A polished AI-narrated video sits in the final review queue, two days before a federal-contractor deadline. The synthetic voice is flawless, the visuals are clean, the script was verified against the policy. Then the Section 508 reviewer opens the auto-generated captions and reads them aloud: the narrator says "do not bypass the interlock," and the caption says "do bypass the interlock." One dropped word, one reversed safety instruction, one course that cannot ship. The reviewer is not being difficult. Accessibility is a legal gate, and this course just failed it on a single line a machine wrote and nobody read. This lesson is about why that gate exists, why AI does not lower it, and how to build so you never meet it by surprise.

The Standard Is Not Optional, and AI Does Not Lower It

The conformance target for digital learning is WCAG 2.2 AA, the Web Content Accessibility Guidelines version 2.2 at the AA level, published as a W3C Recommendation on 5 October 2023. Why you care: it is the line a learning experience has to clear to be usable by people with disabilities, and in many contexts, including US federal and federal-contractor work under Section 508, it is a legal requirement, not a best practice. Section 508 incorporates WCAG by reference, which means the standard is written into procurement and compliance obligations; the version legally incorporated can lag the latest WCAG, so you confirm which version applies to your context rather than assuming the newest. The crucial point for an AI-assisted build is this: the standard attaches to the experience, not to how the experience was made. An AI-narrated video has exactly the same obligations as a human-narrated one. The model's involvement neither adds an exemption nor lowers the bar by a single success criterion.

State the bright line plainly, because the whole lesson hangs on it. An AI-generated experience that fails WCAG 2.2 AA does not ship. Accessibility is a gate, not a polish step, full stop. This is not a softer rule than the iron rule of the program; it is the iron rule applied to access. "The AI generated the captions" is no more a defense to an accessibility auditor than "the AI wrote it" is to a compliance officer. The human owns the conformance, the same way the human owns the content and the assessment.

Accessibility is not the last thing you check before launch. It is a gate the course passes through, and "the AI made the captions" is not a defense to the person standing at the gate.

Why Retrofitting Is the Expensive Mistake

The most common and most costly accessibility error in an AI-assisted workflow is treating access as a finishing task: build the slick video first, then bolt on captions, alt text, and a transcript at the end. Retrofitting is expensive for a structural reason. When you build media without accessibility in mind, you make choices that are hard to reverse, a visual that only works in color, an interaction that only works with a mouse, a key point conveyed only by an image with no text equivalent, a fast cut with no time for a caption to be read. Fixing those at the end means re-rendering, re-scripting, and sometimes rebuilding. Building them in from the first draft costs a fraction, because the script, the visuals, and the interactions are designed to carry their access from the start.

AI changes the economics in a way that cuts both directions. It can generate a first draft of captions, alt text, and a transcript in seconds, which is a genuine gift, the blank-page problem solved. But it generates them at the same speed it generates everything else, which means it generates accessibility errors at the same speed too, and a team moving fast is tempted to accept the AI's access artifacts unread, exactly as they were tempted to accept the AI's content unread. The discipline is identical: AI drafts the captions, the human verifies the captions, and the human owns whether they are correct and conformant. Accessibility built in from the first draft is fast and cheap. Accessibility retrofitted at the end is slow and expensive, and accessibility skipped is a course that does not ship.

The Three Artifacts, and What Each One Actually Requires

Three access artifacts carry most of the load for AI-generated media: captions, alt text, and transcripts. Each is a distinct obligation serving a distinct learner, and each has an AI failure mode you verify against.

Captions: The Auto-Caption Error-Rate Problem

Captions are the synchronized text of everything spoken and every meaningful sound, on a track the learner can turn on. Why you care: for a deaf learner, the caption is the content, and an error in the caption is an error in the lesson. Automatic speech recognition has improved enormously, but its accuracy is uneven and context-dependent: it stumbles on technical jargon, proper nouns, acronyms, accented speech, overlapping audio, and exactly the specialized vocabulary that fills compliance and safety training. The danger is not the obvious garble a human notices; it is the plausible substitution, "interlock" heard as "interlocks," "do not" collapsing to "do," a drug name rendered as a near-homophone, that reads as fluent and reverses the meaning. The error rate is a number to verify in your own context, not a vendor promise to trust, and the only safe assumption is that AI captions are a draft to be read against the verified script word for word before they ship. A single reversed safety instruction in a caption is the lockout/tagout failure of the accessibility world.

Alt Text: Meaning, Not Pixels

Alt text is the short text equivalent a screen reader announces in place of an image, so a blind learner receives what the image teaches. Why you care: a diagram that carries a load-bearing point is invisible to a screen-reader user unless its meaning is in text. The AI failure mode is two-sided. AI can generate alt text quickly, but it tends to describe the picture literally ("a person at a desk") rather than its instructional meaning ("the analyst checks the record's sensitivity level before saving"), and for a decision diagram it may miss the very relationship the diagram exists to show. Worse, a purely decorative image should have empty alt text so a screen reader skips it, and AI may dutifully describe decoration that adds noise, not meaning. Good alt text answers "what does this image teach here," which requires knowing the objective, the same human judgment signaling required in the Mayer lesson. The AI drafts; you decide what the image means in this context and whether it means anything at all.

