Deque Links Axe Engine to AI Coding Agents to Prevent Compliance Risks

Delivering Better, More Accessible Products for Everyone in the Age of AI With Deque Systems — Ai4

Ai4

AI models reproduce existing web design flaws at rapid speed, requiring deterministic rules engines embedded directly into developer coding agents to prevent technical debt and European Union regulatory penalties.

AI-assisted summary of Delivering Better, More Accessible Products for Everyone in the Age of AI With Deque Systems — original source. Check the original for full context.

Source published: 2025-09-22

Detailed notes

## Background Published on YouTube by Ai4 on September 22, 2025, this presentation features Deque Systems CEO Preety Kumar and Chief Product/Technology Officer Dylan Barrell. The session discusses how integrating AI agents with advanced accessibility platforms empowers developers to accelerate software creation while meeting vital business, legal, and usability standards.

## TL;DR Preety Kumar and Dylan Barrell address the velocity paradox of modern software, warning that rapid AI code generation risks exacerbating web accessibility barriers for 1.3 billion people with disabilities. They outline the urgent commercial, regulatory, and technical incentives for building accessible software, noting that structured semantic markup is equally essential for autonomous AI agents. To bridge this gap, Deque presents an integration using the Model Context Protocol (MCP) to embed deterministic axe-core testing directly into developer environments like VS Code. Through a step-by-step demonstration with GitHub Copilot, they illustrate how AI agents can automatically detect, contextualize, and remediate accessibility defects in real time without falling victim to non-deterministic vibe coding.

## Atmosphere & style The presentation strikes a compelling balance between urgent digital civil rights advocacy and pragmatic software engineering. Kumar establishes an earnest, mission-driven atmosphere reinforced by frustrating screen reader audio demonstrations that clearly illustrate real-world barriers, before handing over to Barrell for a relaxed, developer-focused technical walkthrough. The session feels informative, grounded, and practical, using relatable coding anecdotes and clear live tooling demos to demystify automated accessibility remediation.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=4s) Deque Systems centers its mission on making technology accessible to people with disabilities, built on its open-source axe-core rules engine with over 3 billion downloads and a zero-false-positive standard. - [02:33](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=153s) With AI generating roughly 40% of all new code, unchecked velocity risks accumulating unprecedented technical debt that excludes 1.3 billion people with disabilities worldwide. - [02:50](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=170s) The mismatch between AI code generation operating in seconds and traditional accessibility auditing taking weeks leads developers to bypass compliance testing entirely. - [03:20](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=200s) Because 96% of the web is currently inaccessible, AI models trained on existing web repositories inherently replicate and encode accessibility defects into generated code. - [05:52](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=352s) The European Accessibility Act (EAA) makes accessibility non-negotiable by introducing penalties of up to 5% of company revenue across the EU's $20 trillion market. - [06:55](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=415s) Accessibility expands market reach to the one in four individuals with disabilities, exemplified by Microsoft expanding its addressable audience by 33% through accessible design. - [08:10](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=490s) Writing accessible, semantically structured code futureproofs software for autonomous AI agents, which rely on the exact same DOM hierarchy, semantic hooks, and ARIA labels as assistive technologies. - [09:54](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=594s) Solving AI-driven accessibility defects requires 'fighting fire with fire' by injecting domain-specific accessibility agents and rule engines directly into developer IDE workflows. - [13:02](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=782s) The Model Context Protocol (MCP) enables coding agents to proactively query verified accessibility specifications before and during code generation. - [13:30](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=810s) Because AI models are non-deterministic and prone to destructive overwrites during 'vibe coding,' deterministic automated testing is essential to validate and ship production-ready code. - [14:39](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=879s) Integrating the axe MCP server with coding assistants like GitHub Copilot allows agents to identify violations, apply targeted semantic fixes (like dynamic alt text and aria-labels), and re-validate changes in a tight loop. - [18:02](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=1082s) Combining developer intent, AI agent execution, and deterministic accessibility verification allows teams to build accessible products at the speed of modern AI.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=4s) Framing the Future of Accessible Technology and Deque's Mission

Preety Kumar, CEO and founder of Deque Systems, opens the presentation alongside Chief Product Officer Dylan Barrell to address how the ongoing revolution in artificial intelligence can be steered toward expanding digital access rather than restricting it. Grounding the session in a guiding philosophy, Kumar invokes social philosopher Charles Handy, who famously noted that the future is not inevitable, and that people have the power to shape it only if they clearly define what they want that future to become.

For Deque Systems, that defined mission has always been ensuring that information technology is fully accessible to people with disabilities. Central to this effort is `axe-core`, Deque’s open-source accessibility rules engine that has surpassed three billion downloads and is natively embedded across industry-standard environments like Google Chrome and Microsoft technologies. Through the axe Platform, designers, developers, and QA engineers can detect and remediate accessibility barriers at the speed of modern AI workflows. Kumar emphasizes that developers trust the platform primarily because of its commitment to zero false positives—ensuring that surfaced issues represent genuine defects so accessible code quality becomes the default standard for everyone rather than an afterthought.

- **`axe-core` reach:** Over 3 billion downloads; integrated directly into Google Chrome and Microsoft development stacks. - **Core value proposition:** High-speed issue detection with zero false positives to avoid developer friction.

---

[02:33](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=153s) The Velocity Paradox: AI Code Generation vs. Accessibility Lag

The software industry is experiencing another massive paradigm shift, highlighted by GitHub statistics indicating that 40% of all newly written code is now generated by AI. However, this explosive velocity brings a severe catch: without proactive safeguards, AI-generated code risks accumulating unprecedented technical debt that excludes the 1.3 billion people with disabilities worldwide. Kumar warns that ignoring accessibility during this shift would represent a systemic failure not just for software engineering, but for society as a whole.

```mermaid flowchart TD accTitle: The Two-Way Disconnect in AI Code Generation subgraph S1["Traditional Testing"] A["Pre-AI Testing Cycles"] --> B["Takes Weeks to Audit"] end subgraph S2["AI Development"] C["AI Code Generation"] --> D["Ships in Seconds"] end subgraph S3["The Resulting Gap"] B -.-> E["Developers Bypass Testing"] D --> E E --> F["Bias & Semantic Defects Baked In"] end ```

The underlying crisis stems from a two-way disconnect between traditional accessibility practices and modern AI workflows: 1. **Speed disparity:** Accessibility testing traditionally operates in slow, manual, pre-AI cycles measured in weeks, while AI code generators produce output in seconds, prompting developers to bypass compliance steps entirely. 2. **Biased training datasets:** AI models are trained on existing web repositories that already ignore accessibility standards, encoding inherent biases into future outputs.

Because accessibility relies fundamentally on human intent, AI models lacking semantic understanding routinely output illegible color contrast and unnavigable digital forms. Kumar draws an analogy to the everyday frustration of staying in a hotel and struggling to distinguish between the shampoo and conditioner bottles because the low-contrast typography is nearly impossible to read.

---

[05:52](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=352s) Business Imperatives, Legal Stakes, and Semantic Futureproofing

Despite decades of web development, 96% of the web remains inaccessible. Kumar cites World Wide Web inventor Tim Berners-Lee, who asserted at Deque's annual axe-con conference: *"The power of the web is in its universality."* If the AI revolution leaves behind individuals with disabilities, the technology will not be truly revolutionary—it will simply be broken. Navigating this transition is not just about avoiding litigation; it presents a defining choice between allowing AI to become the great equalizer or the great divider.

The business and regulatory risks of ignoring accessibility are escalating dramatically. With the European Accessibility Act (EAA) enforcing penalties up to 5% of a company’s revenue across the European Union's $20 trillion GDP market, compliance has become non-negotiable. Beyond mitigating legal exposure, the commercial upside is massive: - **Untapped market:** One in four individuals has a disability, representing a primary demographic and a broader support network controlling significant discretionary spending. - **Measurable growth:** Microsoft reported expanding its addressable user base by 33% by prioritizing accessibility. - **Brand loyalty:** Younger demographics, including Gen Z and Millennials, actively favor socially responsible companies.

Crucially for developers at an AI conference, coding for accessibility directly futureproofs applications for AI agents. Autonomous AI agents rely on the exact same underlying semantic hooks, ARIA labels, and structured hierarchies that assistive technologies require. Failing to code accessibly locks out both human users and AI agents alike.

---

[09:54](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=594s) Screen Reader Demonstration and "Fighting Fire with Fire"

To demonstrate the real-world frustration of poor semantic markup, Kumar plays an audio clip of a blind user navigating a mock "Mars commuter" booking interface using a screen reader. The user encounters ambiguous radio options, unlabeled buttons reading only *"dot dot dot button to activate"*, broken interactive controls, and an isolated input asking *"Yes. Radio button checked"* without any contextual question—prompting the user to humorously remark, *"What am I saying yes to? I'm confused. Maybe it's a sign I shouldn't be going to Mars."*

Kumar stresses that this friction applies to essential daily tasks across work, study, and commerce. The solution to AI-generated code defects is to *"fight fire with fire"* by utilizing AI agents to catch and remediate accessibility defects in real time at the exact pace code is written.

Handing over the presentation to Dylan Barrell, Barrell outlines how Deque embeds accessibility directly into AI coding workflows within IDEs like Visual Studio Code. Rather than relying on generic LLMs that lack nuanced domain expertise, Deque injects decades of verified accessibility knowledge into the loop:

