse: created by humans for humanity

Original transcript

[0:04] Ocean Beckler, he was one of our very famous football coaches. He said this in 1969. He said, [0:11] "Those who stay will be champions." [0:14] So, thank you. You guys are all AI champions in my book because you stayed for my my last [0:19] presentation here. So, so we'll have a great How many of you attended keynote this morning? [0:24] Uh so some of you quite a few of you did, right? I just want you to know just for clarity's sake, [0:28] uh I have never worked for Sanjay or Sanjay has never worked for me, okay? Uh it'll be an [0:35] honor to work for Sanjay or for all of you, but I just want to clarify considering some bizarre [0:39] conversation that happened in the middle of keynote this morning. Uh there were egos at work. [0:44] At Michigan, we don't have egos. We we we approach everything with humility. So, I stand before you [0:49] in utmost humility because I don't consider myself an AI expert. Uh my In my other life, [0:56] I'm a professor in computer science. Uh I consider myself a student of AI. In my humble opinion, [1:02] with due apologies to any experts in the room, I believe that everyone should be a student of AI [1:08] considering how rapidly this field is growing. I start all my presentations with a Sanskrit phrase [1:16] that my mother taught me. And it goes something like this, "Acharya Devo Bhava." Acharya is [1:22] teacher. Teacher is made in the image of God and should be respected as one would respect God. [1:28] And I could have had one of the LLMs do the translations for you of that Sanskrit phrase, [1:34] but I actually believe in human-centered AI. Human at the center of loop. You heard Dr. [1:39] Fei-Fei Li also talk a lot about human-centered AI. At Michigan, we have taken the approach of [1:46] approaching everything with human-centered AI. So, you're You may be wondering, [1:50] "What have you done at Michigan?" Uh I want you to know, and I say this with utmost [1:55] utmost humility because I'm representing the work of some great colleagues at Michigan. [1:59] Uh We were the first organization, not just higher education institution, first organization [2:05] in the country to launch generative AI services at scale to the entire institution, 140,000 people, [2:13] in August of 2023. So, we've been at it for more than 3 years, and the work continues. [2:20] And I'll share some of our work, ongoing work, as well, [clears throat] as we go forward. [2:24] One of the questions I often ask people is, as we have this conversation is, what's top of mind [2:29] for you when you think about AI? When you think about responsible AI? What is top of mind for you? [2:35] And if we had many hours, we could have debate. I'll take input from you, but we don't. So, [2:39] what I did was I did the next best thing. I stalked many of you in gyms and other places [2:45] and elevators and asked you, "What's top of mind for you as you think about responsible AI?" [2:49] What I heard was, for example, you're worried about AI agents and guardrails, [2:55] that we need to have proper guardrails. I also heard the jobs and jobs losses we heard today. [3:01] You had fears about AI. I also heard, for example, the the idea, the negative impact AI might have [3:09] on our critical thinking skills. So, I heard many, many stories like this. The key there is, [3:15] when you talk about responsible AI in your respective organizations, does your [3:20] responsible AI include a strategy where you're listening to all the voices of your organization? [3:25] At Michigan, we do. We take great pains to ensure that everybody in our community is being heard, [3:32] including giving voice to silence. Because you'll always have a lot of AI people. You'll have people [3:38] on one other end saying, "Oh, we should stop all this madness called AI." Then there'll be a middle [3:43] minority middle group that may not be voicing their opinions. So, what are you doing in your [3:49] organizations to ensure that their opinions are being heard and voiced. For example, in April [3:54] of this year, at Michigan we hosted a full day summit titled everything that's wrong about AI. [4:01] It was sponsored by my office. We invited all the people who were opposed to AI [4:06] to come and share their ideas with us. We did not think we'll be changing any minds, [4:10] but our goal was that maybe we'll understand each other a little bit better. We'll We'll be [4:15] doing one more session summit like this in fall because it was so popular and so well received. [4:21] In your respective organizations, as you think about responsible AI, [4:24] do you have a culture where if you were to host a summit like this that it will be well received? [4:29] And if not, my friends, you have work to do when you think about responsible AI. [4:34] Because just words are not enough, your actions have to follow in my humble opinion. And so, [4:38] bottom line is this, when you think about any AI strategy, it has to be transparent and inclusive. [4:45] For example, if your organization has an AI strategy, do you know what it is? [4:50] And if not, then you should be making efforts. And if you're one of the leaders, if you work [4:54] for your CEO or you are the CEO, then you have to make sure that you have transparent and inclusive [4:59] AI strategy. At Michigan we strive to do that. We are very