Human‑Centered Design for Responsible AI

Ai4

AI-assisted summary of Human‑Centered Design for Responsible AI — original source. Check the original for full context.

Source published: 2026-08-17

Detailed notes

## Background Published by Ai4 on YouTube in August 2026, this presentation features Cinnamon Clark, Chief Diversity, Equity, Inclusion, and Belonging Officer for a California state government agency. The talk outlines practical approaches for embedding DEIB, accessibility, and civil-rights alignment across the AI lifecycle to ensure automated systems remain fair, understandable, and supportive of real people.

## TL;DR Cinnamon Clark argues that AI systems must be intentionally designed around human variation and lived experiences rather than evaluated solely by algorithmic outputs. She outlines a core design framework that prompts teams to examine intended user impacts, accessibility hurdles, and generational communication differences before building automated tools. Highlighting high-stakes harms, such as algorithmic emergency dispatching and the wrongful arrest of Robert Williams via flawed facial recognition, Clark demonstrates the severe cost of ignoring civil rights and user diversity. Ultimately, she calls for proactive cross-functional collaboration, transparent communication, and early accessibility testing to foster organizational trust and deliver reliable technology.

## Atmosphere & style Clark sets an approachable, conversational, and warmly candid tone, using self-deprecating humor—such as joking about a wobbly stage chair—and relatable workplace anecdotes about generational punctuation to disarm technical audiences. The presentation fluidly transitions from engaging, practical observations to somber urgency when addressing the tangible civil rights harms of unchecked algorithmic bias. The session feels like an empowering, clear-eyed strategic briefing that bridges the gap between technical engineering priorities and lived human realities.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=4s) Human-centered AI design focuses on the end users who live with the consequences of automated systems, rather than solely optimizing algorithms and datasets. - [02:14](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=134s) Multidisciplinary teams should evaluate five core design questions upfront regarding who will use the tool, its intended benefits, potential unintended harms, and cross-functional integration. - [02:14](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=134s) DEIB, HR, and civil-rights experts must be integrated from the beginning of development rather than reactively called in after an algorithmic failure or PR crisis. - [04:33](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=273s) Because 'data in is data out,' missing or non-representative data creates high-stakes blind spots in automated hiring, performance evaluation, and emergency medical dispatching. - [07:36](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=456s) Accessibility must be proactive rather than an afterthought, accounting for assistive technologies like screen readers, language translation, and cultural idioms. - [07:36](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=456s) Lived experiences and generational communication norms diverge significantly, causing the same digital prompts and punctuation to be interpreted in radically different ways. - [11:05](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=665s) Accounting for human physical, cognitive, and infrastructural variation creates a virtuous cycle that reinforces business robustness, error tolerance, usability, and trust. - [11:05](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=665s) Operational availability does not guarantee true accessibility if institutional gatekeeping or biased supervisory approvals prevent employees from participating. - [15:31](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=931s) Embedding DEIB into AI development is not a zero-sum trade-off; creating equitable tools for vulnerable populations elevates broader productivity, outcomes, and GDP. - [18:02](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=1082s) Deploying biased algorithms without civil rights validation causes severe harm, as seen in the wrongful arrest of Robert Williams caused by facial recognition with a known high error rate on Black faces. - [19:24](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=1164s) Failing to communicate transparently about AI tools triggers workplace communication breakdowns, dismantles employee trust, and increases costly turnover. - [20:10](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=1210s) To ensure equitable outcomes, organizations must explicitly identify who tools are not designed for, shift accessibility checks to the design phase, and test across diverse cohorts.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=4s) Introduction and the Foundational Reality of Human-Centered Design

Cinnamon Clark opens the session by introducing her role as the Chief Diversity, Equity, Inclusion, and Belonging (DEIB) Officer for a state government agency in California. Setting an educational and collaborative tone, Clark expresses appreciation for the opportunity to have open discussions among professionals with dissimilar backgrounds. In her capacity as an organizational strategist, she emphasizes that her objective is not to forcefully change minds, but rather to foster awareness and education around designing automated systems that operate reliably for real people. Too often, discussions in the artificial intelligence space focus almost exclusively on data sets, algorithms, and high-level outputs, completely neglecting the end user who must live with the consequences of those systems.

Clark establishes three baseline disclaimers to frame the session. First, all insights shared represent her own professional experience and education rather than the official stance of her government employer. Second, the presentation is meant to spark dialogue rather than serve as a rigid, step-by-step rulebook, because "there's no one set of rules or explanations to human problems." Third, she reminds the audience of a fundamental truth: humans are humans, regardless of whether they operate in high tech, healthcare, or government.

- Clark's core objective as a strategist centers on education and awareness rather than forcing perspective shifts. - Human problems resist simplistic, formulaic solutions, requiring flexible, human-centered approaches across all industry sectors.

---

[02:14](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=134s) Five Crucial Questions to Ask in the Design Phase

While managing a wobbly stage chair with humor—promising that if she falls, everyone can laugh once she gets past the initial embarrassment—Clark introduces the core framework she relies on in policy, legislation, and public service. Before building or deploying automated tools, multidisciplinary product teams must interrogate five fundamental questions:

```mermaid flowchart TD accTitle: Five Core Questions for Human-Centered AI Design subgraph Discovery ["1. Target Audience & Impact"] Q1["Who is going to use this product or tool?"] --> Q2["How will what we develop impact those users?"] end subgraph Evaluation ["2. Value vs. Harm"] Q2 --> Q3["How will the technology improve their lives?"] Q3 --> Q4["How might it introduce unintended challenges or harm?"] end subgraph Delivery ["3. Lifecycle Integration"] Q4 --> Q5["Are collaborative teams engaged from inception to deployment?"] end ```

Clark stresses that human resources, civil rights, and DEIB professionals are historically brought in only after disaster strikes—when an employee files a formal complaint, an algorithm misbehaves, or a major credibility, trust, or public relations crisis hits the press. Integrating cross-functional perspectives early ensures teams proactively identify potential harm before products reach the public.

- Teams must continually evaluate the fundamental "who, what, when, where, and why" of their AI systems. - Proactive engagement of HR and compliance teams prevents downstream credibility, trust, and PR fallout.

---

[04:33](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=273s) High Stakes: How AI Decisions Impact Real Lives

Technology is not designed for technology itself; it is created by and for human beings, who ultimately choose which solutions make their lives easier. Organizations across sectors are rapidly integrating automated decision-making engines into high-stakes operational workflows. AI tools are already screening job applicants, managing onboarding, deciding promotions, scoring worker performance, and prioritizing dispatch services in municipal Emergency Medical Services (EMS) and fire rescue operations based on neighborhood data.

Because the data feeding these models directly dictates the outcomes generated ("data in is data out"), any lack of representative data or diverse perspectives during development creates dangerous blind spots. Clark points out that algorithms are already orchestrating everyday life, from social media feeds to the conference app scheduling her itinerary based on registered interests. When technology dictates access to livelihoods and emergency services, creators carry an ethical obligation to center the full spectrum of human experience.

- AI algorithms currently determine real-world outcomes in hiring, performance ratings, and municipal EMS emergency dispatching. - Biased or non-representative input data directly translates into biased real-world outputs.

---

[07:36](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=456s) Understanding User Diversity, Generational Gaps, and Lived Experience

Centering the human experience requires builders to thoroughly understand the practical realities of their users, including physical abilities, language, and cultural context. For instance, digital platforms that fail to support screen readers effectively lock out low-vision users from vital public information. Similarly, systems relying heavily on cultural idioms, slang, or localized pop culture references alienate non-native English speakers, immigrants, and indigenous populations.

To illustrate how lived experience and communication norms diverge, Clark shares an anecdote from her doctoral dissertation research on generational workplace communication. As a member of the "Xennial" cohort (straddling Gen X and Millennials), Clark notes that when she confirms an assignment over text, she types `"Sure."` followed by a period, intending it as an efficient signal that means "I'm on it, boom." Her younger colleagues, however, interpret the punctuation as passive aggression, asking, "Why is there a period? I thought you were mad at me." Lived experience also shapes institutional access:

> "I do not experience the organization as an executive, as somebody that's on the entry level of the organization. Because if I ask for my computer to be fixed, people are jumping out of the woodwork to make that happen. My assistant asks for that, and she may not get an answer for weeks."

- Interface designs must accommodate assistive technologies like screen readers rather than treating accessibility as an afterthought. - Generational and cultural variations in language and punctuation radically alter how messages and prompts are received. - Organizational hierarchy creates blind spots; executive user experiences rarely match the day-to-day friction faced by entry-level personnel.

---

[11:05](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=665s) Designing for Human Variation and the Virtuous System Cycle

Designing inclusive AI requires accounting for physical, cognitive, environmental, and infrastructural differences. Clark jokes that if a workplace requires running five miles a day, "this girl is not running 5 miles a day." Cognitive diversity is equally vital; as neurodivergence is increasingly destigmatized in professional settings, AI workflows must accommodate non-linear ways of thinking. Additionally, structural realities—such as an employee being marked tardy because public transit lacks infrastructure to reach a remote suburban office—must be factored in rather than penalized by automated monitoring tools.

Clark details how accommodating human variation fuels a continuous, virtuous cycle across five core system pillars:

```mermaid flowchart LR accTitle: Virtuous Cycle of Human Variation in AI Systems A["Business Robustness"] --> B["Error Tolerance"] B --> C["Usability"] C --> D["Accessibility & Availability"] D --> E["Reliability & Trust"] E --> A ```

- **Business Robustness & Error Tolerance**: Diverse inputs expand the system's operational parameters, catching edge cases before deployment. - **Usability**: Clark cites the frustrating cycle of iPhone updates that immediately require two subsequent patches to repair basic usability flaws as an example of poor design foresight. - **Accessibility vs. Availability**: A corporate professional development program may be officially "available," but if participating requires sign-off from two biased supervisors, it remains entirely inaccessible to certain employees. - **Reliability and Trust**: Demystifying AI is crucial for bridging trust deficits across generations, such as addressing the skepticism of older generations who view AI as a monolithic threat despite using automated tools for years.

---

[15:31](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=931s) Practical DEIB Definitions and the Non-Zero-Sum AI Advantage

Clark reframes Diversity, Equity, Inclusion, and Belonging (DEIB) as operational necessities rather than ideological concepts, noting that building inclusive systems has everything to do with expanding usability and audience reach across geographic and demographic lines:

| DEIB Pillar | Core Definition | Concrete Workplace & AI Application | |:--- |:--- |:--- | | **Diversity** | The collective differences among people, encompassing both innate and acquired traits. | Expands data set representativeness and catches systemic algorithmic blind spots. | | **Equity** | Providing individuals with the specific, tailored resources they need to thrive. | Flexible work schedules, ergonomic accommodations, and fair distribution of critical services. | | **Inclusion** | The active practice of ensuring people feel invited and integrated. | Clark illustrates this with a relatable analogy: "Men, you ever been invited to a baby shower? What's the first question you ask? Are there other guys there?" | | **Belonging** | The psychological safety of feeling accepted and valued as an authentic participant. | Strengthens institutional trust, leading to measurable gains in workforce retention and productivity. |

Embedding equity throughout the AI lifecycle is not a zero-sum trade-off. Referencing earlier remarks from Dr. Brown, Clark reiterates that designing technology to support underrepresented or vulnerable populations elevates outcomes, productivity, and economic value for everyone.

- Representative datasets improve equity in automated municipal resource allocation, such as ambulance and fire dispatching. - Inclusive design directly contributes to broader economic indicators, including increased gross domestic product (GDP) and organizational productivity.

---

[18:02](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=1082s) Civil Rights, Algorithmic Harm, and the Detroit Police Case Study

Clark warns that ignoring legal compliance, civil rights, and disparate impact during development leads to catastrophic real-world harm. To demonstrate the danger of deploying flawed predictive technologies, she points to the wrongful arrest of Robert Williams in Detroit. Williams, a Black man, was arrested on his front lawn in front of his wife and daughters after an automated facial recognition algorithm misidentified him from blurry surveillance footage.

Crucially, the Detroit Police Department deployed the facial recognition tool despite knowing the software suffered from an estimated 96% error rate when attempting to identify Black faces. Clark highlights that among technical and engineering audiences, proceeding with deployment despite such a catastrophic failure rate represents an unacceptable breakdown of basic ethics and oversight. She concludes by reminding creators that simply making a tool available does not mean it is accessible, reliable, or safe, urging technologists to question potential bias and civil rights impacts early and often throughout the design process.

- Robert Williams was wrongfully arrested in Detroit after facial recognition erroneously matched his photo to surveillance footage. - The facial recognition software was deployed despite internal knowledge of an estimated 96% error rate on Black faces. - Failure to validate algorithms against civil rights standards creates severe legal liabilities and inflicts profound personal harm on innocent users.

[19:24](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=1164s) The Cascade of Communication Breakdowns and Organizational Costs

Failing to communicate openly and frequently about automated tools and AI initiatives triggers a direct, damaging chain reaction inside an organization. When leaders do not routinely discuss what systems are being implemented, how those systems impact the workforce, and how they address fundamental questions of fairness and purpose, a critical opportunity is lost. This silence inevitably breeds workplace communication breakdowns, which in turn dismantle organizational trust.