```mermaid flowchart LR accTitle: Developer Workflow with axe Platform AI Integration subgraph IDE["Developer IDE (e.g., VS Code)"] Dev["User / Developer"] <--> Agent["AI Coding Agent"] end subgraph DequeCore["axe Platform Intelligence"] KB["Industry Accessibility Knowledge & Standards"] --> Assistant["axe Assistant"] Assistant <--> Agent end ```

- **Domain expertise:** The world's certified accessibility professionals train on Deque’s curriculum to prepare for industry certification exams. - **Benchmark performance:** Deque’s specialized axe Assistant consistently outperforms standard general-purpose models like GPT when answering complex accessibility and compliance questions.

[13:02](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=782s) Model Context Protocol, the Inconsistency of "Vibe Coding," and the Need for axe Testing

The Model Context Protocol (MCP) bridges the gap between AI agents and deep domain knowledge. Through MCP, an AI agent can proactively query the axe platform with questions such as what the accessibility requirements are for implementing a specific form, retrieve those exact specifications, and immediately apply them when generating code. However, relying on AI generation alone is insufficient because AI models are inherently non-deterministic. If an agent is asked to perform the same task twice on the same day, it will produce two completely different implementations.

Illustrating this inconsistency, Dylan shares a personal anecdote from the previous weekend when he experimented with "vibe coding" a Pac-Man clone. Every time he asked the model to make small tweaks, it was unable to edit the existing code cleanly and instead replaced the entire codebase. In doing so, it discarded previous elements that worked well, broke gameplay functionality, and altered the visual aesthetics—right down to making the ghosts look different and fail to operate correctly—ultimately forcing him to abandon the attempt. Because AI code generation remains unpredictable and "nobody's going to ship production code that hasn't been tested," the axe platform supplies consistent, automated testing within a tight "code, test, fix, loop" to deliver accessibility compliance at the speed of AI.

```mermaid flowchart TD accTitle: Accessibility at the Speed of AI Loop subgraph DevWorkflow["Developer Workspace"] Dev["Developer"] Agent["AI Agent (VS Code Copilot)"] Codebase["Source Code & Localhost App"] end subgraph DequeAxe["axe Platform"] MCP["axe MCP Server"] Engine["axe Rules, Testing & Remediation Engine"] end

[14:39](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=879s) Demo: Automated Remediation in VS Code with Copilot Agent and axe MCP

A recorded demonstration inside Visual Studio Code shows GitHub Copilot operating in agent mode alongside a web application running locally on port 3033. When prompted to analyze `localhost:3033` for accessibility and fix any discovered issues, Copilot consults its available tools, identifies the axe MCP tool, prompts the user for execution permission, and inspects the live application. The scan uncovers three critical barriers that would severely impede screen reader users: poor color contrast, missing alternative text on images, and interactive links that lack an accessible name.

Copilot receives precise remediation instructions back from the axe MCP server and systematically navigates the project's source code to apply targeted fixes: - **Color Contrast:** Recognizing a contrast failure on a button with white text, the agent modifies the corresponding CSS file to apply a slightly darker background color. - **Dynamic Alt Text:** The agent pinpoints a missing `alt` attribute inside a reusable recipe card component. Although the component renders eight separate times on the page, the agent identifies the underlying component and dynamically assigns the recipe's title from the existing data model rather than hardcoding static text. - **Accessible Icon Links:** In the footer component, the agent identifies GitHub and LinkedIn icon links that lacked accessible names. To preserve the visual design without altering page layout, it injects descriptive `aria-label` attributes onto the link elements.

Rather than unilaterally overwriting files in a blind "vibe coding" style, the agent presents each diff for user approval. Once the edits are staged, the agent asks for permission to re-run the axe MCP validation tool to confirm that all three violations are resolved. This setup combines the agent's strength in contextual code navigation with the deterministic rigor of axe-core testing.

[18:02](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=1082s) Accessibility at the Speed of AI and Audience Wrap-Up

The complete workflow unites developer intent, AI execution, and deterministic accessibility verification into a single feedback loop: the developer sets instructions, the AI agent consults the axe platform via MCP for rules and testing, the agent refines the source code, and the developer reviews, edits, or approves the final pull request.

- Attendees interested in the technology are directed to visit Deque Systems at **Booth 149** in the exhibit hall. - A QR code is provided for audience members to join the waitlist for the **axe MCP** product or schedule an enterprise sales demonstration. - After a brief pause for audience questions with no hands raised, the presenters thank the audience and conclude the session.

## Notable quotes > [05:52](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=352s) The power of the web is in its universality. — Tim Berners-Lee

> [09:54](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=594s) What am I saying yes to? I'm confused. Maybe it's a sign I shouldn't be going to Mars.

> [09:54](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=594s) fight fire with fire — Preety Kumar

> [13:02](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=782s) nobody's going to ship production code that hasn't been tested — Dylan Barrell

## Mentioned Deque Systems, Preety Kumar, Dylan Barrell, Charles Handy, Tim Berners-Lee, axe-core, axe Platform, axe Assistant, axe MCP, Google Chrome, Microsoft, GitHub, GitHub Copilot, Visual Studio Code, European Accessibility Act (EAA), GPT, Model Context Protocol (MCP), LinkedIn

## Video description Preety Kumar, CEO/Founder, Deque Systems Dylan Barrell, Chief Technology Officer, Deque Systems

Abstracts

The web was already inaccessible before AI. Now, with digital content creation accelerating exponentially, the problem is getting worse, not better. And that’s a serious issue. Accessible products aren’t just higher code quality, and they aren’t just better for people with disabilities—they’re better for everyone.