open, we share information freely. [5:05] And when we talk about AI and what we did with our implementation, you know, [5:10] we are curious. This morning you heard Geoffrey Hinton talk about how, for example, young kids, [5:17] if they're given an AI tutor, they can uh learn twice the speed because they're very curious. They [5:23] ask a lot of questions. You heard him say that this morning, right? At Michigan we deployed AI [5:27] tutors for faculty members who wanted it in August of 2023. It takes 4 minutes for a faculty member [5:34] to configure an AI tutor for their class should they choose to do it. It's available for everyone. [5:38] So, we did a lot of small experiments and it gives us a place to begin. So, don't let uh you know, [5:46] AI can be overwhelming, but small experiments give you a start. I wanted to share I can share about [5:51] 100 stories, but I want to share one story of a small experiment at Michigan, and this relates to [5:57] uh a dear friend of mine and an amazing neurosurgeon named Dr. Todd Holland. Todd Holland, [6:03] all of this work has been published in Nature and other prestigious journals, so this is all [6:07] documented, so you can check that out later. Todd and his team developed a tool called fast glioma. [6:15] Fast glioma. That's the tool that they developed. And prior to that, when Todd and his team would [6:23] operate on patients, patients who have malignant tumor in their brain, one of the challenges was [6:29] that often times surgeons cannot be very aggressive in removing the tumor because if you're [6:36] too aggressive, you can lose some bodily functions if you take out good tissue by mistake. So often [6:41] surgeons would be very conservative, so they would often send patients back, this is global, [6:46] with about 25% tumor remaining still in the brain that could be malignant. So the technique [6:51] that they developed using AI essentially does the following. While the surgery is happening, [6:57] in progress, continuous images are taken of the brain, that information is sent to our [7:02] data center, it is analyzed in real time using machine learning AI high-performance compute, [7:07] and in under 10 seconds, the information comes back to Todd Holland and his team uh [7:16] giving them information about how much malignant tumor is remaining. Using this technique, [7:20] they have been able to reduce the amount of tumor that patients are sent back, malignant tumor, [7:25] from 25% down to 4%. 25% down to 4% using AI. That's responsible AI. Managing data properly, [7:36] data security, patient data is never compromised. Uh appropriate information is processed, [7:41] timely feedback is back in the hands of surgeons, and really what this results in is many many lives [7:48] saved. So, when our CFO or others want to talk to me about ROI, there is your ROI for you, okay? [7:56] Saving lives. Who can argue with that, right? Many many lives saved. So, bottom line is this, all of [8:01] this is been created by humans, all of us, all of you. Human-centered computing, human-centered AI. [8:08] For humanity. Transforming and saving lives one person at a time, one person at a time. That's [8:15] the power of AI if used responsibly. And when I talk about human-centered AI, I often talk about [8:21] AI plus you as greater than you. It is augmenting you. It's not replacing you. The you still exist. [8:28] And AI plus you can be greater greater than AI because as you work with AI, [8:33] you're actually making that AI better to better help you. The bottom line is you. You are the [8:39] center. So, when people talk to you about job losses, when people in my humble opinion, [8:44] the human ingenuity that each one of us brings, the uniqueness that each one of us brings [8:50] is not replaceable. AI is never going to be able to replace uh So, Sanjay doesn't like waffling. [8:56] How is that for a statement, okay? Uh It's never going to be able to replace, bottom line. [9:01] So, one of the other questions I often ask is, you know, what's your AIQ? I ask this question to my [9:07] students, I ask this question to my community, my staff members, others. And when people see AIQ, [9:12] they say, "Oh, are do you mean prompt engineering? Do you mean this?" No, no, no, I mean a lot of [9:16] different things. Number one, of course, we need to know and be aware of all the different [9:20] AI tools that exist. Second, we also need to know when not to use AI. When not to use AI. [9:30] I don't need you to raise a hand, but I bet anytime these days anybody asks you a question, [9:34] or if you're trying to do anything, the temptation is to go ask your favorite LLM. [9:40] I would encourage you to with intentionality try not to do that. Don't lose your thinking skills. [9:46] That negative impact on critical thinking skills is really really important, especially [9:51] if you have kids in the house, they model your behavior. And so so you need to think about that. [9:57] When we talk about AIQ, we also talk about the negative impact AI could have on our environment. [10:02] On you know, the bias that exists in in our society and AI amplifies that bias. So my dream is [10:08] any student that graduates from any institution of higher education should have the highest possible [10:15] AIQ, which means they should understand all of this. Isn't that a great goal for [10:20] our