Once trust breaks down, employees become disengaged and actively look to leave the organization. This attrition drives up turnover and inflates recruitment, hiring, and onboarding costs, creating an expensive cycle of organizational disruption. To prevent this compounded cost, teams must proactively communicate early and often, intentionally involving stakeholders and collaborating across departments before systems are deployed.

```mermaid flowchart TD accTitle: The Cost of Communication Breakdown in AI Deployments A["Failure to Communicate AI Intent & Impact"] --> B["Workplace Communication Breakdown"] B --> C["Erosion of Organizational Trust"] C --> D["Increased Employee Attrition"] D --> E["Rising Hiring & Replacement Costs"] ```

- **Early and frequent communication:** Engaging staff early on prevents rumor and alienation, ensuring that stakeholders understand the intent, impact, and safeguards behind new tools. - **The compounding cost of lost trust:** The erosion of workplace trust has tangible financial consequences, as replacement and hiring costs mount when disillusioned staff choose to leave.

---

[20:10](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=1210s) Actionable Steps for Centering Humans in AI Design

To build AI systems that truly work for everyone, organizations must implement practical, human-centered interventions from day one. Leaders and developers need to deliberately pause and ask two foundational questions: who is the system designed for, and who is it *not* designed for? Accessibility reviews must be embedded into the initial design phase rather than brought in reactively as damage control after an issue or harm has already occurred. Furthermore, testing must be conducted with genuinely diverse cohorts spanning different ages, generations, geographic regions, and countries.

Ultimately, achieving equitable and effective outcomes requires building cross-functional teams that keep the human experience at the forefront. Technology is fundamentally not designed for tech's sake—it is designed for human beings, and each individual human ultimately chooses which tools they will embrace to improve their lives. Shifting the focus to this human element fundamentally transforms project outcomes from start to finish.

- **Examine user inclusion and exclusion:** Explicitly identify who the AI tool serves and who might inadvertently be excluded or disadvantaged by its design. - **Shift accessibility left:** Incorporate accessibility and civil-rights checks at the earliest stages of development rather than as a post-launch remediation effort. - **Diversify user testing cohorts:** Test systems with user groups that reflect diverse ages, generations, physical locations, and national backgrounds. - **Assemble cross-functional teams:** Combine technical, DEI, civil-rights, and domain-specific perspectives to guide development. - **Acknowledge human agency:** "Technology, like I said earlier, is not designed for tech, it's designed for humans. And each human decides which tech that they're going to use to improve their lives."

## Notable quotes > [00:04](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=4s) there's no one set of rules or explanations to human problems. — Cinnamon Clark

> [07:36](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=456s) I do not experience the organization as an executive, as somebody that's on the entry level of the organization. Because if I ask for my computer to be fixed, people are jumping out of the woodwork to make that happen. My assistant asks for that, and she may not get an answer for weeks. — Cinnamon Clark

> [07:36](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=456s) Why is there a period? I thought you were mad at me.

> [11:05](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=665s) this girl is not running 5 miles a day. — Cinnamon Clark

> [15:31](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=931s) Men, you ever been invited to a baby shower? What's the first question you ask? Are there other guys there? — Cinnamon Clark

> [20:10](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=1210s) Technology, like I said earlier, is not designed for tech, it's designed for humans. And each human decides which tech that they're going to use to improve their lives. — Cinnamon Clark

## Mentioned Ai4, California State Government, Cinnamon Clark, Detroit Police Department, Dr. Brown, Emergency Medical Services (EMS), iPhone, Robert Williams

## Video description Cinnamon Clark - Chief Diversity, Equity, Inclusion and Belonging Officer, Government Agency

What happens when an AI system makes a decision about you that you don’t understand? If AI is shaping our workplaces, how do we make sure it actually works for the people in them? When technology makes choices at scale, who gets left out—and how do we fix it? AI is powerful, but it isn’t human, so how do we design it for humans? If AI influences opportunity, access, and fairness, what does it take to make those systems trustworthy?

I will explain why human-centered design matters now more than ever, especially as organizations rely on automated tools to make decisions that affect real people. The talk will include clear, practical ways to bring Diversity, Equity, Inclusion, and Belonging, accessibility, and civil-rights alignment into each stage of the AI lifecycle so systems are understandable, fair, and supportive of everyone who uses them. Attendees will learn straightforward approaches for reducing bias, improving communication, and building AI tools that people can trust. The talk will encourage leaders to design AI with people at the center so technology strengthens—not harms—equity and belonging.

[Source](https://www.youtube.com/watch?v=EFRMrRpYP9Q)

中文简报

以人为本的负责任AI设计 Cinnamon Clark 认为，人工智能系统的设计必须有意识地围绕人类差异与切身体验展开，而不能仅凭算法输出来进行评估。她提出了一个核心设计框架，促使团队在构建自动化工具之前，审视预期的用户影响、无障碍阻碍以及代际沟通差异。通过强调算法紧急调度问题以及因有缺陷的人脸识别技术导致 Robert Williams 被错误逮捕等高风险危害，Clark 揭示了忽视公民权利和用户多样性所带来的沉重代价。最终，她呼吁开展前瞻性的跨职能协作、透明沟通以及早期无障碍测试，以增进组织信任并交付可靠的技术。以人为本的 AI 设计核心关注受自动化系统影响的真实终端用户，跨学科团队应在早期评估产品受众、预期收益与潜在危害，而非在算法失控或遭遇公关危机后才被动介入。 (2:14) “输入决定输出”（Data in is data out），由于 AI 已广泛应用于招聘筛选、绩效评估以及紧急医疗调度（EMS）等高风险决策场景，非代表性或缺失的数据将直接导致严重的现实盲区与偏见。 (4:33) 系统设计必须主动兼容屏幕阅读器、认知多样性及代际沟通差异，充分考量人类生理与环境的多样性能形成良性循环，增强业务稳健性、容错率与用户信任度。 (11:05) 底特律警方因人脸识别错误抓捕 Robert Williams 的案例表明，在明知算法对黑人面部有约 96% 错误率的情况下仍强行部署，不仅践踏民权，还会造成灾难性的现实伤害与法律责任。 (18:02)

Reading summary

Cinnamon Clark argues that artificial intelligence initiatives often falter because technical teams prioritize datasets and algorithms over end-user realities, urging organizations to integrate civil rights, lived experience, and cross-functional oversight from inception rather than as reactive damage control. Proactive Interrogation Versus Reactive Damage Control (2:14) Human resources, civil rights, and DEIB professionals are historically summoned only after algorithmic bias triggers a legal grievance, public relations crisis, or loss of organizational trust. To avert downstream harm, product teams must evaluate five fundamental questions during the design phase: who will use the tool, how it will impact them, how it improves their lives, what unintended harms it may introduce, and whether cross-functional teams remain engaged throughout the lifecycle. Rather than treating civil rights compliance and accessibility as post-launch fixes, Clark advises teams to shift evaluation left into the earliest design phases. Deliberately identifying who a product is built for—and who it inadvertently excludes—protects organizations from operational fallout while honoring human agency, as users ultimately decide which technologies they will adopt. High