In my talk, I’ll explain how the integration of AI agents and advanced digital accessibility tools empowers developers to build better and more accessible products faster, while delivering against a wide range of essential business objectives—greater market share, increased loyalty, minimized financial and legal risk, compliance, and more.

If the question is, how can we build faster and better, AI plus digital accessibility is the answer.

[Source](https://www.youtube.com/watch?v=w7Qn3SmhBD4)

中文简报

Deque将Axe引擎接入AI编程智能体以规避合规风险 AI模型正快速复现既有的网页设计缺陷，亟需将确定性规则引擎直接嵌入开发者编程智能体，以防范技术债务和欧盟监管处罚。 Preety Kumar与Dylan Barrell探讨了现代软件的速度悖论，警告称AI代码的快速生成可能会加剧全球13亿残障人士面临的网页无障碍访问壁垒。他们指出了构建无障碍软件在商业、监管和技术层面的迫切动力，并提到结构化语义标记对自主AI智能体同样不可或缺。为填补这一鸿沟，Deque展示了如何利用模型上下文协议（MCP）将确定性axe-core测试直接嵌入VS Code等开发环境。通过结合GitHub Copilot的逐步演示，他们阐明了AI智能体如何在不陷入非确定性“氛围编码”（vibe coding）的情况下，实时自动检测、结合上下文分析并修复无障碍缺陷。随着 AI 生成近 40% 的新代码，传统耗时数周的无障碍审计与秒级代码交付形成严重脱节，加之模型基于 96% 缺乏无障碍支持的现有网页训练，正加速固化缺陷并可能孤立全球 13 亿残障人士。 (2:33) 《欧洲无障碍法案》（EAA）出台了最高可达企业营收 5% 的合规罚则；规范的语义结构与 ARIA 标签不仅能帮助微软等企业开拓 33% 的潜在市场，也是未来自主 AI Agent 正确解析网页交互的底层保障。 (5:52) 针对“氛围编码（vibe coding）”每次运行结果不可控且易破坏原有代码的问题，Deque 倡导“以 AI 治理 AI”，通过模型上下文协议（MCP）将经过数千亿次下载验证的 axe 规则引擎无缝注入 IDE。 (13:02) 现场演示展示了 VS Code 中的 GitHub Copilot Agent 如何借助 axe MCP 工具自主扫描本地应用，定向修复 CSS 对比度缺陷、动态为卡片组件绑定 Alt 文本并补全图标链接的 ARIA 标签。 (14:39)

Reading summary

Preety Kumar and Dylan Barrell of Deque Systems argue that the surge in AI-assisted code generation risks worsening digital exclusion unless accessibility verification is integrated directly into automated developer workflows. To counter this, they propose pairing non-deterministic generative coding agents with deterministic accessibility testing engines via the Model Context Protocol. The Velocity Gap and Biased Training Data (2:33) Software development is undergoing a rapid transition, with GitHub statistics indicating that AI models now generate 40% of newly written code. However, Kumar warns that this speed creates an acute velocity mismatch. While AI tools produce code in seconds, traditional accessibility auditing relies on manual cycles that can take weeks, frequently leading engineering teams to skip testing entirely. This dynamic threatens to accumulate technical debt that directly excludes the world's 1.3 billion people with disabilities. The issue is compounded by underlying training datasets. Because an estimated 96% of the web currently fails accessibility standards, generative models ingest and reproduce these historical design flaws. Lacking intrinsic semantic understanding, AI code generators regularly introduce severe interface bugs, such as unnavigable forms, missing interactive context, and low-contrast text that leaves digital interfaces unusable for screen-reader users. Economic, Regulatory, and Agentic Stakes (5:52) Accessibility compliance is rapidly shifting from a voluntary best practice to a legal and operational mandate. Kumar highlights the European Accessibility Act (EAA), which introduces regulatory penalties reaching up to 5% of company revenue across the European Union's $20 trillion GDP market. Beyond legal exposure, commercial incentives are substantial: one in four people has a disability, and Microsoft reported expanding its addressable user base by 33% after prioritizing accessible product design. For software architectures built around artificial intelligence, accessibility standards offer an additional technical benefit: semantic futureproofing. Autonomous AI agents navigate digital interfaces using the exact same ARIA attributes, semantic markup, and structured hierarchies that screen readers require. Applications constructed without accessible semantic foundations will consequently lock out both human users relying on assistive technology and autonomous AI agents attempting to interface with the software. Bridging Agent Non-Determinism With Deterministic Rules (13:02) To address accessibility issues created by rapid AI generation, Kumar advocates "fighting fire with fire" by embedding accessibility intelligence directly into development tools. However, Chief Product Officer Dylan Barrell cautions against relying solely on generic LLMs or unstructured "vibe coding." AI models are inherently non-deterministic, often producing conflicting implementations across identical prompts or breaking working architectures during refactoring. Deque addresses this unpredictability by linking coding agents to specialized domain tooling via the Model Context Protocol (MCP). Through an axe MCP server, agents inside development environments like Visual Studio Code can proactively retrieve established accessibility specifications and run automated checks. Grounding generative workflows in an engine like axe-core—which has recorded over three billion downloads—ensures high-speed issue detection without false positives, providing a deterministic safety rail for non-deterministic code generation. Automated IDE Remediation and Human Oversight (14:39) A technical demonstration inside Visual Studio Code illustrates how GitHub Copilot operates alongside the axe MCP server on a locally hosted web application. Prompted to audit and fix accessibility defects, the agent queries the MCP server, identifies three critical violations, and systematically resolves them in the source code rather than rewriting entire files. The agent adjusts CSS to correct low-contrast text, dynamically injects component data properties into missing image alt attributes rather than hardcoding static text, and adds descriptive aria-labels to unlabelled icon links in the footer. Crucially, the workflow preserves developer agency: the AI agent generates staged diffs for human review and re-runs deterministic verification checks through the MCP server to validate compliance before code is committed. Deque's framework demonstrates that while AI coding agents can accelerate remediation, generative models alone cannot guarantee compliant code due to their inherent non-determinism. Ensuring digital accessibility at scale requires coupling automated code generation with deterministic rules engines, automated verification loops, and final human review.

精炼阅读摘要