organization too? Your staff members, your employees, your colleagues, we should all be [10:25] thinking about this way when we think about AIQ. So the bottom line is don't use this [10:29] as just something you just use it for everything. Be very intentional and focused about when to use [10:35] and when not to use AI. At Michigan, we preach this, we practice it, we show examples every day. [10:41] The bottom line is when you do this, you don't just adopt AI, you're helping shape it then. [10:46] You have the power to shape AI. Because if you just sign up with a vendor start using a product, [10:52] are you actually helping shape that AI? You need to be in the middle [10:57] and human judgment must remain central while you're shaping the AI. Otherwise, [11:01] somebody else will do it for you. And you can do that. And when it comes to responsible AI, [11:07] it has to be human-centered. Responsible AI has to be human-centered. Would you indulge [11:13] and say with me all of you? Responsible AI has to be human-centered. Can you all say that once? [11:19] Let's go, good students. Thank you. Thank you. Thank you for indulging me on that one. And [11:24] you know, as organizations, we have two choices. Choice number one is we can wait and let vendors [11:30] and outside forces define our future. The other choice is we can do the small experiments we [11:36] talked about. We can learn, we can establish those guardrails that we talk about, and help shape the [11:42] future. Michigan has chosen to take that approach. The bottom line is this, it's created by humans [11:49] for humanity transforming lives. And of course you heard a lot about agentic AI and agentic [11:55] AI sphere. So what Michigan has been doing in general, our Michigan ecosystem looks like this. [12:01] Uh we have multiple tools available. One of our intellectual properties is called Mazy. That's a [12:07] code less platform that's built at Michigan on top of LLMs. But the way we approached it was [12:14] we are vendor agnostic. I can dump any vendor tomorrow. I can use open source models. I can [12:19] use cloud. I can use uh you know, open AI. I can use llama four. All of that's available. [12:26] So the question I have for you is at your different organizations, do you have a vendor [12:32] agnostic strategy? Because each time a model comes up with an upgrade, if you're saying uh-oh, [12:39] then you may have a problem. You may want to think about that. The other thing is because of [12:44] all the circular funding that's going on among the key players, there's always concern around [12:50] are we in a world of Enron bubble? And if we are, I don't want the future of Michigan, [12:57] the future of incredible students all across the world to be dependent on that one LLM. [13:03] And that's why we have always taken this approach. Uh and I'm happy to get into technical details [13:07] with you to whatever details you want. We can do whiteboard sessions or come visit us in Ann Arbor. [13:13] Maybe not in winter, but you're welcome in winter as well. Uh and we can have fun together. [13:17] The key is when we did all of this, our data is private because we believed that [13:23] data sovereignty is very, very important. We did not want it share our data with anybody. We have a [13:28] large health system. We sit on gold mine of data. That data is being used to change and save lives [13:34] every day. So we didn't want to share data with anyone. So everything that's built at Michigan, [13:38] all of our data stays at Michigan. That's again a key parameter to responsible AI. The bottom [13:44] line is those who control your data are going to control your future. So you got to ask yourself [13:50] that question and think about it. So responsible AI requires a thoughtful data strategy. [13:56] AI doesn't exist without a thoughtful data strategy. And so you got to be focused on that. [14:02] So going forward in October, we can have a big announcement coming out of Michigan. We've been [14:06] working on it for a while. We're actually going to launch a Maize and App Store. Just like your [14:12] Apple App Store. A Maize and App Store that will be launched that will be Michigan branded to be [14:16] useful for Michigan community initially. And think about it. Michigan has brilliant 63,000 [14:22] citizen developers developing agents to be useful for that humanity. Created by humans [14:30] for humanity transforming and saving lives. That's what our kids will do [14:35] because they will have the highest possible AIQ hopefully if we do all our things right. That's [14:39] the question you should all be asking yourself. And bottom line is this, we are creating an [14:44] environment where we don't need to have subject matter expertise. A procurement expert can do [14:49] procurement agent. A nurse A nurse can do a an agent for helping nurses. And I could go [14:55] on and on and on. That's the environment we have created at Michigan and we're building. But again, [15:00] this is built for an organization not on a around any specific model. Models will come and go. [15:06] You know, we we we are using open source models. All of the ones that were mentioned and discussed [15:11] here. And that's what open way does well. But that's the idea. Models will come and go, [15:16] but we need to focus on that. And when we talk about responsible AI, it's not one