Stakes and Algorithmic Bias in Real-World Systems (4:33) Automated decision-making systems increasingly govern life-altering public and private functions, including job candidate screening, employee promotions, worker performance scoring, and municipal emergency dispatching for fire and medical rescue services. Because algorithmic outputs directly reflect input data, non-representative datasets generate dangerous blind spots that dictate access to livelihoods and public safety resources. To illustrate the real-world harm of deploying flawed predictive tools, Clark highlights the wrongful arrest of Robert Williams in Detroit, who was misidentified in front of his family by a facial recognition system. The Detroit Police Department deployed the software despite internal knowledge of an estimated 96 percent error rate when analyzing Black faces, demonstrating the catastrophic legal and ethical consequences of bypassing civil rights standards prior to launch. Accounting for Lived Experience and Systemic Variation (7:36) Designing reliable AI requires addressing cognitive, generational, physical, and infrastructural user differences. Digital tools that lack screen reader compatibility exclude low-vision users, while reliance on regional idioms and pop culture alienates non-native English speakers. Generational communication norms also introduce friction; in her doctoral research, Clark observed that ending a text confirmation with a period signaled efficiency to her own cohort, whereas younger colleagues interpreted the punctuation as passive aggression. Organizational hierarchy and external infrastructure further distort technological utility. While executive IT requests receive immediate support, entry-level staff often wait weeks, and automated monitoring systems can penalize employees delayed by municipal public transit gaps. Clark notes that availability does not equal accessibility: an internal development program may officially exist, but requiring approval from biased supervisors renders it practically inaccessible to targeted employees. Operational DEIB and Mitigating Organizational Distrust (15:31) Clark reframes diversity, equity, inclusion, and belonging (DEIB) from abstract concepts into operational necessities that expand audience reach and system robustness. Embedding equity across the AI lifecycle operates as a non-zero-sum advantage, where designing tools that support underrepresented or vulnerable populations elevates broader usability, organizational productivity, and gross domestic product. Conversely, organizational silence around AI adoption triggers a costly internal cascade. When leadership fails to communicate how automated systems operate, what they intend to achieve, and how safeguards are structured, workplace communication deteriorates. This breakdown erodes institutional trust, leading to employee disengagement, elevated attrition, and compounding recruitment and onboarding expenses. Clark concludes that automated tools must be engineered to serve people rather than technology for its own sake, while emphasizing that technical formulas alone cannot resolve nuanced human problems. Ensuring equitable and reliable outcomes depends on organizations consistently testing systems across diverse cohorts and maintaining transparent, multidisciplinary collaboration from discovery through deployment.

精炼阅读摘要

加利福尼亚州政府机构DEIB官员Cinnamon Clark指出，人工智能系统的构建不能仅聚焦算法与数据，必须从最初设计阶段践行以人为本，否则将直接导致算法偏见、民权侵害以及严重的组织信任危机。早期跨职能审查与全生命周期五问 (2:14) Clark认为，技术团队以往往往在算法出错、员工提出正式投诉或爆发公关危机后，才被动引入人力资源、民权和多元平等（DEIB）专家进行补救。为避免灾难性后果，跨职能协作必须从项目构思阶段即刻介入，而不是作为事后补救措施。在该框架下，产品团队在构建或部署自动化工具前必须系统性评估五项核心问题：该产品服务于谁、对这些用户产生何种影响、如何改善其生活、可能带来哪些非预期的挑战或危害，以及跨职能团队是否参与了从立项到部署的全生命周期。及早整合多学科视角，有助于团队在产品推向公众前主动识别并消除潜在危害。高风险应用场景与算法伤害现实后果 (4:33) 自动化决策正在快速渗透至关乎个体生计与安全的高风险领域，包括求职简历筛选、入职管理、员工绩效评分、晋升决定，乃至基于社区数据的市政紧急医疗服务（EMS）和消防救援调度。Clark强调“输入什么数据就决定输出什么结果”，缺乏代表性的训练数据或开发团队缺乏多元视角，会直接将现实偏见固化为算法结论。底特律警方因人脸识别错误而抓捕黑人居民Robert Williams的案例即为明证。Clark指出，该执法部门在明知人脸识别系统针对黑人面孔存在约96%预估错误率的情况下依然选择部署，导致Williams在其妻女面前被当场误捕。这一案例表明，未经过严格民权标准验证的算法不仅会给机构带来严峻的法律责任，更会对无辜民众造成毁灭性的人身伤害。用户差异代际隔阂与组织层级盲区 (7:36) 以人为本的设计要求技术开发者充分理解现实用户的多元背景，包括身体能力、语言偏好和文化背景。例如，未适配屏幕阅读器的平台会直接将低视力群体排斥在公共信息之外；依赖特定文化俚语或地方流行语的系统也会给非英语母语者和原住民群体带来使用障碍。此外，神经多样性与基础设施差异（例如因公共交通匮乏导致偏远办公室员工迟到）亦应纳入系统考量，而非直接由算法机械判定为违规扣分。认知盲区还普遍存在于代际沟通与组织层级之中。Clark分享其博士论文研究指出，在即时通讯中回复带有句号的确认用语，在年长世代看来代表高效利落，但年轻一代往往会将其解读为冷暴力或被动攻击。同时，高管在组织内享有设备即刻报修的特权，而基层员工可能数周得不到反馈；若开发者仅以高层视角构建系统，将完全忽视基层人员面临的日常摩擦。制度化包容与规避沟通失效的循环 (11:05) Clark将DEIB重塑为提升系统可用性与覆盖面的运营必需品，并提出涵盖业务稳健性、容错能力、可用性、可达性与可用度、以及可靠性与信任的良性循环。以“可达性”与“可用度”的区别为例，一项专业发展培训虽然名义上存在，但若要求两名具有偏见的管理层审批签字，则对特定员工而言仍不可达。包容性设计并非零和博弈，为弱势群体提供定制化资源同样能提升整体生产力甚至GDP。相反，若组织在引入AI工具时缺乏公开透明的沟通，极易引发信息阻断与员工疑虑，进而侵蚀组织内部信任。信任崩塌将直接导致员工离职率上升，大幅推高再招聘与培训的组织成本。因此，技术团队必须在设计阶段就明确系统“为谁设计以及不为谁设计”，并在不同年龄、代际与地域的多样化群体中开展充分测试。 Clark总结强调，技术并非为了技术自身而存在，工具的最终选择权与采纳权始终掌握在真实人类手中。本简报内容仅代表其个人专业经验，并非政府官方政策；虽然其倡导在早期纳入民权与包容性审查，但在高错误率算法依然可能因行政或商业冲动被强行落地的现实环境下，如何建立具备强效法律约束力的问责机制仍是亟待解决的课题。

Detailed notes — zh

## Background 本内容由频道 Ai4 于 2026 年 8 月 17 日发布在 YouTube 平台上。视频记录了加利福尼亚州某政府机构首席多元、平等、包容与归属感官（Chief DEIB Officer）辛纳蒙·克拉克（Cinnamon Clark）的主题演讲，页面简介指出，她探讨了当 AI 深度介入职场与决策时，如何将以人为本的设计、民权合规与 DEIB 原则融入 AI 全生命周期，以构建公平、透明且值得信赖的负责任 AI 系统。 ## TL;DR 克拉克在演讲中深刻阐述了“技术是为人而设计，而非为技术本身设计”的核心理念，呼吁研发团队在产品立项初期就要追问受众画像、现实影响以及潜在的系统性障碍。她通过代际沟通符号误解、高管与基层的 IT 支持反差、苹果系统补丁等生动案例，辨析了“可用”与“真正可及”的本质区别，并对 DEIB（多元、平等、包容、归属感）进行了去标签化的实践定义。此外，她以底特律警方明知人脸识别错误率高达 96% 仍导致无辜黑人被误捕的惨剧为戒，告诫技术团队必须前置民权法案与法律合规审查，并通过早期透明沟通与跨职能多元化测试，避免组织信任崩塌与社会伤害。 ## Atmosphere & style 整场演讲充满务实、清醒且极具感染力的现场互动氛围，克拉克以其作为战略家和前 HR 的敏锐洞察，将严肃的制度与技术话题讲得生动透彻。她善于用自嘲高脚椅不稳、调侃代际标点符号误解和婴儿洗礼派对等生活化幽默活跃气氛，拉近与观众的距离；而在谈到底特律人脸识别误捕惨剧及算法决定生死时，语气瞬间转为凝重而严厉，形成了从轻松共鸣到发人深省的情感弧线，给人一种兼具制度严谨性与深厚人文关怀的震撼体验。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=4s) AI 开发者常过度聚焦底层数据集与算法产出，却容易忽略终端用户及其真实生活影响；技术终究因人而生，需将人的本质需求置于核心。 - [02:31](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=151s) 产品设计初期必须反复追问核心受众是谁、会产生何种具体影响、能否改善生活，以及是否会在无意中引入系统性障碍甚至造成伤害。 - [02:31](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=151s) 在立项之初就引入跨职能与多元视角的协同团队，能从源头规避危机，避免产品上线后陷入公关危机、品牌声誉崩塌或信任破产。 - [04:33](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=273s) “输入什么数据，产出什么数据”，若开发流程缺少对不同群体、能力差异和地域特性的考量，自动化系统将直接造成群体性排斥与偏见。 - [04:33](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=273s) 当 AI 深度介入招聘晋升乃至紧急医疗救助等关乎普通人生计、安全与生死的场景时，研发者必须承担起以人为本的强制性设计责任。 - [07:36](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=456s) 代际沟通差异、文化俚语及组织层级权限都会导致巨大体验鸿沟；决策者若仅凭自身舒适的高管体验设计工具，无法察觉基层的真实摩擦。 - [11:06](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=666s) 脱离真实使用场景的设计必然导致可用性（Usability）崩塌，上线后频繁发布系统补丁正是前期忽视用户体验的典型缩影。 - [11:06](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=666s) 必须深刻辨析“理论可用（Available）”与“真正可及（Accessible）”，系统设计必须主动识别并拆除审批门槛与人际偏见等隐形壁垒。 - [14:45](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=885s) 包容性设计绝非零和博弈，纳入多元视角不仅不会损害优势群体，反而能提高数据代表性、减少算法差错并提升整体生产力。 - [14:45](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=885s) DEIB 深刻关乎系统可用性与广泛痛点解决：多元为客观差异，平等是精准赋能，包容是不被排斥的本能，归属感是被接纳的情感状态。 - [18:02](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=1082s) 底特律警方明知人脸识别算法对黑人存在高达 96% 错误率仍强行部署并造成误捕，证明脱离民权审视的技术会带来灾难性恶果。 - [19:24](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=1164s) 技术落地缺乏透明沟通会破坏组织信任并引发离职潮与高额招聘替代成本，必须建立高频沟通机制与多方参与协作。 - [20:10](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=1210s) 以人为本的落地实践要求将无障碍审查前置至研发初期、覆盖多元代际与地域的用户测试，并通过跨职能团队保障系统公正性。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=4s) 开启对话与三大免责声明：技术因人而生演讲者辛纳蒙·克拉克（Cinnamon Clark）开场介绍了自己的身份——加利福尼亚州某政府机构的首席多元、平等、包容与归属感官（Chief DEIB Officer）。她谈到，作为一名战略规划专家，她的核心目标从来不是“强行改变人们的想法”，而是促进公众教育与认知普及。在当下的技术讨论中，AI 开发者们往往全神贯注于底层数据集与预设的技术产出，却常常把最关键的一环——终端用户及其真实生活影响——抛在脑后。面对这个由不同背景专业人士组成的会场，克拉克强调，仅仅是将这些话题摆上台面展开对话，就已经迈出了关键的第一步。