Deque Systems指出，随着AI生成代码占比迅速攀升，缺乏无障碍规范的训练数据正在加剧数字排斥，必须通过将确定性测试工具与领域知识接入AI工作流来同步解决无障碍合规问题。 AI代码生成加速与无障碍技术债的矛盾 (2:33) 软件行业正经历快速转变，GitHub统计数据显示目前已有40%的新编写代码由AI生成。然而，Deque Systems首席执行官Preety Kumar警告，AI模型的训练数据主要来源于过去忽视无障碍规范的现有网络资源，导致生成的代码频繁出现色彩对比度不足、表单无法导航等底层语义缺陷，进而可能将全球13亿残障人士排除在外。当前网络环境已有96%的内容未能实现无障碍访问。这一危机源于传统开发与现代AI工作流之间的双重脱节：一方面，传统的无障碍审计通常需要数周的手动测试周期，而AI代码生成只需数秒，促使开发者为了追求交付速度而跳过合规测试；另一方面，通用大语言模型缺乏对人类意图与无障碍语义规范的深度理解。Kumar借用在酒店辨识洗发水和护发素瓶身低对比度文字的日常困扰为例，说明缺乏良好语义和视觉对比的代码会给日常工作、学习和商业带来实际障碍。无障碍合规的商业回报、法律约束与Agent协同 (5:52) 忽视无障碍标准的商业和监管风险正在迅速上升。欧洲无障碍法案（EAA）在涵盖20万亿美元GDP的欧盟市场内，对违规企业执行最高达全球营业额5%的罚款，使合规成为强制性要求。在市场层面，全球四分之一的人口存在某种残障，其自身及支持网络构成了庞大的可支配支出群体；例如，微软曾报告通过优先考虑无障碍功能，其目标用户群扩大了33%，同时年轻群体也更偏好具有社会责任感的品牌。面向AI开发场景，无障碍代码对自主AI Agent同样具有技术层面的前瞻意义。Kumar与Dylan Barrell指出，自主AI Agent在解析和操作界面时，依赖与辅助技术完全相同的底层语义标记、ARIA标签和结构化层级。若未能实现无障碍编码，不仅会阻碍人类辅助技术用户，也会导致未来的AI Agent无法正常识别并执行任务。基于MCP协议与确定性规则约束AI随机性 (13:02) 通用大语言模型具有非确定性，同一任务在不同调用下可能产生截然不同的实现代码。Deque首席产品官Dylan Barrell通过尝试对吃豆人克隆程序进行“氛围编程”（vibe coding）的个人经历指出，模型在调整局部逻辑时容易完全重写代码库并破坏既有功能，这种不稳定性使得未经严格测试的AI代码无法直接投入生产。为了让AI编码保持合规，Deque利用模型上下文协议（Model Context Protocol, MCP）将专业无障碍知识库和axe规则引擎接入IDE（如VS Code）。已录得超30亿次下载且内嵌于Google Chrome和微软技术栈的axe-core规则引擎，主打零误报检测能力。通过MCP，AI Agent在编写代码前可主动查询特定组件的合规规范，并借助axe Assistant超越通用GPT模型的垂直领域性能，实现以AI治理AI代码缺陷。 IDE内由Agent驱动的自动化检测与修复闭环 (14:39) 在Visual Studio Code环境的实际演示中，GitHub Copilot以Agent模式运行，通过axe MCP工具对运行在本地端口3033的Web应用进行检查。扫描发现了三项严重阻碍屏幕阅读器用户的缺陷：按钮色彩对比度不足、组件图像缺失替代文本（alt属性），以及页脚交互图标链接缺少无障碍名称。获取axe MCP服务器提供的具体修复方案后，Copilot直接在源码层级实施精确修复，而非进行破坏性重写：它微调了CSS文件以提高文本背景对比度；为在页面上渲染八次的食谱卡片组件动态绑定数据模型中的标题作为alt文本，避免硬编码；并在页脚社交图标链接中添加了aria-label属性。最后，Agent主动请求权限重新运行axe MCP测试进行验证，形成“编码、测试、修复”的闭环反馈。该方案展示了将人类意图、AI生成与确定性规则验证相结合的协同路径，但企业在实际生产落地中仍高度依赖开发者对Agent提交代码差异的最终人工审核，以规避AI非确定性带来的潜在风险。

Detailed notes — zh

## Background 本视频由频道 Ai4 于 2025 年 9 月 22 日发布在 YouTube 平台。视频记录了 Deque Systems 创始人兼 CEO Preety Kumar 与 CTO Dylan Barrell 的主题演讲，主要探讨如何将 AI Agent 与数字无障碍工具深度融合，助力开发者在 AI 时代更高效地构建普惠所有人且兼顾商业与合规价值的高质量产品。 ## TL;DR 演讲指出，当前 AI 虽以秒级速度生成海量代码，但传统无障碍测试的滞后与训练偏差正加剧全球 13 亿残障人士的技术排斥，并让企业面临《欧洲无障碍法案》（EAA）高达全球营收 5% 的合规罚款。同时，无障碍语义标签也是未来 AI Agent 理解和操作页面的核心接口，不做无障碍优化将导致代码无法面向未来。针对 AI 编程的不稳定性与破坏性重写问题，Deque 推出了基于 Model Context Protocol（MCP）的 axe 平台集成方案。现场实机演示了 VS Code Copilot 在 axe MCP 赋能下，自动检测并精准修复色彩对比度、动态 alt 属性和 aria-label 缺陷的全流程，展现了以确定性测试闭环在 AI 速度下实现无障碍合规的前沿实践。 ## Atmosphere & style 整场演讲兼具强烈的社会责任感与硬核的技术实操属性，开篇通过真实的屏幕阅读器订票音频演示营造出令人感同身受的无力感，迅速唤醒现场对数字无障碍痛点的重视。随后的 CTO 演示穿插着关于“氛围编程”（Vibe Coding）失控的幽默自嘲，让严谨的工程挑战显得格外接地气。整体叙事节奏明快、逻辑层层递进，既有宏观法律与商业层面的警示，又有基于 MCP 的前沿工具链落地演示，给与会者带来极具说服力与前瞻性的科技冲击。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=4s) Deque Systems 旗下的开源引擎 Axe Core 拥有超 30 亿次下载量并深度整合入 Chrome 和微软生态，其秉持“零误报”原则保障代码默认具备可访问性。 - [02:33](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=153s) GitHub 上已有 40% 的新增代码由 AI 生成，但 AI 模型往往缺乏对人类意图的理解，导致生成的代码可能将全球 13 亿残障人士隔绝在数字化体验之外。 - [03:05](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=185s) 传统无障碍测试需耗时数周，与现代 AI 秒级的生成速度存在严重代差，导致开发者在追求上线效率时频繁跳过无障碍验证。 - [05:52](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=352s) 全球合规监管迅速收紧，《欧洲无障碍法案》（EAA）针对单项缺陷的最高罚款可达企业全球营收的 5%，对进入 20 万亿美元规模的欧盟市场构成刚性约束。 - [06:55](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=415s) 数字无障碍蕴含巨大的商业机遇，全球残障群体及其网络拥有 1.9 万亿美元的可支配支出，微软通过无障碍优化实现了 33% 的用户基数增长。 - [08:05](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=485s) 无障碍语义标签（Semantic Hooks）是代码面向未来的关键，不仅供视障人士的屏幕阅读器解析，也是未来自主 AI Agent 操作网页所必需的底层接口。 - [09:54](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=594s) 面对 AI 大规模生成不可访问代码的挑战，必须“以火攻火”，利用专业的 AI Agent 在毫秒级实时捕捉并自动修复无障碍缺陷。 - [13:02](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=782s) AI 编程在“氛围编程”（Vibe Coding）时存在输出不一致与全量重写破坏已有逻辑的缺陷，未经自动化测试验证的代码绝不能直接上生产线。 - [13:40](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=820s) 通过 Model Context Protocol（MCP），axe 平台为 AI Agent 提供精准的无障碍标准规范与测试工具，构建“编码—测试—修复—验证”的闭环。 - [14:39](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=879s) 在实机演示中，AI Agent 结合 axe MCP 成功实现了 CSS 对比度调色、动态绑定组件属性作为图片 alt、添加 aria-label 补全链接名称等多项精准修复。 - [18:02](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=1082s) “以 AI 的速度实现无障碍”模式在释放 AI 生产力的同时，通过开发者把关（Accept / Edit / Reject）彻底杜绝了合规盲区与技术不确定性。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=4s) 引言与数字无障碍的愿景演讲开场，Deque Systems 创始人兼首席执行官 Preety Kumar 向听众致意，并介绍了技术总监兼产品负责人 Dylan Barrell。本次演讲的核心议题聚焦于：在 AI 与新一代技术浪潮席卷而来的当下，我们如何确保技术是在“拓展”可访问性与接入渠道，而非“限制”它们。为了为整场讨论定下基调，Preety 引用了她极为推崇的社会哲学家 Charles Handy 的名言： > “未来并非不可避免，我们能够影响它——你和我都可以影响它。但前提是，我们必须清楚自己希望未来成为什么样。” Preety 介绍了 Deque Systems 的核心使命：让信息技术无障碍地服务于全球所有残障人士。作为无障碍领域的领头羊，Deque 开发了知名的开源规则引擎 Axe Core。该引擎目前拥有超过 30 亿次下载量，已深度嵌入至 Google Chrome 以及微软的核心技术体系中，几乎无处不在。依托 Axe Core 的 Axe 平台，设计师、开发者和测试人员得以在“AI 级别的速度”下以极小的工作量发现并修复无障碍缺陷。开发者青睐该平台的核心原因在于其坚持“零误报”（Zero False Positives）的原则，确保报告的每一个缺陷都是真实存在的问题。Preety 强调，高质量的代码在默认状态下本就应当对每个人都友好可用。随着 AI 的飞速演进，行业必须跟上这一节奏，决不能为了追求 AI 的生成速度而牺牲代码应有的社会责任与规范。 - **核心数据与产品指标**： - 开源规则引擎：Axe Core - 全球下载量：已超过 30 亿次 - 生态整合：深度集成至 Chrome 与 Microsoft 工具链中 - 核心设计原则：零误报（Zero False Positives），确保高质量代码默认普惠所有人 --- ### [02:33](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=153s) 软件范式转变与 AI 代码的技术债隐患整个软件行业正经历一场前所未有的巨变浪潮。GitHub 数据显示，如今其平台上已有 40% 的新增代码是由 AI 辅助生成的。这标志着软件开发领域有史以来最重大的范式转变。然而，在惊人的生成速度背后隐藏着严重的隐患：AI 带来的代码生成速度极可能演变成巨大的技术债务，将全球 13 亿残障人士彻底阻隔在数字化体验之外。如果将这部分群体排除在外，这不仅是 AI 编程工具的技术失误，更是整个人类社会的系统性溃败。当前数字无障碍（Digital Accessibility）的实践方式严重脱节，依然停留在“前 AI 时代”的缓慢节奏中。传统的无障碍测试往往以“周”为单位推进，周期冗长，完全无法嵌入到现代 AI 辅助的秒级开发工作流中。这种速度上的巨大代差导致开发者在追求上线效率时直接选择绕过无障碍测试。更严重的问题在于两者之间的“双向脱节”：不仅测试方式脱节，AI 自身的代码生成逻辑也存在固有缺陷。AI 模型的训练数据往往忽视了残障人士的实际需求，从而将偏见内置到生成逻辑中。由于无障碍设计高度依赖“人类意图”（Human Intent），没有理解人类意图的 AI 生成的代码往往充满低级错误，例如对比度极低导致无法辨认的文本、或者让视障屏幕阅读器完全无法解析的无标签表单。Preety 以听众身处酒店的日常细节幽默比喻道：就像人们在酒店洗澡时，经常因为洗发水和护发素瓶身上字迹对比度极低而难以分辨一样，AI 正在批量生成这种无法正常辨认的代码。 ```mermaid flowchart TD accTitle: AI快速生成与传统无障碍测试的脱节循环 subgraph AI_Speed["AI 高速生成工作流 (秒级)"] A["开发者输入 Prompt 提示词"] --> B["AI 快速生成业务代码"] B --> C["缺乏无障碍意图与对比度设计"] end subgraph A11y_Gap["传统无障碍验证 (周级)"] C --> D{"传统人工/慢速测试"} D -- "耗时数周 / 拖慢进度" --> E["开发者直接跳过测试"] E --> F["系统性无障碍技术债累积"] end ``` - **行业现状与痛点**： - GitHub 平台上 40% 的新增代码由 AI 生成 - 全球受影响残障人口总数达 13 亿 - 速度鸿沟：AI 生成代码仅需数秒，传统无障碍测试需耗时数周 - 网页现状：全网仍有 96% 的网站存在无障碍缺陷 --- ### [05:52](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=352s) 商业价值、法律合规与“AI Agent 接口”新视角万维网发明者 Tim Berners-Lee 在出席 Deque 的年度 Axe-con 大会时曾指出：“Web 的力量在于其普适性（Universality）。普适性不是锦上添花，而是 Web 具有变革性力量的根本原因。”如果 AI 革命将残障群体排斥在外，它将失去革命性的意义，最终变成一个支离破碎的系统。Preety 指出，无障碍设计绝不仅仅是为了应付诉讼，而是一个决定 AI 成为“伟大的平等器”还是“巨大的隔离墙”的历史契机。在法律与合规层面，全球监管正迅速收紧。以近期备受瞩目的《欧洲无障碍法案》（European Accessibility Act，简称 EAA）为例，其监管力度堪比 GDPR：每个缺陷均可被处以罚款，最高罚金可达企业全球年营业额的 5%。考虑到欧盟拥有高达 20 万亿美元的 GDP 市场体量，任何企业都无法承受被该市场排除在外的代价。除合规之外，商业层面的收益更为显著。全球每四个人中就有一人存在某种形式的残障，若加上其亲朋好友，这一群体的可支配支出高达 1.9 万亿美元，是全球最大的未充分开拓市场。微软曾公开表示，通过提升产品无障碍性，其用户基数直接增长了 33%。此外，注重无障碍设计有助于提升品牌忠诚度，并在年轻一代（Gen Z 与千禧一代）中赢得更高的社会责任声誉。 Preety 向现场的 AI 开发者抛出了一个极具震撼力的技术论点：无障碍设计是让代码“面向未来”（Future-proof）的关键。无论是视障人士使用的屏幕阅读器，还是未来在网络上自主执行任务的 AI Agent，都需要依赖相同的语义化标签（Semantic Hooks）。如果不做无障碍优化，系统不仅会将人类用户拒之门外，同样也会让 AI Agent 无法理解和操作页面。 - **关键商业与法律数据**： - 欧洲无障碍法案（EAA）最高罚款：可达公司全球营收的 5% - 欧盟 GDP 市场规模：20 万亿美元 - 全球人群画像：每 4 人中就有 1 人存在残障（1/4 比例） - 全球残障群体可支配支出：1.9 万亿美元 - 微软案例：无障碍举措助力用户基数扩大 33% --- ### [09:54](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=594s) 屏幕阅读器实测演示与 Axe 智能代理方案为了让与会者直观感受无障碍缺失的糟糕体验，Preety 在现场播放了一段盲人用户使用屏幕阅读器尝试预订火星旅行车票的真实音频演示： ``` [屏幕阅读器快速播报] Mars commuter one-way radio button checked. One of seven. [用户独白] 我最好订一张往返票。 [屏幕阅读器] Round-trip radio button checked. To change the selection, press space. Dot dot dot button to activate. Press enter. [用户困惑] 这个“点点点（dot dot dot）”按钮是什么？ [用户按下空格] Space. Space. 看来什么都没发生，它根本不起作用。 [屏幕阅读器] Button to activate. Press enter. [用户] 这个按钮连个标签（Label）都没有，按一下试试看。 [用户按下空格] Space. 还是没反应。 [屏幕阅读器] Dot dot dot button. To activate, press enter. [用户] 怎么又绕回刚才这里了？ [屏幕阅读器] Box 12 p.m. To change the selection, use the arrow key. 10 night 7 a.m. [用户] 那就选早上 7 点出发吧。 [屏幕阅读器] "Yes" radio button checked to change this. [用户抓狂] 这里的“Yes”到底是对什么说是？我彻底搞糊涂了……可能这就是个信号，告诉我根本不该去火星。 ``` Preety 强调，这种令人绝望的体验不仅发生在订票场景，在日常办公、学习和娱乐中比比皆是；更重要的是，缺乏语义结构的页面同样会让 AI Agent 彻底陷入混乱。因此，面对 AI 大规模生成不可访问代码的现状，我们必须“以火攻火”（Fight fire with fire）——利用 AI Agent 以相同的生成速度，在代码编写的毫秒级实时捕捉并自动修复无障碍缺陷。随后话筒交给了 CTO Dylan Barrell。Dylan 展示了典型的 AI 编程工作流（如用户在 VS Code 等 IDE 中与 AI 编程助手交互生成代码的循环）。Deque 将 Axe 平台的核心能力注入到了这一工作流中，其首要支柱便是行业首屈一指的“权威知识库”。全球所有认证无障碍专业人士（Certified Accessibility Professionals）均通过 Deque 的课程进行备考，其自主研发的 Axe Assistant 在专业无障碍认证考试的答题测试中表现已经超越了通用的大语言模型 GPT。 ### [13:02](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=782s) AI 编程的不一致性挑战与 Model Context Protocol (MCP) 的结合在现代开发流程中，Deque 的 axe 平台通过 Model Context Protocol（MCP）为 AI Agent 提供实时的无障碍上下文支持。开发人员使用的 AI Agent 可以直接向底层协议提问，例如查询“以无障碍方式实现该表单的具体要求是什么”，系统便会返回准确的技术规范，Agent 进而依据这些标准生成符合规范的代码。然而，当前的 AI Agent 虽然极擅长快速生成代码，却普遍存在严重的输出不一致性问题——即使在同一天内要求它执行两次完全相同的任务，它也往往会给出两种截然不同的实现方式。演讲者分享了自己周末尝试使用新模型进行“氛围编程”（Vibe Coding）制作《吃豆人》（Pac-Man）克隆版的亲身经历：每当要求模型调整某些功能时，AI 往往无法进行精准的局部增量修改，而是倾向于直接重写全部代码；重写后，之前令人满意的特性往往被意外破坏，屏幕画面变样，幽灵角色的外观与移动逻辑彻底错乱，最终不得不让人放弃。这一痛点清晰地表明：没有任何团队敢在未经严格测试的情况下将 AI 生成的代码直接推向生产环境。为了解决这一可靠性缺失，axe 平台通过 MCP 提供了高一致性的自动化测试能力，将 AI 纳入“编码—测试—修复”的闭环，实现真正兼具速度与合规性的高质量开发。 ```mermaid flowchart TD accTitle: axe 与 AI Agent 的无障碍闭环协作流程 subgraph DevEnv["开发环境 (Agent Mode)"] A["开发者提示词 / 任务指令"] --> B["AI Agent (如 Copilot)"] B --> C["代码修改 / 修复建议"] end subgraph AxeEngine["axe Platform & MCP"] D["MCP Server 工具发现与鉴权"] --> E["无障碍规范与测试分析"] E --> F["修复方案推荐与回归验证"] end B <-->|请求上下文与工具调用| D C -->|提交回归测试| E ``` - **一致性困境**：AI Agent 缺乏可预测的输出稳定性，全量重写代码容易导致已有功能损坏与 UI 退化。 - **闭环测试核心**：通过 axe 平台提供确定性的测试基准，构建严密的“代码—测试—修复—验证”（Code-Test-Fix-Validate）工作流。 --- ### [14:39](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=879s) 现场实机演示：VS Code Copilot 与 axe MCP 的自动化无障碍修复演讲者通过一段在 VS Code 中结合 Copilot Agent Mode 的实操录屏展示了这套系统的实际运作。在演示中，终端正运行着位于 `localhost:3033` 的本地 Web 应用程序，侧边栏为代码编辑器。当向 Copilot 下达指令“分析 localhost:3033 的无障碍问题并修复所发现的缺陷”后，Agent 首先利用 MCP 协议检索可用工具，定位到 axe MCP 工具并请求开发者授权运行，随后对本地运行的应用展开深度扫描。扫描迅速定位出三类对屏幕阅读器（Screen Reader）等辅助技术用户影响极为严重的无障碍缺陷，并给出了针对性的修复策略： 1. **色彩对比度缺陷（Color Contrast）**：白色文本按钮因背景色过浅导致可读性不足。AI Agent 准确识别出该样式位于 CSS 样式文件中，自动提取并应用了明度更暗的背景颜色以满足无障碍对比度阈值，同时主动请求开发者确认采纳而非盲目覆盖。 2. **图片缺少替代文本（Missing Alt Text）**：在页面中复用了 8 次的“食谱卡片”（Recipe Card）组件中，图片标签缺少 `alt` 属性。Agent 聪明地定位到该单一组件模板，并且没有采取生硬的硬编码文本，而是从组件代码中提取动态的食谱名称绑定为图片的 `alt` 属性。 3. **缺少可访问名称的链接（Links with No Accessible Name）**：页脚（Footer）组件中指向 LinkedIn 和 GitHub 的图标链接缺少文本描述。Agent 识别出这是视觉图标链接，为了在不破坏现有页面排版布局的前提下提供语义信息，自动为其添加了 `aria-label` 属性。在完成所有代码修复建议后，Agent 再次向开发者申请权限，通过 axe MCP Server 执行二次回归验证，确保所有修复均达到标准且未引入新问题。这一过程充分展现了 AI 理解业务意图与代码结构的能力，结合底层 axe 测试引擎的确定性保障，构建了一个安全可信的修复闭环。 - **三项关键修复成果**： - **CSS 调色**：提高白色文字按钮的背景对比度。 - **动态语义注入**：为 Recipe Card 组件绑定动态食谱名称作为图像 `alt`。 - **无侵入可访问名称**：通过 `aria-label` 为页脚社交媒体图标链接赋予无障碍名称。 - **开发者控制权**：保留人类审核环节（Accept / Edit / Reject），避免纯粹 Vibe Coding 带来的黑盒风险。 --- ### [18:02](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=1082s) “AI 速度下的无障碍合规”总结与展台互动演讲者将这一整套协同体系定义为“以 AI 的速度实现无障碍”（Accessibility at the Speed of AI）。其核心逻辑在于形成开发者、AI Agent 与 axe 平台的三方高效协作：开发者下发意图指令；AI Agent 借助 axe 平台获取规范指导、修复上下文与测试工具；Agent 将修正代码应用至项目；最后由开发者决定接受、编辑或拒绝特定改动。这种模式既释放了生成式 AI 快速重构代码的生产力，又彻底杜绝了无障碍合规中的盲区与不确定性。在演讲尾声，团队向现场与会者发出邀请： - **展台交流**：欢迎前往展厅 149 号展位（Booth 149）与团队进行深入技术探讨。 - **产品试用**：现场提供二维码供大家申请加入 **axe-MCP** 产品的专属候补名单（Waitlist），或预约销售团队的产品演示。 - **结语**：在确认现场观众没有进一步提问后，演讲者向全场参会者致谢并圆满结束本次演讲。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=4s) 未来并非不可避免，我们能够影响它——你和我都可以影响它。但前提是，我们必须清楚自己希望未来成为什么样。 — Charles Handy > [00:04](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=4s) 高质量的代码在默认状态下本就应当对每个人都友好可用。 — Preety Kumar > [05:52](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=352s) Web 的力量在于其普适性（Universality）。普适性不是锦上添花，而是 Web 具有变革性力量的根本原因。 — Tim Berners-Lee > [05:52](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=352s) 无障碍设计绝不仅仅是为了应付诉讼，而是一个决定 AI 成为“伟大的平等器”还是“巨大的隔离墙”的历史契机。 — Preety Kumar > [09:54](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=594s) Mars commuter one-way radio button checked. One of seven. — 屏幕阅读器 > [09:54](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=594s) 这个“点点点（dot dot dot）”按钮是什么？ — 演示用户 > [09:54](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=594s) 这里的“Yes”到底是对什么说是？我彻底搞糊涂了……可能这就是个信号，告诉我根本不该去火星。 — 演示用户 > [09:54](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=594s) 以火攻火（Fight fire with fire）——利用 AI Agent 以相同的生成速度，在代码编写的毫秒级实时捕捉并自动修复无障碍缺陷。 — Preety Kumar > [14:39](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=879s) 分析 localhost:3033 的无障碍问题并修复所发现的缺陷 — 开发者提示词 > [18:02](https://www.youtube.com/watch?v=w7Qn3SmhBD4&t=1082s) 以 AI 的速度实现无障碍（Accessibility at the Speed of AI） — Dylan Barrell ## Mentioned Preety Kumar, Dylan Barrell, Charles Handy, Tim Berners-Lee, Deque Systems, Axe Core, Axe 平台, Axe Assistant, axe-MCP, Model Context Protocol (MCP), Google Chrome, Microsoft, GitHub, VS Code, GitHub Copilot, 欧洲无障碍法案 (EAA), GDPR, Axe-con, Pac-Man ## Video description Preety Kumar, CEO/Founder, Deque Systems Dylan Barrell, Chief Technology Officer, Deque Systems Abstracts The web was already inaccessible before AI. Now, with digital content creation accelerating exponentially, the problem is getting worse, not better. And that’s a serious issue. Accessible products aren’t just higher code quality, and they aren’t just better for people with disabilities—they’re better for everyone. In my talk, I’ll explain how the integration of AI agents and advanced digital accessibility tools empowers developers to build better and more accessible products faster, while delivering against a wide range of essential business objectives—greater market share, increased loyalty, minimized financial and legal risk, compliance, and more. If the question is, how can we build faster and better, AI plus digital accessibility is the answer. [Source](https://www.youtube.com/watch?v=w7Qn3SmhBD4)

Source quotes