thing. [15:21] It is all of these things. It's ethics. It's all of these things. The bottom line is this, and I'm [15:26] being sensitive to time. When you think about AI, you if you choose to ignore responsible AI, [15:33] you do so at your own peril. You do so at your own peril. So, this is the message you need to [15:38] give back to your organization. Now, I'm sure you've heard lots of these headlines, right? [15:44] Oh, singularity is here, okay? Many people have done this, different people. And what does the [15:52] community hear? Community hears all of these things. When people make these huge blanket [15:58] statements with no context, community So, when you talk about this opposition to data center, [16:04] when you talk about opposition to AI, that's the result of people not practicing responsible AI. [16:10] Sure, talk about singularity is here, but take time to explain what that actually means. [16:15] Don't just make bunch of statements that could get picked up by media very quickly. [16:19] You've heard these statements, too. This was from 2019, by the way. Our model is so kind of [16:25] this thing that we're just not going to release it to anybody. This was in 2019, by the way, [16:30] at the time of ChatGPT 2.0, okay? And then then, of course, other stories followed, [16:35] right? With Metis, we heard similar story, right? And then we heard this about the [16:41] agents breaking into hacking into different things, and so on. And then we heard this, okay? [16:47] Responsible AI requires separating signal from noise. Does your organization have the ability to [16:52] separate signal from noise? So, as you go back to your respective homes, organizations Monday, here [16:57] are a few questions I want you to think about. And I'm not going to read every one of them, [17:02] but is human judgment central to your AI strategy? I would urge that you need to make sure that is a [17:06] priority. That human-centered AI is critical to what you do in your organizations. What about [17:12] sovereign data strategy? You need to make sure you have that. What about keeping your implementation, [17:18] everything you're doing vendor agnostic, where you build your own orchestration such that you [17:23] are not dependent on any vendors? What about Are you ready to handle signals and a lot of noise? [17:31] Singularity is here. Tomorrow there's going to be one more statement made by somebody. Can you [17:36] differentiate between signals and noise? Does your organization has a culture of separating signals [17:41] from noise? If not, we have all work to do when you go back. Uh what about ethics? How are you [17:46] integrating that in your AI strategy? Or is ethics an afterthought? It cannot be an afterthought. [17:51] Each organization, each company, you have some core values. That should be tied to whatever [17:56] ethical considerations you have. Uh Michigan has our core values, so we stick to those when we [18:00] think about ethics. Observability is critical when you think about security and AI overall. [18:06] So, these are some of the things you should be thinking about. The bottom line is this, [18:10] when we see say this, it's built for you know, today your your infrastructure, [18:16] but it has to be ready for tomorrow. It's built for today, has to be ready for tomorrow. I have [18:21] about a 3-minute video and have 1 minute left. What's the verdict? Show video or don't show? [18:27] Okay, all right. So, here we go. This shows some of the work that's going on at Michigan. I use AI. [18:43] I use AI because it's important to me to work with the tools that are going to shape our future. [18:49] I use all the great AI tools here at Michigan to create a digital twin of myself, [18:53] trained on the courses that I teach, which I make available to students, [18:57] answer their questions 24/7, practice assignments with them, and inject [music] more interactivity [19:04] into the classroom than ever before. I use AI for more informed and efficient procurement process. [19:10] I've used U-M Maize to compile my notes and use them to study for midterms and finals. [19:17] I use AI when I integrate Mayzie into the student engineering team resource hub to enable [19:24] on-demand learning. I use AI to transform complex health data for actionable insight, [19:32] fostering meaningful impact in the communities we serve, [19:36] enabling efficient, effective, and equitable care, especially when resources are limited globally. [19:43] I use AI a lot to help make complex ideas simpler when I especially when I'm learning new concepts. [19:50] I use AI in the classroom to allow students to have an interaction with a virtual patient. [19:57] I use AI to create a virtual assistant for my Google Drive. It answers students' questions like, [20:02] "Are there office hours today?" and "Can I have an extension?" It cuts down on a lot of emails. [20:08] I use AI to help inspire my creativity. [20:12] AI empowers my students to work through technical challenges and rapidly advance their research. [20:41] So, let me stop there and um thank all of you and I just want to remind you one thing as you leave. [20:47] Remember, this is all created by humans for humanity, [20:51] transforming and saving lives every day. So, we all have work to do and together we can [20:56] implement responsible AI. Thank you for indulging me. [applause]