进入正题前，克拉克以政府雇员和战略家的严谨（并带着几分幽默）给出了三条免责声明：第一，本次演讲的所有观点仅代表个人基于教育与从业经历的见解，不代表其雇主或加州政府的官方立场；第二，世上没有解决所有复杂人性问题的标准使用手册，本次分享旨在启发思考而非提供现成的教条指南；第三，也是最根本的一点，“人就是人”（Humans are humans）。克拉克打趣地询问全场听众早晨起床时是否都确认过自己是人类，并指出无论身处科技圈、医疗系统还是公共服务部门，人的本质需求从未改变。克拉克还调侃了演讲台上的高脚椅很不稳当，笑着对观众打招呼：“如果我待会儿摔下去了，大家尽管笑，因为等我尴尬完了，我自己也会笑出来的。” ### [02:31](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=151s) 产品设计初期应追问的核心问题结合自己在加利福尼亚州参与公共政策、法规制定及公共服务语言规范的实践，克拉克提出，研发团队在立项之初就必须反复推敲几个关键问题，而不是把它们当成事后补救的注脚： 1. **谁会使用这款产品或工具？** 在算法、数据和预设方案之外，真正的受众究竟是谁？ 2. **这项技术将对目标人群的生活产生怎样的具体影响？** 3. **它能否真正切实地改善这些人的生存状态与工作效率？** 4. **它是否会在无意中引入意想不到的系统性障碍，甚至造成现实伤害？** 克拉克提到，自己曾从事人力资源管理工作，而企业内部的 HR 往往扮演着“消防员”的角色——只有在出现矛盾、收到正式投诉或局势失控时才被紧急召集介入。若能将跨职能的协同团队在构想初期就引入研发流程，让代表多元视角的专业人员自始至终坐在一张桌子旁，就能从源头规避危机。这不仅关乎合规，更能直接避免产品上线后陷入公关惨剧、信任破产或品牌声誉崩盘的泥潭。 ```mermaid flowchart TD accTitle: 人类多样性对AI系统价值的循环反馈 subgraph Design["早期设计与多元考量"] A["包容性设计输入<br/>能力/语言/认知/生活经验"] --> B["代表性数据集<br/>减少偏见与数据盲区"] end subgraph Core["系统鲁棒性与体验"] B --> C["高容错率与高可用性"] C --> D["企业与技术稳健度提升"] end subgraph Impact["信任与落地循环"] D --> E["实现真无障碍与平权准入"] E --> F["构筑公众信任与合规护城河"] F --> A end ``` ### [04:33](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=273s) 为什么以人为本的设计在AI时代至关重要克拉克向台下抛出提问：“以人为本的设计为何如此重要？”面对观众短暂的沉默，她笑问大家是不是一个“不爱互动的害羞听众群”，随后一针见血地指出核心逻辑：“技术不是为技术本身设计的，技术是为人设计的。”现实中，决定哪款技术工具能够留存并推广的，永远是活生生的人类，他们只会选择那些能让自身生活更轻松、能真正解决痛点的工具。当前，自动化工具和人工智能已经渗透进组织的微观肌理，从招聘简历初筛、员工入职、晋升决策、绩效评估，到社会资源分配的极端场景——例如紧急医疗系统（EMS）和消防力量依据算法与社区维度决定先去救助谁。克拉克严肃提醒，当人工智能在幕后决定一个普通人的生计、安全乃至生死时，开发者肩负着不可推卸的责任： - 必须认识到“输入什么数据，产出什么数据”（Data in is data out）。如果开发流程缺少对不同群体、能力差异和地域特性的考量，系统将直接抛弃一整片人群。 - 自动化算法的影响不仅存在于社会新闻中，每个人每日打开社交媒体、使用行业大会的议程推荐软件时，都在受其影响。 - 越是广泛应用自动化决策，越需要强制性地将“人的真实体验”置于系统研发的绝对核心。 ### [07:36](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=456s) 真实体验与代际沟通：何为真正的以人为本为了将“以人为本”拆解为可操作的认知，克拉克列举了多种常被技术专家忽视的人群特性。以无障碍为例，开发网页或移动端应用时，若未考虑低视力人群的屏幕朗读器（Screen Reader）适配，该工具对于这部分群体而言就完全等于“不可用”。语言层面，英语非母语人群（ESL）和跨国移民往往无法理解系统界面中充斥的俚语、行业黑话或特定流行文化梗。克拉克分享了她正在撰写的有关职场代际沟通博士论文中的一个鲜活案例：她属于介于 X 世代与千禧一代之间的“Xennial”群体。在工作中回复即时消息时，她习惯敲下一个带有句号的单词：“Sure.”（好的。），在她看来，这代表“没问题，我正处理，这就办好，搞定”。然而，这让年轻一代的下属极为恐慌与焦虑，他们会战战兢兢地问：“这个句号是什么意思？你是不是在生我的气？”克拉克在台上大声反问听众：“难道我说谎了吗？这就是现实差异！”这表明，即便是最微小的符号习惯，在不同生活经历与文化语境的人眼中，也会引发截然相反的认知。在组织权限和地位差异上，克拉克举了一个内部技术支持的反差案例： - 当她作为管理层高管打电话要求维修电脑时，IT 团队会立刻蜂拥而至争相解决； - 但如果同样的问题由她的行政助理提出，报修工单可能会被搁置数周无人问津。 - 高管如果仅凭自身舒适的日常体验去设计工作流或内部 AI 工具，根本无法察觉基层员工每天遭遇的制度摩擦。 ### [11:06](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=666s) 人类多样性循环：从认知神经到“可用”与“可及”之辨讨论延伸到人类多样性的多个维度：身体机能差异（克拉克幽默自嘲：“如果一份工作要求我每天跑五英里，那我肯定跑不下来，我得先告诉大家这一点”）；认知与神经多样性（Neurodiversity），随着社会去污名化的推进，越来越多的神经发散型员工在职场显现；以及复杂的地域与生存环境。克拉克提到，很多管理者会严厉斥责迟到的基层员工，却完全无视公共交通基础设施的匮乏——这些员工需要辗转倒乘公交车与地铁才能到达偏远的办公区，而决策者可能每天只是舒服地开着私家车上下班。在技术产品化方面，克拉克以苹果公司的系统更新做类比：“在座有人用 iPhone 吗？这简直人尽皆知。你每次更新完 iOS 系统，脑子里都会冒出疑问：‘这谁想出来的脑残设计？这怎么用？根本用不了！’然后紧接着苹果就得连发两次系统补丁，来修复最初那个补丁带来的灾难。”这就是忽视真实使用场景导致系统可用性（Usability）崩塌的典型缩影。克拉克随后深刻辨析了“可用（Available）”与“真正可及（Accessible）”之间的鸿沟： - 一项服务或培训项目存在于公司内网、理论上面向全员开放，这仅仅是“Available”； - 但如果员工想要报名，必须层层报批直属领导及上层主管，而该员工恰好与主管存在人际矛盾，或是主管本身存在主观偏见，那么这项资源对该员工而言就根本不是“Accessible”。 - 在技术系统设计之初，开发者必须系统性地识别并拆除这类隐形门槛，否则产品的覆盖面将大打折扣。 ### [14:45](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=885s) DEIB不是零和博弈：定义与组织赋能克拉克指出，多元、平等、包容与归属感（DEIB）在产品构建中绝非仅仅关乎“种族标签”，其底层逻辑在于系统可用性、无障碍覆盖、跨地域适用性，以及打造一个能够广泛解决大众痛点的卓越工具。她引用了现场参会的布朗博士（Dr. Brown）的观点，强调包容性设计绝非“零和博弈”（Not a zero-sum game）。在 AI 研发中纳入全方位的视角，不仅不会损害优势群体的利益，反而会整体拉高数据代表性、降低算法差错、提升产品性能，进而推动企业生产力与宏观经济增长。为了厘清容易混淆的概念，克拉克给出了通俗的现场定义： - **多元（Diversity）**：人与人之间存在的客观差异，一部分是与生俱来的（如种族、神经特质），一部分是后天形成的（如教育背景、地域经历）。 - **平等/公平（Equity）**：确保每个人都能获得其生存和繁荣所需的精准资源与工具。在职场中，这表现为提供符合人体工程学的办公设备、灵活的工作制度以及匹配个人需要的医疗保障。 - **包容（Inclusion）**：克拉克向台下男士提问：“男同胞们，你们被邀请参加婴儿洗礼派对（Baby Shower）时，问出的第一个问题是什么？是不是‘现场还有别的男的去吗？’”全场爆发出笑声。克拉克总结道，这就是人类希望被包容、不想被当成异类排斥在外的本能需求。 - **归属感（Belonging）**：每个人发自内心渴望被组织、群体接纳与认可的情感状态。 ### [18:02](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=1082s) 民权与合规底线：底特律人脸识别误捕惨剧的教训在演讲第一部分的尾声，克拉克严正告诫在座的技术专家，在 AI 产品的设计全周期中，必须及早把民权法案、隐私法规与法律合规红线纳入核心考量，绝不能等到产品商业化甚至酿成灾难后再找法务救火。她举出了发生在密歇根州底特律的惨痛真实案例：非裔男子罗伯特·威廉姆斯（Robert Williams）在自家草坪上，当着妻子和年幼女儿的面被警察直接上手铐逮捕。事件的起因是警用人脸识别系统在扫描模糊监控录像后错误地锁定了他的面孔。更具讽刺意味和令人震惊的技术事实是，底特律警方在部署并使用该系统时，早就明确知晓这套商业人脸识别算法在识别黑人群体时存在高达 96% 的惊人错误率（Error Rate）。克拉克以此案例痛陈：带着已知高达 96% 的致命缺陷强行上线技术产品，最终直接摧毁了一个无辜公民的正常生活与全家人的尊严。这正是技术脱离了人性审视、漠视民权红线所造成的恶果。她重申，如果在系统开发早期没有建立多元质疑机制，技术带来的将不是智能进步，而是成规模的灾难与偏见输出。 ### [19:24](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=1164s) 沟通断裂与组织信任的连锁反应在推进技术变革与系统落地时，如果未能抓住机会就所做的工作、带来的实际影响以及前述核心议题进行频繁且持续的沟通，组织就会陷入严重的沟通断裂。这种沟通上的缺失会直接摧毁员工对团队与管理层的信任，而一旦“信任崩塌，人们就会想要离开组织”。员工流失随即会引发高昂的重新招聘与培训成本，陷入持续消耗组织资源的恶性循环。因此，管理者与技术团队必须在早期就主动建立透明沟通机制，吸纳员工参与，从源头推进跨部门的深度协作。 ```mermaid flowchart TD accTitle: 缺乏沟通导致的组织信任危机链条 subgraph S1["沟通缺失"] A["缺乏频繁与透明的沟通"] end subgraph S2["信任与流失"] B["职场沟通断裂"] C["组织信任崩塌"] D["员工离职意愿上升"] end subgraph S3["组织代价"] E["招聘与人员替代成本激增"] end A --> B B --> C C --> D D --> E ``` - **早期透明沟通**：就技术变更的目的、影响范围和核心问题保持高频同步。 - **全员参与协作**：在方案初期引入多方利益相关者，避免单向灌输引发抵触。 - **避免隐性成本**：通过维护组织信任来降低员工离职率及后续的高额招聘开支。 ### [20:10](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=1210s) 即刻行动：构建以人为中心的 AI 落地实践面对 AI 与自动化工具的应用，Clark 给出了当下即可付诸实践的具体行动准则。核心第一步是明确拷问系统的定位——弄清“该系统是为谁设计的，又将谁排除在外”。同时，必须将无障碍（Accessibility）审查前置到研发初期，杜绝“出现问题后再打补丁”的被动做法。在测试与评估阶段，必须引入涵盖不同年龄、所处地域、国家及代际背景的多元化用户群体，并组建跨职能团队协同攻关。 Clark 总结强调，技术本身绝非为了技术而设计，它的终极目的始终是服务于人类，最终也是由每一个真实的人去决定选择何种技术来改善自己的生活。只要在整个生命周期中始终关注人性化要素，“从起点到终点的所有结果都将迎来彻底的转变”。 - **前置无障碍评估**：在系统设计与开发早期完成无障碍合规与可用性审查。 - **多元化用户测试**：覆盖不同年龄、地区、国家和代际的真实用户画像，全面排查偏见与体验盲区。 - **组建跨职能团队**：打破技术与业务壁垒，融合多元视角确保系统的公正性与可用性。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=4s) 人就是人 — 辛纳蒙·克拉克 > [00:04](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=4s) 如果我待会儿摔下去了，大家尽管笑，因为等我尴尬完了，我自己也会笑出来的。 — 辛纳蒙·克拉克 > [04:33](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=273s) 技术不是为技术本身设计的，技术是为人设计的。 — 辛纳蒙·克拉克 > [04:33](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=273s) 输入什么数据，产出什么数据 — 辛纳蒙·克拉克 > [07:36](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=456s) Sure. — 辛纳蒙·克拉克 > [07:36](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=456s) 这个句号是什么意思？你是不是在生我的气？ — 年轻一代员工 > [11:06](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=666s) 如果一份工作要求我每天跑五英里，那我肯定跑不下来，我得先告诉大家这一点 — 辛纳蒙·克拉克 > [11:06](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=666s) 这谁想出来的脑残设计？这怎么用？根本用不了！ — 辛纳蒙·克拉克 > [14:45](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=885s) 男同胞们，你们被邀请参加婴儿洗礼派对时，问出的第一个问题是什么？是不是‘现场还有别的男的去吗？’ — 辛纳蒙·克拉克 > [19:24](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=1164s) 信任崩塌，人们就会想要离开组织 — 辛纳蒙·克拉克 > [20:10](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=1210s) 该系统是为谁设计的，又将谁排除在外 — 辛纳蒙·克拉克 > [20:10](https://www.youtube.com/watch?v=EFRMrRpYP9Q&t=1210s) 从起点到终点的所有结果都将迎来彻底的转变 — 辛纳蒙·克拉克 ## Mentioned 辛纳蒙·克拉克 (Cinnamon Clark), 加利福尼亚州政府 (State of California), 苹果公司 (Apple), iPhone, iOS, 布朗博士 (Dr. Brown), 罗伯特·威廉姆斯 (Robert Williams), 底特律警察局 (Detroit Police Department), Ai4 ## Video description Cinnamon Clark - Chief Diversity, Equity, Inclusion and Belonging Officer, Government Agency What happens when an AI system makes a decision about you that you don’t understand? If AI is shaping our workplaces, how do we make sure it actually works for the people in them? When technology makes choices at scale, who gets left out—and how do we fix it? AI is powerful, but it isn’t human, so how do we design it for humans? If AI influences opportunity, access, and fairness, what does it take to make those systems trustworthy? I will explain why human-centered design matters now more than ever, especially as organizations rely on automated tools to make decisions that affect real people. The talk will include clear, practical ways to bring Diversity, Equity, Inclusion, and Belonging, accessibility, and civil-rights alignment into each stage of the AI lifecycle so systems are understandable, fair, and supportive of everyone who uses them. Attendees will learn straightforward approaches for reducing bias, improving communication, and building AI tools that people can trust. The talk will encourage leaders to design AI with people at the center so technology strengthens—not harms—equity and belonging. [Source](https://www.youtube.com/watch?v=EFRMrRpYP9Q)

Source quotes