[0:04] Charles Handy: 未来并非不可避免，我们能够影响它——你和我都可以影响它。但前提是，我们必须清楚自己希望未来成为什么样。 [0:04] Preety Kumar: 高质量的代码在默认状态下本就应当对每个人都友好可用。 [5:52] Tim Berners-Lee: Web 的力量在于其普适性（Universality）。普适性不是锦上添花，而是 Web 具有变革性力量的根本原因。 [5:52] Preety Kumar: 无障碍设计绝不仅仅是为了应付诉讼，而是一个决定 AI 成为“伟大的平等器”还是“巨大的隔离墙”的历史契机。 [9:54] 屏幕阅读器: Mars commuter one-way radio button checked. One of seven. [9:54] 演示用户: 这个“点点点（dot dot dot）”按钮是什么？ [9:54] 演示用户: 这里的“Yes”到底是对什么说是？我彻底搞糊涂了……可能这就是个信号，告诉我根本不该去火星。 [9:54] Preety Kumar: 以火攻火（Fight fire with fire）——利用 AI Agent 以相同的生成速度，在代码编写的毫秒级实时捕捉并自动修复无障碍缺陷。 [14:39] 开发者提示词: 分析 localhost:3033 的无障碍问题并修复所发现的缺陷 [18:02] Dylan Barrell: 以 AI 的速度实现无障碍（Accessibility at the Speed of AI）

Source quotes

[5:52] Tim Berners-Lee: The power of the web is in its universality. [9:54] What am I saying yes to? I'm confused. Maybe it's a sign I shouldn't be going to Mars. [9:54] Preety Kumar: fight fire with fire [13:02] Dylan Barrell: nobody's going to ship production code that hasn't been tested

Original transcript

[0:04] Welcome everybody. Thanks for joining [0:07] us. [0:08] I am Priti Kumar, the CEO and founder of [0:12] DQ Systems and this is Dylan, our chief [0:16] product officer. We are here today to [0:19] talk about how the next era of [0:22] technology and AI [0:25] expands access and not restricts it. [0:29] Before [0:31] we dive in, I want to ground today's [0:34] talk in a very simple truth. [0:38] And maybe I'll say it the way that one [0:40] of my favorite social philosophers, [0:42] Charles Handy, said it. He was good at [0:45] predicting the future. I'll tell you [0:47] that. And he said, "The future is not [0:50] inevitable. [0:52] We can influence it. You and I can [0:55] influence it. [0:59] we can only influence it if we know what [1:02] we want it to be. [1:05] So [1:07] as we think about it um at DQ our [1:11] mission has been to make information [1:13] technology accessible to people with [1:14] disabilities and we are the makers of [1:18] Axe Core. It's an open-source uh rule [1:21] engine that has over three billion [1:23] downloads. It's uh embedded in Chrome. [1:26] It's embedded in Microsoft technologies. [1:28] It's ubiquitous. Essentially, [1:32] our product line Axe platform [1:35] allows designers, developers, testers to [1:38] be able to find and fix accessibility [1:41] issues at the speed of AI [1:44] with minimal effort of course. Um, and [1:48] the reason that developers love our [1:50] platform is because of zero false [1:52] positives. We we are very careful to [1:55] make sure that every defect that we [1:57] surface is really a defect. A and [2:01] I guess this is how [2:05] good code quality should actually what [2:07] it really should look like by default [2:10] for everyone. It's just that simple. [2:14] Today AI is racing ahead and we have to [2:18] keep up with the pace of AI. We have to [2:21] make sure that AI speed doesn't get [2:24] responsibility of what AI should look [2:26] like to to you know be bypassed. [2:31] So let's get started. [2:33] All right. Living what what age are we [2:36] living in? We living in a tidal wave of [2:38] change in software yet again. Right. [2:42] We've gone through this before. [2:44] uh GitHub git well yeah GitHub I think [2:47] said that 40% of all new code that [2:50] they're generating today is being [2:51] generated by AI 40% of all new code that [2:56] they write and so this is the biggest [2:59] paradigm shift that we have ever seen [3:02] okay but here's the catch [3:06] AI's velocity [3:08] may end up resulting in a lot of [3:11] technical debt for all of [3:14] 1.3 billion people with disabilities [3:17] globally if we exclude them from what we [3:20] doing it would be a systemic failure not [3:23] just of AI coding but of us as a [3:26] society. [3:32] All right. So what's going on here? What [3:34] what are we talking about? Digital [3:36] accessibility is not keeping up with the [3:38] speed of AI. Here's the problem. digital [3:42] accessibility the way that is done [3:43] today, it's really very very slow, slow [3:46] test cycles and it's really operating in [3:48] a pre-AII kind of a world. It doesn't [3:51] integrate with AI assisted dev [3:54] workflows. And so what happens [3:57] developers bypass it and what does that [4:00] do to us? It takes weeks, [4:05] not seconds. AI generates codes in [4:07] seconds and manual testing, [4:10] accessibility testing is done in weeks. [4:13] It's a big disconnect. But the [4:16] disconnect, by the way, guys, is both [4:18] ways. It's a two-way disconnect. Not [4:20] just accessibility testing and how it's [4:22] being done, but how AI is writing code. [4:27] We are shipping features at velocity [4:30] that humans [4:33] with. Hello. Okay. All right. And but [4:36] the problem is the AI train training [4:39] data that we have it ignores people's [4:42] people with disabilities and builds in [4:45] inherent biases. [4:48] So bias baked in code that is produced [4:52] is not going to be accessible and [4:56] accessibility is especially vulnerable [4:58] to this because accessibility is about [5:02] human intent. Human intent is what [5:05] produces accessibility. [5:08] Um, [5:10] so [5:12] think about AI that's generating code [5:15] that doesn't have any idea of things [5:17] like accessibility, like color contrast. [5:20] Have you guys I mean you guys are in a [5:23] hotel right now, right? Have you ever [5:25] struggled with trying to read a shampoo [5:28] bottle or a conditioner bottle and [5:30] confuse the two because the writing is [5:32] so difficult to decipher? It's so faint. [5:36] Well, AI is generating code with color [5:39] contrast that is not legible. It's [5:42] creating inaccessible forms that [5:44] somebody who's blind simply cannot [5:46] navigate. [5:48] Okay. So, [5:52] Okay. So why am I going on about this? [5:54] Why is this important? Why should all of [5:56] you care about this? Well, 96% of the [5:59] web is still inaccessible. [6:03] The inventor of the worldwide web, Tim [6:06] Berners Lee, when he was presenting at [6:08] our annual conference, Axe Con, he said, [6:11] "The power of the web is in its [6:14] universality. [6:16] It's not a nice to have. It's the very [6:20] reason that the web became so [6:22] transformative. [6:24] So the next revolution that we are [6:27] experiencing, AI, if it excludes people [6:31] with disabilities, it's not going to be [6:33] revolutionary. It's going to be simply [6:36] broken. [6:38] And it's not just about avoiding [6:40] lawsuits. It's a once-in-a-lifetime [6:43] chance for all of you to decide whether [6:45] AI becomes the great equalizer or the [6:49] great divider. Your choice. [6:53] So, you know, um the risk of getting [6:58] this blueprint wrong is getting the [7:01] stakes are getting higher and higher. [7:02] New