{
  "video": {
    "id": "ZccJsdOE54M",
    "title": "The Cost of Accessibility in AI — Jemima Abu, CAIS",
    "duration": 602,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "ZccJsdOE54M",
    "title": "The Cost of Accessibility in AI — Jemima Abu, CAIS",
    "one_liner": "In a single informal run at the speaker's workplace, an agent asked to implement a Figma design with an accessibility skill loaded cost 58% less and finished in a third of the time, because without guidance AI copies a codebase's inaccessible habits and writes verbose custom code instead of using native semantic elements.",
    "summary": "Jemima Abu argues that building inaccessible products wastes tokens, time and money: a product people can't use is useless, and AI makes accessibility easier, not harder. In a one-off experiment, the same 'implement this design from Figma' prompt run with her accessibility skill was 58% cheaper and took a third of the time. She says this is not empirical proof. Her explanation is that agents without guidance absorb human 'it works for me' bias and write verbose custom components when semantic elements already exist. She closes with a four-step process for getting agents to produce accessible code.",
    "key_points": [
      "WebAIM found that 96% of the homepages of the top 1 million websites have some form of accessibility issue.",
      "Experiment: a simple prompt with no guidelines ('implement this design from Figma') was run twice on her work codebase, once without an accessibility skill and once with her own. With the skill, cost dropped 58% and the run took a third of the time. She says this was a one-off, unconstrained experiment and can't serve as empirical data.",
      "Human bias becomes AI bias: a codebase built with an 'it works for me' mindset teaches the agent that accessibility doesn't matter, and every generated change makes the codebase more inaccessible.",
      "Without guidance, AI defaults to verbose code and reinvents the wheel, for example building an accordion or modal from scratch instead of using an existing semantic element. Recreating native behaviour takes more features, more testing and more time.",
      "Blueberry muffin analogy: adding ARIA labels to a finished site doesn't make it accessible. Accessibility has to be baked in from the start, across business idea, design and QA. Leaving it to the end wastes time on regressive testing and refactoring.",
      "Business and legal cost: 'none of our users rely on assistive technologies' is a self-fulfilling prophecy, like a business with no door. In the US, the Americans with Disabilities Act requires public-facing products to be accessible, or the business faces fines and lawsuits.",
      "Her four-step process: define accessibility requirements for your platform, audit and fix the existing codebase before adding an agent, build a reusable component system or design system, and enforce a mandatory accessibility skill.",
      "A lightweight bare-minimum skill: use semantic HTML wherever possible and ARIA only when semantic HTML can't express the behaviour; make anything mouse-operable keyboard-operable too; label elements properly for screen readers; and have the agent audit against these steps before running the code. Manual testing is still needed."
    ],
    "takeaways": [
      "Load a mandatory accessibility skill for any agent work that generates UI, such as Chrome's Modern Web Guidance skill or a lightweight checklist of your own for vibe-coding and small projects.",
      "Before adding an AI agent to a codebase, audit it for accessibility (with automated tools or the agent itself) and fix existing issues, so the agent doesn't copy them as best practice.",
      "Tell agents to prefer native semantic HTML/CSS over custom JavaScript components and to use ARIA only when semantic HTML can't express the behaviour.",
      "Centralise UI in a reusable component library or design system, so an accessibility fix is made once instead of in every file.",
      "Build accessibility in from the start and at every stage (idea, design, QA), have the agent self-audit before running code, and still do manual testing."
    ],
    "topics": [
      "accessibility",
      "ai-agents",
      "agent-skills",
      "vibe-coding",
      "semantic-html",
      "token-cost",
      "design-systems",
      "prompt-engineering"
    ],
    "tools": [
      "WebAIM",
      "Figma",
      "Modern Web Guidance skill (Chrome)",
      "Google Chrome",
      "Microsoft Teams",
      "Netflix (Old Enough)",
      "ARIA",
      "CAIS"
    ],
    "quotes": [
      {
        "text": "stop wasting tokens on building products that people can't use.",
        "at": "00:05",
        "url": "https://www.youtube.com/watch?v=ZccJsdOE54M&t=5s"
      },
      {
        "text": "there was also a 58% reduction in cost when using that skill.",
        "at": "01:31",
        "url": "https://www.youtube.com/watch?v=ZccJsdOE54M&t=91s"
      },
      {
        "text": "Firstly, human bias becomes AI bias.",
        "at": "03:45",
        "url": "https://www.youtube.com/watch?v=ZccJsdOE54M&t=225s"
      },
      {
        "text": "It's not that your users don't rely on assistive technologies, it's that people with assistive technologies cannot use your website in the first place, so they just never go to it.",
        "at": "06:26",
        "url": "https://www.youtube.com/watch?v=ZccJsdOE54M&t=386s"
      }
    ],
    "words": 2814
  },
  "summary_url": "/#ZccJsdOE54M",
  "transcript": {
    "html": "/transcripts/ZccJsdOE54M.html",
    "txt": "/transcripts/ZccJsdOE54M.txt",
    "vtt": "/transcripts/ZccJsdOE54M.vtt"
  }
}