[0:04] 辛纳蒙·克拉克: 人就是人 [0:04] 辛纳蒙·克拉克: 如果我待会儿摔下去了，大家尽管笑，因为等我尴尬完了，我自己也会笑出来的。 [4:33] 辛纳蒙·克拉克: 技术不是为技术本身设计的，技术是为人设计的。 [4:33] 辛纳蒙·克拉克: 输入什么数据，产出什么数据 [7:36] 辛纳蒙·克拉克: Sure. [7:36] 年轻一代员工: 这个句号是什么意思？你是不是在生我的气？ [11:06] 辛纳蒙·克拉克: 如果一份工作要求我每天跑五英里，那我肯定跑不下来，我得先告诉大家这一点 [11:06] 辛纳蒙·克拉克: 这谁想出来的脑残设计？这怎么用？根本用不了！ [14:45] 辛纳蒙·克拉克: 男同胞们，你们被邀请参加婴儿洗礼派对时，问出的第一个问题是什么？是不是‘现场还有别的男的去吗？’ [19:24] 辛纳蒙·克拉克: 信任崩塌，人们就会想要离开组织 [20:10] 辛纳蒙·克拉克: 该系统是为谁设计的，又将谁排除在外 [20:10] 辛纳蒙·克拉克: 从起点到终点的所有结果都将迎来彻底的转变

Source quotes

[0:04] Cinnamon Clark: there's no one set of rules or explanations to human problems. [7:36] Cinnamon Clark: I do not experience the organization as an executive, as somebody that's on the entry level of the organization. Because if I ask for my computer to be fixed, people are jumping out of the woodwork to make that happen. My assistant asks for that, and she may not get an answer for weeks. [7:36] Why is there a period? I thought you were mad at me. [11:05] Cinnamon Clark: this girl is not running 5 miles a day. [15:31] Cinnamon Clark: Men, you ever been invited to a baby shower? What's the first question you ask? Are there other guys there? [20:10] Cinnamon Clark: Technology, like I said earlier, is not designed for tech, it's designed for humans. And each human decides which tech that they're going to use to improve their lives.

Original transcript

[0:04] My name is Cinnamon Clark. I'm the Chief Diversity, Equity, Inclusion, and Belonging [0:07] Officer for the state of well, in the state of California for like they said a government agency. [0:14] And I'm delighted to be here. Much like Janine said, I'm just excited that these things are on [0:21] the agenda to talk about and have conversations because when we have the conversations, [0:27] even in a group of folks that have dissimilar backgrounds, it starts the conversation. [0:33] And as a strategist, my goal is always education and awareness, not changing people's minds. So, [0:39] I'm glad you're here and I'm very thankful that you're here. And without further ado, [0:45] human-centered design, designing systems that work reliably reliably for real people. I think often [0:52] times when we start talking about um AI, we're thinking about data sets and potentially outcomes. [0:59] And sometimes we don't really take a look at the end user and how that impacts them. So, [1:06] I'll talk about a little bit of that today. Also, I wouldn't be a strategist [1:10] or a government employee without giving you a few disclaimers. The first disclaimer is that [1:14] anything I say here is my opinion based on my experience and education and all of those things, [1:21] not that of my employer. The second thing is that this is really just to start the conversation. [1:28] There's not going to be a step-by-step, you know, guide to follow. There's no one set of rules or [1:34] explanations to human problems. And the third is really that humans are humans. Each one of [1:41] you woke up a human I'm assuming this morning, right? We're good? Okay. So, humans are humans. [1:47] And often times we're like, oh, we're in tech. We're in health care. Humans are humans regardless [1:52] of where they work. And that's the goal here is to provide you with just a little bit of [1:57] foundational information about human-centered design and responsibly designing um AI. So, [2:04] I'll save you all the biographical details. That's in all the flyers. You can catch me on LinkedIn or [2:09] wherever you see me. The agenda here you see is um you know, we're going to talk about five questions [2:14] that you should really consider in the designing of your product and thinking about your end user. [2:20] And I want to open this talk up. These chairs are really messing with me. So, if I fall, [2:25] you can laugh because I'm going to laugh after I get done being embarrassed, right? Thank you. [2:31] So, the first question that I want you to ask or that I encourage folks to ask, especially in [2:38] my line of work. I work in policy, legislation, and all of the language around serving humans in [2:44] the people in the state of California. And who is going to use this product or tool? [2:51] Who is it? Right? We have our data sets, we have our ideas, we have the solution we're trying to [2:56] solve, but who's going to use it? The second question is how will whatever we are developing [3:04] impact the people we want it to get to? So, how are our products impacting those folks? [3:13] The third question is, excuse me, how the technology or the product, [3:21] there's lots of language going on this week, um how that product will improve the lives of [3:26] the folks that we're trying to get it to. And the last question is how will what I'm creating [3:33] either improve the lives of the folks that will be the end user and or [3:40] how it can introduce unintended challenges or introduce harm. These are questions that aren't [3:48] often spoke about in the design phase. As an HR professional, I saw a few HR folks in the room. [3:54] We're always called in at the end when there's something going wrong. When somebody files a [3:58] complaint or there's something going on. But if you're using collaborative teams and things of [4:04] that nature in the beginning, there's a good chance that you'll have the right people in [4:09] the room from the beginning to the end. So that at the end, just like the previous panel Janine said, [4:15] you know, you don't get to the end and you have to raise your hand and there's either [4:19] uh [snorts] publicity issue, a credibility issue, a trust issue, or whatever may come with that. So, [4:25] the five questions, basically the who, what, when, where, why of your product. [4:33] Why human-centered design matters. Why does it matter? Anybody want to take a guess? [4:40] Or is this one of those non-participatory uh audiences? Just let me know. Okay. [laughter] [4:47] Okay. So, it matters because we're all humans. We're all humans. Technology is not designed for [4:53] technology. Technology is designed for humans. And humans choose which which technologies make their [5:01] lives easier, that improve the things that they need it to improve, right? And so, organizations [5:07] are using automated and AI tools. I've heard countless people this week talk about how the um [5:14] their organizations are integrating and implementing and and launching these solutions. [5:20] Um and they're making choi- they're making decisions about things like [5:25] you know, promotions, onboarding, who gets hired, the way we screen our applicants, right? And [5:32] often times, um it just it depends on who's at the table, just like the previous uh panel was talking [5:37] about. We're making decisions about who, you know, how we rate performance. We're even making [5:45] decisions about who gets support and in emergency manage management systems, who gets help from EMS. [5:53] Right? Based on neighborhoods and things of that nature. So, it's really important to think about [5:59] all of these ways that AI is impacting our day-to-day lives. When AI makes decisions [6:05] about people, it's our responsibilities, it's our responsibility, excuse me, to be responsible [6:11] and to really center the human