regulations are being created every [7:05] single day and this is to make the [7:08] information highway accessible. The most [7:10] recent one is the European Union. Have [7:13] any of you dealt with GDPR? If you did, [7:16] you know that's really serious stuff. [7:19] The European Commission takes its [7:21] regulations very seriously. [7:24] And the European Accessibility Act has [7:28] fines for every defect actually and it [7:31] can be up to 5% of a company's revenue. [7:35] So this is not, you know, something that [7:36] you can just sniff at, right? we have to [7:38] do this and the European GDP is 20 [7:42] trillion. [7:44] That's a very large market as well. [7:47] So can't ignore that either. [7:50] All right, but let's talk about the [7:52] upside. I I mean, you know, it's great [7:54] compliance, all of that good stuff. But [7:56] one in four people have disabilities. [7:59] It's the largest untapped market. And [8:01] that does not even include the friends [8:02] and families of people with [8:04] disabilities. It's a $1.9 billion dollar [8:08] in discretionary spending. So this is [8:10] not a detour. It's not some side street. [8:14] It's actually the main street. So [8:17] designing for all is the right thing to [8:19] do. It helps you get it. Well, it helps [8:23] you expand reach, reach a loyal customer [8:26] base and strengthens your brand. [8:28] Microsoft said it expanded their user [8:31] base. Accessibility expanded the user [8:34] base by 33%. [8:37] 33%. [8:40] So accessibility the whole product is [8:45] going to get stronger as a result of [8:47] that. It is going to reduce your [8:49] compliance risk and it's going to you [8:52] should want to avoid costly retrofitting [8:54] and lawsuits. Okay. It increases your [8:57] market share and reaches audiences that [9:00] your competition can't reach. Um, it [9:03] increases loyalty and is generally [9:06] better for your brand. And if you think [9:08] about the Gen Xers and the millennials [9:10] or Gen Z's, I think yes, they are [9:13] consistently going to choose your brand [9:15] if you're a socially responsible u uh [9:18] company really. So, but here's the [9:21] kicker, guys. You all are here at an AI [9:24] conference and here's what I'm going to [9:26] tell you. [9:29] If you code with accessibility in mind, [9:31] you're going to futureproof your code. [9:34] Okay? Because not just humans but AI [9:38] agents need the same semantic hooks that [9:41] we are inserting for accessibility. So [9:44] you are going to actually be if you [9:47] don't do accessibility, you're going to [9:49] lock out humans and AI agents. [9:54] So, I'm going to show you what an [9:55] inaccessible experience is like. You [9:58] know, [10:00] somebody who's blind, navigating a [10:02] website is what you're going to do. What [10:04] you're going to hear now [10:10] is fast. I better book a reservation. [10:13] >> Mars commuter one-way radio button [10:15] checked. One of seven. [10:16] >> I better look for a round trip. [10:18] >> Round- trip radio button checked. To [10:20] change the selection, press space. Dot [10:23] dot dot button to activate. Press enter. [10:26] >> I wonder what this dot dot dot button [10:28] is. [10:28] >> Space. Space. [10:29] >> I guess it's nothing cuz it's not [10:31] working. [10:31] >> Button to activate. Press enter. [10:34] >> I wonder what this button here is. It [10:36] doesn't have a label. Let's push it. [10:38] >> Space. [10:38] >> Doesn't seem to be working either. [10:40] >> Dot dot dot button. To activate, press [10:42] enter. [10:42] >> Well, I was just here. [10:44] >> Box 12 p.m. To change the selection, use [10:47] the arrow key. 10 night 7 a.m. [10:50] >> Let's leave at 7 a.m. [10:51] >> Yes. Radio button checked to change [10:54] this. [10:54] >> What am I saying yes to? I'm confused. [10:58] Maybe it's a sign I shouldn't be going [10:59] to Mars. [11:03] >> Okay, so this is about travel, but it's [11:05] also about work. It's about your [11:08] studying. It's about your playing. It's [11:10] about everything. And you're going to be [11:13] actually the same thing is true for AI [11:15] agents. So [11:17] what do we need to do as a group here? [11:19] We need to fight fire with fire. If AI [11:23] creates inaccessible code at scale, we [11:26] must be part of the solution. Don't you [11:28] think? And we can use AI agents to catch [11:31] and fix accessibility defects in real [11:34] time in at the same speed as we are [11:37] creating code. And the only way to close [11:40] this gap is to do that. And we're going [11:43] to have Dylan show you that now. So [11:45] Dylan, [11:47] >> thank you. [11:50] >> Thank you. Thank you, PY. Can I have the [11:52] clicker? [11:56] >> Thanks. [11:59] Thank you, PY. So, oh, this is a [12:01] terrible slide. Um, can you see that? [12:05] All right. So, um, all of you probably [12:07] used AI coding agents. Uh, this is what [12:10] a typical, uh, workflow looks like. You [12:13] have a user working with a co AI agent [12:15] that's embedded into an IDE like VS Code [12:19] for example and the user gives the AI [12:21] agent some instructions to create some [12:23] code. Uh the agent goes away creates [12:25] that shows it up in the IDE and there's [12:27] a sort of uh loop that that you go [12:29] through in that. So uh what we do at DQ [12:32] is we with the axe platform we add some [12:34] capabilities into that. So the X [12:36] platform has two main things that it [12:38] does. The ones is the one uh uh [12:41] capability is knowledge. We have the [12:43] highest quality data and knowledge in [12:45] the industry. Uh all the uh certified [12:48] accessibility professionals in the world [12:50] take our courses in preparation for the [12:52] certification exam. And in fact, our AX [12:56] assistant uh outperforms GT GPT when [12:59] answering questions for that [13:00] certification exam. And we make that [13:02] available through the model context [13:04] protocol to your AI agent. So your AI [13:06] agent can ask a question like um what [13:09] are the requirements for for [13:10] implementing this form in an accessible [13:12] way and can get those requirements back [13:14] and then use those requirements when [13:16] implementing the code. So that's one of [13:18] the ways that we uh make that available [13:21] uh to you. The second way is AI agents [13:23] are great at generating code but they're [13:25] not very consistent. So if you ask an AI [13:28] agent to do something uh twice on the [13:31] same day, it'll do it in two different [13:33] ways. Um, ju this just this weekend I [13:35] was trying out GPT5 trying to get it to [13:38] create