Module 4, Nov 11, Seminar and lab. The Digital Story is due next week.

Production and accessibility

Background

The Digital Story is due next week, and today is the last supervised build time. Holding a video back until it is fully accessible has a real cost: auto-generated captions run roughly 90 percent accurate, cost nothing, and reach a viewer that a video delayed for a perfect caption pass never reaches at all, and a systematic review of assistive technology in education finds real, measurable inclusion gains once tools are actually used, imperfect or not. The same evidence base carries a warning about what never gets measured: a systematic review of 11 randomized trials of digital health interventions for children in underserved and rural settings found improved outcomes alongside a near-total absence of cost-effectiveness or equity-disaggregated analysis.

An independent benchmark of eleven speech-recognition services found word-error rates swinging widely by service and audio, from a few percent to around one word in five, and the people who depend on captions most are the ones the worst of that range hits hardest, so an uncorrected caption track can fail its own audience while looking finished.

The state of the wider web sets the base rate. The 2026 WebAIM Million, the eighth annual automated scan of the home pages of the top 1,000,000 websites, found detectable WCAG 2 failures on 95.9 percent of them, averaging 56.1 errors per page, up 10.1 percent from 2025 and reversing several years of slow improvement. Low-contrast text is the single most common failure, appearing on 83.9 percent of pages.

The promise that AI closes this gap automatically has already met a regulator. In January 2025 the FTC ordered accessiBe, an AI accessibility-overlay vendor, to pay $1 million for deceptive claims that its product could make any website WCAG-compliant.

Key ideas

The legal floor

The ADA, Section 508, and WCAG make media accessibility concrete: synchronized captions are a Level A requirement for prerecorded video, audio description of meaningful visuals is Level AA, and a full transcript is the text alternative. Captions mean verbatim speech, speaker identification, and meaningful non-speech sound. The federal guidance goes down to interface placement: caption and audio-description controls must sit at the same menu level as the volume control.

Caption quality is regulated separately from caption presence. The FCC’s four legally binding quality standards, accuracy, synchronicity, program completeness, and placement, have been in force since January 2015, and the National Association of the Deaf traces them to a petition it led in 2004. A caption track can exist and still fail every one of them.

The floor is enforced in court. 3,948 ADA website lawsuits were filed in 2025, up 23.8 percent from 2024, with four states accounting for 86.65 percent of filings. The beneficiaries of compliance go far beyond Deaf and hard-of-hearing viewers, taking in sound-off phone viewing, non-native speakers, noisy rooms, and low bandwidth.

Language has a floor of its own. The Plain Writing Act of 2010 requires federal public-facing content to be written for its audience, and the Robert Wood Johnson Foundation makes the disability case with a working example: New Disabled South’s plain-language dashboard, which translates bills across 14 Southern states, has drawn more than 3,000 monthly users since its November 2023 launch, and a $100,000 Georgia plain-language campaign drove enrollment in Medicaid home- and community-based-services waivers.

The standards become usable once they turn into checkable items. The A11Y Project’s checklist ties each item to a WCAG success criterion: every image needs an alt attribute, a page gets one top-level heading, and normal text needs a 4.5:1 contrast ratio at Level AA, with 3:1 for large text, the thresholds the WebAIM Contrast Checker tests from a pasted pair of colors.

Even the standards vocabulary stays unsettled. A 2025 follow-up to a 2003 paper now cited about 1,300 times found accessibility, usability, and universal design still used inconsistently across research and practice, and proposes treating the first two as the operational tools that put the third into effect.

Automatic captions

An independent benchmark of eleven commercial speech-recognition services tested them on real higher-education lecture recordings instead of clean lab speech, because vendors claim human parity while Deaf and hard-of-hearing users keep reporting serious accuracy problems in practice. The measured error rates varied too much across services and recordings to trust any single vendor number, and the people who depend on captions are the ones those errors hit hardest.

The standards bodies have already drawn the conclusion. The W3C states that automatically generated captions do not meet accessibility requirements unless they are confirmed to be fully accurate, and Section508.gov flags that auto-captioning does not yet provide an equivalent experience for Deaf and hard-of-hearing viewers. In lab you generate captions, correct them by hand, and count your own errors; that count is what the revote runs on.

Production lanes

The course workflow has two lanes. The easy lane is cloud, text-based editing: Clipchamp (free with your U-M Microsoft account, no watermark at 1080p, auto-captions included) or Vrew, which transcribes your footage and lets you cut by deleting sentences. The offline lane is software you install and control: DaVinci Resolve for editing (free version, no watermark), Audacity for audio, and Subtitle Edit running Whisper offline for captions. One rule sorts you into a lane: footage containing an identifiable vulnerable person stays offline, because cloud editors upload it to someone else’s servers. “Free” also deserves a second look before you commit: CapCut moved auto-captions and SRT export behind a roughly $20-a-month paywall in its 2025 restructure, and Vrew’s free-tier limits change often.