experience in that development and in that entire process. [6:18] Human-centered design is exactly how we get there. Like I said, there's nothing super technical [6:24] about what I'm going to tell you. It's just a reminder about centering the human experience. [6:29] Human design, um, excuse me, human design mat human-centered design matters more now than ever. [6:37] Um, I don't know a lot of you stayed from the previous panel, but human-centered [6:41] design ensures that we are looking at diverse populations, how they use what we're developing, [6:49] how it impacts their lives, and the data in, just like we've heard all week, is the data out, right? [6:55] And if we don't have diverse audiences, different levels of abilities, even the regional element, [7:00] um, of folks in that development, we're going to miss the opportunity to miss [7:06] to miss a whole population of people essentially. Organizations are relying on these automated tools [7:12] to make decisions about each one of us. If you use social media, you already know this. If you [7:18] use the app for this particular conference, you also know this because it built my whole agenda [7:23] based on what my interests are and what my talk was about. And when AI makes decisions [7:29] about people, again, we have the responsibility to ensure that we're doing that responsibly. [7:36] And so, what does it actually mean? What does human-centered design actually means? [7:40] I think I've talked about it a little bit already, but it's centering the human experience. And my [7:46] goal is to make this as simple as possible. You need to know who is using your product [7:52] or your tool. Period. What are their abilities? Right? Because if you're developing, I don't know, [7:58] a webpage or whatever app, if people with the screen reader that may have low vision [8:04] can't see it, can't use it, that information is deemed inaccessible to those folks, right? [8:11] Also, languages, English as a second language, or even folks that migrated from other countries [8:21] in the world, right? When we build tools or we build policies or programs that use idioms or even [8:28] even pop culture and language, that's lost on a particular part of your your your audience because [8:34] we don't speak the same language always. Um and also, knowing the lived experience, the stressors. [8:42] Lived experience is a big one. I heard um I heard a woman I had a conversation earlier today [8:48] and she was talking about we were talking about I'm writing my dissertation right now and we're [8:52] talking about um generational communication in um generational communication in the workplace. [9:00] And I was saying I'm of the exino exinnial generation and for me, when I text back, [9:07] it's usually like you ask me a question, I'm like, "Sure." Period. Right? That means I'm [9:11] on it. Just tell me what you need, right? But my younger colleagues are completely thrown by this, [9:16] right? They're just like, "What does that mean? Why is there a period?" And they think that it's [9:21] Am I lying? Am I lying? Exactly. Because I'm like, "Yeah, I'm on it. Boom." [9:27] But then I talked to them, they're like, "Well, I thought you were mad at me." And I'm like, [9:32] "No, I told you I was sure. I got it. I'm I'm on it, right? Here's what you were looking for." [9:37] And so, it's really important to understand that the folks that are using your tools [9:41] may not have the same perspective or the same experiences as you. And so, we have to remember [9:46] that early on in the design phase. We need to know their abilities. We need to know their languages, [9:51] their cultures. Again, I work for an organization in California that serves all Californians, [9:58] right? And if we don't have the language to work with the indigenous population or maybe other [10:04] immigrant populations, then we're going to miss the opportunity to serve them with responsible [10:10] solutions or Excuse me, responsible collaboration, and we're just going to miss the opportunity. [10:17] You know, lived experience is a big one. Ensuring that we have the right people at the right table. [10:23] And also, no one has the answers. No one has all the answers. So, as much as myself [10:29] and my executive team want to sit and just solve all of life's problems, guess what? [10:34] I don't experience I do not experience the organization as an executive, as somebody [10:39] that's on the entry level of the organization. Because if I ask for my computer to be fixed, [10:43] people are jumping out of the woodwork to make that happen. My assistant asked for that, and [10:48] she may not get an answer for weeks. And so, we have to keep those things in mind. I know [10:52] that's not AI specific, but we have to remember that we're dealing with humans, and having a [10:57] human-centered approach is so important. Inclusive AI, designing human variation. Excuse me. [11:06] And so, we have to remember that there's different levels of abilities, right? Bless you. Bless you. [11:13] Bless you, doctor. [laughter] Um there's different levels of abilities, right? [11:19] If I'm integrating into a workplace that requires me to run 5 mi a day, this girl is not running [11:25] 5 mi a day, okay? I'm just letting you know, right? And so, everybody has a different level [11:29] of ability, and we have to think about that in the very beginning. Same with language. The way we use [11:35] language. I could make a joke about, I don't know, I can say "returner" or something, right? Maybe a [11:41] third of the people in this room understand that, right? And so, we have to be careful in the way [11:46] in which we communicate. Cognition. We heard in the last session about neurodivergent employees, [11:53] and that's a thing, right? And we're probably, as it's destigmatized, we're probably going to [11:57] see more of that in coming years, right? And understanding that people think differently. [12:03] And I'm one of those individuals, right? Culture has a big play in it. Environment. I just talked [12:10] to you about the woman in the Midwest versus how I how I say, "Hey, I'm on it" in California, right? [12:16] Um and lived experience. I think we don't do the best job leveraging the lived experience of folks. [12:24] Think about it. I'm mad at an employee because the employee's late to work. Employee has to use the [12:30] bus, the I don't know, the subway. All of these things, but we don't have the infrastructure to [12:36] get them to this remote office location. So, we haven't considered, you know, folks that have to [12:41] rely upon like public transportation. And so, we just have to think about all of the things because [12:47] we want our tools and our products to reach the widest audience possible. And if we don't [12:55] essentially, um consider those folks, we're going to have a missed opportunity. And so, [13:01] this slide here you'll see why human variation matters in AI system. And it's a cycle, [13:07] but it continues going on and on. So, designing for human variation improves business robustness, [13:15] right? Because you have a more inclusive way of thinking, you've expanded the um [13:21] the thought process and the lived experience of folks. Error tolerance. That's a little [13:25] more technical than I'd like to get into today, but if you have a wider [13:30] um data set and diverse populations error error tolerance, excuse me. Usability. Something as [13:38] simple Have you ever used something Anybody have an iPhone here? This is notorious, right? [13:44] You ever get an update from your iPhone? You're like, "Who thought of this? What is this? [13:48] How am I going