a clone of Pac-Man for me. So [13:40] doing some vibe coding with that and uh [13:42] I was struck by the inconsistency [13:44] because every time I would ask it to do [13:46] some changes um it would often not be [13:49] able to actually edit the code and it [13:51] would have to replace all the code. And [13:53] when it did that uh a lot of the things [13:55] that were good about it or the things I [13:57] liked about the previous version were [13:59] actually changed and the the the screen [14:02] looked different. the little ghosts look [14:05] different, didn't operate correctly, and [14:07] I actually gave up on it. But, um, [14:09] that's that's been getting better and [14:11] better, but it's still something that [14:12] we're going to have to deal with. [14:13] Nobody's going to ship code that hasn't [14:15] been tested, right? Who's going to ship [14:17] production code that hasn't been tested? [14:19] So, the second thing that we provide as [14:21] part of the AX platform is consistent [14:23] testing uh that you can apply in sort of [14:26] a code, test, fix, loop, and that's the [14:29] testing capability here of the AX [14:31] platform. And so this is how we envision [14:34] uh compliance accessibility compliance [14:37] at the speed of AI. [14:39] So let me show you what this looks like. [14:41] I took a little video of myself uh in VS [14:45] Code using Copilot in agent mode. And [14:47] you have the code on the side. You have [14:49] the local host uh instance of my [14:52] application running in the uh terminal [14:54] here. You have I have some basic copilot [14:56] instructions. And I'm going to ask it to [14:58] analyze my localhost 3033 for [15:01] accessibility and fix any issues that it [15:03] finds. So the first thing it's going to [15:05] do is look for any tools that are [15:07] available through MCP and other other [15:09] things and instructions to figure out [15:11] how to do this. It found the our MCP [15:13] tool and it asked for permission to to [15:15] run that. And now it's analyzing that [15:18] application running on my local machine. [15:20] It found three different types of [15:22] issues. It found color contrast issues. [15:24] It found missing alt text uh for images [15:26] and links that have no accessible name. [15:29] So now those are pretty uh uh uh [15:32] critical issues for a screen reader [15:34] user. They wouldn't be able to see those [15:35] links. They wouldn't know what they are. [15:37] They wouldn't be able to see those [15:38] images and understand what they are. So [15:40] now given that um it's asked for it also [15:43] receives information back from us on how [15:44] to fix that. And now it's going and [15:46] actually applying the AI agent is [15:48] looking at the code and trying to locate [15:49] that because one of those is a is a [15:52] color contrast issue. It knows that it's [15:54] got to look in the CSS and it's going to [15:56] look to the CSS file. It's also found uh [15:59] the image problem inside a recipe card [16:01] component. Even though that recipe card [16:03] occurred eight times on that page, it's [16:06] determined that that's in that one [16:07] recipe card. And then the the third [16:09] thing it's doing is those links without [16:11] a name were in the footer for LinkedIn [16:13] and and GitHub and it's found the [16:15] component, the sort of footer component [16:17] that's responsible for that, and it's [16:19] analyzing that code right now. So pretty [16:21] soon you'll see it pop up some [16:23] suggestions here. Uh the the rest of [16:26] this is really just the agent uh doing [16:28] its work applying the changes to the [16:30] actual code. So here it picks a slightly [16:33] darker background color. Uh the the the [16:35] text is white on that button. And so in [16:38] order to get a higher contrast, it's [16:39] applying a darker background color to [16:41] that. And obviously it's using uh this [16:43] not in a vibe coding mode where it's [16:45] just applying the fixes but it's [16:46] actually asking me to uh uh accept them [16:49] or change them. The second thing it's [16:52] done is it's found that that recipe [16:54] actually has a name and instead of [16:56] applying sort of a hardcoded [16:59] alternative text to that image, it's [17:00] applying that recipe name from the code [17:03] as a as a good old text for that image. [17:05] And then the third thing it's doing uh [17:07] because those are icon links one [17:09] pointing to GitHub and one pointing to [17:11] LinkedIn it knows that it needs to apply [17:13] a non vis visual uh alternative text to [17:16] that. So it's chosen area label as a [17:19] technique for applying accessible name [17:20] without changing the layout of the page. [17:23] And now after that it's asking for [17:25] permission to validate this using the [17:27] MCP server. Once again I'm going to give [17:29] it permission to do that. Uh and then [17:32] it'll validate that. And that's really [17:33] an example of how you can do the coding, [17:36] the testing, the the fixing of that and [17:39] the validation of that. And then it's [17:41] sort of in a tight loop uh where you're [17:43] using the AI agent and its uh strength [17:45] of understanding what the user wants to [17:47] do, understanding the code, but you have [17:49] the security of a uh a trustworthy [17:53] consistent testing and remediation tool [17:56] at the back end which will give you the [17:58] results, the guaranteed results that you [18:00] that you want. [18:02] So this is what we call accessibility to [18:04] the speed of AI. Uh you give the AI [18:07] agent some instructions to do what you [18:09] want it to do. It asks the axe platform [18:11] for help either through requirements or [18:13] through understanding how to fix it or [18:15] through testing. Uh the AI agent takes [18:18] that information and applies it to the [18:19] code. Uh and then the developer can of [18:22] course decide what to do with that [18:23] whether to accept that or uh edit it or [18:26] reject that particular change. [18:29] So, if you want to talk about this more, [18:31] uh we're at uh booth 149 in the uh [18:35] exhibit hall. Uh you can also take that [18:37] QR code if you want to uh sign up for [18:40] the wait list. We have a wait list for [18:41] the AXMCP product right now or if you [18:43] want a demo from our salespeople. And I [18:46] think we have one minute 30 seconds for [18:49] if there's any question. Anybody has a [18:51] question? [18:56] >> No questions. What? Okay. Well, thank [18:58] you very much for your time. [18:59] >> Yeah. Thank you. Thank you. Thank you so [19:02] much.