Accessibility overlays

An overlay is a JavaScript widget a site owner pastes in so that a vendor’s software can retrofit accessibility on the fly, and the sales pitch is that one snippet replaces actual remediation. The Overlay Fact Sheet, signed by more than 1,000 people including WCAG and ARIA specification authors and accessibility leads at Google, Microsoft, and Apple, argues that overlays cannot achieve WCAG compliance and often make sites worse for assistive-technology users, because a widget that rewrites a page’s structure collides with the screen reader already interpreting it. A WebAIM practitioner survey the fact sheet cites found 67 percent of accessibility professionals, and 72 percent of respondents with disabilities, rate overlays not at all or not very effective.

Two outside checks agree. The FTC’s January 2025 order against accessiBe found the company’s compliance claims deceptive, and 24.9 percent of the ADA website suits filed in 2025 targeted sites that had already installed a widget, so the product does not reliably deliver even the litigation shield it is sold as. An opinion piece in Communications of the ACM sets the case in a wider argument: AI systems carry ableist bias, from training data that associates disability with negative content to unsubstantiated claims that a product’s AI makes things accessible, and disabled people need direct roles in building, regulating, and deploying these systems.

Imagery ethics

Accessibility decides who can use your video; imagery ethics decides how the people in it appear. SAIH’s social-media guide, which grew out of a 2017 collaboration with the satirical Instagram account Barbie Savior, asks travelers and volunteers to hold the same posting ethics abroad that they would at home: no photographing strangers’ children without consent, and no poverty imagery that reduces a person to a stereotype while centering the poster as a savior.

Stella Young’s 2014 TEDxSydney talk applied that standard inside disability representation, coining “inspiration porn” for images of disabled people doing ordinary things, framed to motivate non-disabled viewers. She called it objectification because it uses one group for the benefit of another.

Whether such imagery even raises money is an empirical question. Two experiments in Business & Society measured how poverty-porn imagery, spokesperson identity, and message framing affect attention and donation intention, one with eye tracking (n=236) and one with a survey (n=667).

The NVDA case

A screen reader turns what’s on screen into speech or braille; for a blind computer user it is the way into the machine. For years the standard tool, JAWS, cost around $1,000. Two blind developers, Mick Curran and Jamie Teh, answered with NVDA, a free, open-source screen reader that now has roughly 250,000 users across 175 countries, in more than 55 languages. In WebAIM’s tenth screen-reader survey (1,539 respondents, fielded December 2023 to January 2024), NVDA had overtaken JAWS as the most commonly used desktop screen reader, 65.6 percent to 60.5 percent, though JAWS still narrowly leads as respondents’ primary tool, 40.5 to 37.7 percent, and VoiceOver dominates mobile at 70.6 percent.

The same analytic standard applies to successes as to harms: whose participation shaped this tool, and would it replicate without a disabled-led development team. The design record behind those questions is older than screen readers. Elise Roy, a disability-rights lawyer who began losing her hearing at age 10, argues that designing for disability first surfaces solutions that outperform designing for the norm, and her evidence includes text messaging, developed for deaf and hard-of-hearing users before it became universal.

Angela Glover Blackwell’s curb-cut history starts with Berkeley disability activists pouring illegal concrete ramps in the early 1970s before the city adopted official curb cuts in 1972, and traces the earliest US installations to Battle Creek and Kalamazoo, Michigan, around 1945. Blackwell adds a caution: “benefits everyone” should never replace “serves the excluded” as the justification. NVDA also appears in the seminar’s cold open, where it reads an image-only title card aloud and produces silence.

Before class

In class

Seminar

Vote, then a cold open on exclusion: a 30-second clip plays with the sound off and no captions, then a screen reader reads an image-only title card, and the room names who just got shut out and whether the maker decided that on purpose or by default. A short mini-lecture covers the legal floor and the NVDA case. The bulk of the seminar belongs to a guest session on inclusive and universal design (speaker confirmed on Canvas before the term); come with one question about your own story’s exclusions.

Lab: troubleshooting clinic and the Access Bug-Bounty

You arrive with the rough cut assigned in Week 10, so lab time goes to fixing and auditing instead of building from zero. The lab opens with a 15-minute troubleshooting clinic on the hardest edit problem each person hit, with novices paired with tool-comfortable classmates at rotating stations (Resolve, Audacity, Canva). Then the accessibility pass on your own draft: generate captions with Subtitle Edit and Whisper (or YouTube Studio on an unlisted upload as the fallback), correct them by hand while counting your errors in the first 30 seconds, and import the corrected .srt into Resolve. The Access Bug-Bounty follows: swap drafts with a peer, hunt for access failures against the checklist, and hand back three prioritized, specific fixes (“0:42, white text on light sky fails contrast” counts; “make it more accessible” doesn’t). Color calls get checked against the WebAIM Contrast Checker, which returns a WCAG pass or fail for any pair. A ten-minute usability walk closes the pass: one partner tries to find your own partner organization’s intake number or its volunteer sign-up on its real website, on a phone or with NVDA running, narrating out loud, while the other only watches and logs every hesitation or failure; the last two minutes are swap and compare. Watching one real person use the site turns up problems a checklist review passes over. The revote follows, checked against your own error count. The completed checklist is submitted with your Digital Story and graded.