to use this? This doesn't even work." And then you have two more updates so that [13:52] you can fix the original update. Who thought of this, right? That's a prime example of usability. [13:59] Same with accessibility. Just because it's available doesn't mean that everybody can access [14:04] it. And I know folks in different communities understand exactly what I'm talking about, [14:09] but I'll make it simple. Just because there's a program out there that's available, let's say, [14:16] I don't know, a professional development program in your organization. [14:20] If you have to ask your supervisor and your supervisor's supervisor to participate and you had [14:26] a couple bad interactions with that person or that person has bias or whatever the things may be, [14:31] just because it's accessible doesn't mean it's available because it's available doesn't mean [14:36] it's accessible to you. So, you want to be thinking about making easier, [14:40] removing barriers, and thinking about those things in the very beginning. [14:45] Reliability and trust because I can't tell you how many times I've talked to my boomer mother about [14:51] AI and she thinks it's one thing. AI. Right? AI has so many things. We've been using AI for years. [14:59] But there's a lack of trust and there's a lack of understanding and it's our responsibility [15:04] in forms like this to really talk through that, but also in the communication, the deployment, [15:10] and the collaboration, and the ways that we develop, implement, and um and work through uh [15:17] AI in our organizations and just in the world. And then being inclusive in the design. [15:23] Who's going to use it? How are they going to use it? Where are they going to use it? Is it [15:26] going to be on a desktop? Is it going to be on a Is it going to be on an iPhone? Um it's really [15:31] important to think about these things. I would be remiss if I didn't talk about diversity, equity, [15:35] inclusion, and belonging um in my talk. And you know, it's not a workshop. And being inclusive in [15:44] what it is you were thinking about and developing and building and hoping that the end user finds [15:50] valuable has nothing to do with race. It has everything to do with the things that we talked [15:55] about earlier. Usability, accessibility, larger demographic area, and who doesn't want a tool that [16:02] solves everybody's problems? That's why you guys are working on them, right? So diversity improves [16:08] the data set representativeness. We've talked about that ad nauseam, so I'm not even going [16:12] to get into that. Um equity improves outcomes of fairness, because these tools are making decisions [16:18] about something so simple like who gets service from an ambulance or a fire truck. Right? Equity [16:25] is important. Belonging in your organization improves trust. And when we embed inclusive [16:32] practices and equity in the entire AI life cycle, the outcomes are [sighs] just endless, right? More [16:41] productivity, better GDP like we talked about in the last session. Just better outcomes in general. [16:47] And just like Dr. I think it was Brown, but I know she's in Dr. She's right here. She said, "It's not [16:52] a zero-sum game. If we include everybody and we think about all of the things that we can possibly [16:57] think about in development, it's going to improve everything for everybody. It's not zero-sum." [17:05] Quick definitions of diversity, equity, inclusion. Diversity really quickly are [17:08] the things that make us different. Some things we're born with, other things we're not born with. [17:14] Equity, ensuring that everybody has what they need. In the workplace, ensuring that everybody [17:18] has the tools that they need to thrive. That could be making sure that they have [17:21] flexible work arrangements. That could be even accommodations for things like ergonomics and [17:28] things of that nature. Making sure people have the right healthcare. That's inclusive. That's [17:33] inclusive and it's also equitable to ensure that everybody has what they need to survive [17:38] and to thrive in the workplace. Inclusion, who's ever Men, you ever been invited to a baby shower? [17:43] What's the first question you ask? Are there other guys there? Exactly. You want to be included, too. [17:49] You're not excluded, right? So, it's a basic human need. And belonging, who doesn't want to belong? [17:56] I don't know. You don't have to answer that, but really it's just that simple. [18:02] I'd be remiss if I didn't talk about civil rights really quickly. I only have about 2 minutes left. [18:06] Civil rights, you have to ensure that bless you. In the design of your AI product, tool, [18:12] whatever the thing may be, that you're looking at civil rights, you're looking at the law, you're [18:15] looking at compliance. I'm not here to talk about compliance. I'm not here to talk about the law. [18:19] However, I am here to say that that is something that you should be looking at and thinking about [18:23] very early on. One quick example, Detroit, Mr. Williams, he was arrested in front of his whole [18:31] family, black male, arrested in front of his whole family because facial recognition system [18:37] picked him out of a crowd. [laughter] Excuse me, it's not funny, right? But it's a perfect [18:42] example of how we didn't get it right and we knew that there were errors in the system [18:46] and we still deployed the product and somebody's somebody's life was negatively impacted. And oh, [18:52] by the way, 96% the Detroit Police Department knew that there was a 96% error rate. I mean, here with [18:59] technical folks, you know that that is problematic on a whole other level. Okay? So, if we're not [19:05] asking the questions early and often, we're going to miss the opportunity. Biased AI outcomes, [19:11] I just talked about that. I got about 30 seconds. Accessible design failures, again, just like I [19:17] was saying a couple slides back, just because it's available does not mean it's accessible. [19:24] Human-centered redesign successes. Um, sorry. I want to get to the what you can do now. Sorry. [19:31] Um, workplace communication breakdowns. When we don't use the opportunity to communicate [19:37] frequently and often about what we're doing, how it impacts, and how it looks at those [19:42] five questions we talked about earlier, we're missing the opportunity. It creates communication [19:46] breakdowns. When communication breaks down, trust breaks down. When trust breaks down, [19:50] people want to leave the organization. When people want to leave the organization, it costs more [19:54] money to hire more people, and it just goes on and on. And there's a cost associated with that. [20:00] So, we want to make sure that we are communicating early, [20:03] we are involving people, we're collaborating early, and everything of that nature. [20:10] And my time is up, but what you can do now. Ask who the system is designed for, [20:17] and who it isn't. Include accessibility reviews early, not when there's a problem. [20:22] Test your system with diverse users. Diversity in age, location, region, country. [20:29] All of the things. Generations. Build cross-functional teams. And once you look at [20:39] the human element of what it is you're trying to do, I guarantee the outcomes are going to change, [20:45] from start to finish. Humans are humans. Humans, well, technology, like I said earlier, is not [20:52] designed for tech, it's designed for humans. And each human decides which tech that they're [20:59] going to use to improve their lives. Thank you. Sorry for going over time, and have a lovely day.