Transcripts: The Full Accessible Record

A transcript is a complete text version of the media: the narration plus any meaningful information that appeared only on screen, including descriptions of important visuals. Why you care: it serves screen-reader users, learners who absorb better by reading, learners on poor connections, and anyone searching the content later, and it is often the most efficient accessibility artifact to produce well because it doubles as a study aid and a searchable record. The AI failure mode is the silent omission: a generated transcript captures the spoken words perfectly and quietly drops the information that was only ever on screen, the figure in the diagram, the label on the warning, the step shown but not narrated, leaving a blind learner with a transcript that is fluent and incomplete. Verifying a transcript means checking not just that the words are right but that nothing visual-only was lost.

Building It In: A First-Draft Workflow

Building accessibility in rather than bolting it on is a workflow, not a virtue. The shift is to treat the access artifacts as part of the draft, generated and verified alongside the content, not as a separate late phase. Concretely, that means a few habits.

• Script for access first. Write narration that does not depend on an unspoken visual ("as you can see here"), so the transcript and a sound-off learner both receive the point. The script is where accessibility is cheapest to build in.
• Generate captions, alt text, and transcript with the draft, then verify each against the source. Read captions against the verified script word for word. Rewrite alt text to the instructional meaning. Check the transcript for visual-only omissions. Treat every AI access artifact as a draft, never a deliverable.
• Verify the experience, not just the artifacts. WCAG 2.2 AA is more than captions and alt text: color is not the only way meaning is carried, contrast is sufficient, the content works by keyboard, motion can be paused, captions have time to be read. AI-generated interactions and visuals can fail these silently.
• Record conformance. Keep an accessibility conformance record, the artifact, sometimes formalized as a VPAT or ACR, that lets an auditor see what was checked and by whom. This is the access counterpart to the SME sign-off log: it answers "who verified accessibility before this shipped."

A definition for the record: a VPAT (Voluntary Product Accessibility Template), whose completed form is an ACR (Accessibility Conformance Report), documents how a product meets accessibility standards. Why you care: in procurement and audit, "we made it accessible" is a claim; the ACR is the evidence, and an auditor reads the evidence.

Two cautions keep this workflow honest. First, an automated accessibility checker, the kind built into an authoring tool, is necessary but not sufficient. It catches structural problems, a missing alt attribute, a heading out of order, insufficient contrast it can measure, but it cannot tell you whether the caption matches what was actually said or whether the alt text conveys the meaning the diagram exists to teach. Those are judgments about meaning, and meaning is exactly what a machine cannot verify for you. A green checker result on an unread caption track is a false sense of safety. Second, accessibility is not only a video problem. The same discipline applies to an AI-drafted interactive, a generated knowledge check, a chat-based tutor: can a screen-reader user complete the interaction, is the question operable by keyboard, does the tutor's output reach a learner who cannot see the screen. AI can generate an inaccessible interaction as easily as an inaccessible video, and the gate is the same.

There is also a quiet alignment between accessibility and good instructional design that makes building it in less of a tax than it first appears. Scripting narration that stands on its own, the access habit, also produces clearer explanations for every learner. A transcript doubles as a searchable study aid and a revision record. Captions help the learner in the open-plan office with the sound off, not only the deaf learner. Alt text that names the instructional meaning of an image forces you to articulate why the image is there at all, which often reveals that a decorative image should simply be cut, the coherence principle and the accessibility requirement arriving at the same answer from two directions. Accessibility built in from the first draft is not a separate burden bolted onto the work; done well, it is part of what makes the work good.

A Worked Example: Before and After

Return to the AI-narrated video in the review queue, and watch the two paths that led there.

Before (accessibility as the finish line). The team built a polished synthetic-narrated video fast, verified the script against the policy, and felt finished. Captions were auto-generated and accepted unread. Alt text on the key decision diagram read "a flowchart," conveying none of the decision logic to a screen-reader user. No transcript was made, because the video felt self-contained. With two days to the deadline, the 508 reviewer found the reversed "do not bypass the interlock" caption, the meaningless alt text, and the missing transcript. The course failed the gate. The fix, under deadline pressure, meant re-reading every caption, rewriting alt text, building a transcript, and re-rendering, at a cost and stress far beyond what building it in would have taken.

After (accessibility from the first draft). The same team scripted for access, avoiding "as you see here," so the narration carried its own meaning. Captions were generated and then read against the verified script line by line, catching the reversed interlock instruction before render. Alt text on the decision diagram was rewritten to its instructional meaning, the actual classification rule the diagram teaches. A transcript was generated and checked for visual-only omissions. The experience was verified for contrast, keyboard operation, and caption timing, and an accessibility conformance record was attached naming who checked what. When the 508 reviewer opened the course, every artifact passed, and the conformance record answered the "who verified this" question before it was asked. Same tool, same speed, completely different fate, because accessibility was the first draft, not the last check.

The lesson is not that AI cannot help with accessibility; it generates the first drafts of captions, alt text, and transcripts genuinely fast, and that is a real gift. The lesson is that AI generates accessibility errors at the same speed it generates everything, that the standard does not bend for a machine, and that the human owns the conformance. Build access in from the first draft, verify every artifact against the source, keep the record, and the gate becomes a formality you walk through instead of a wall you hit two days before launch.

Key Takeaways

• WCAG 2.2 AA (W3C Recommendation, 5 October 2023) is the conformance target; Section 508 incorporates WCAG by reference and the legally incorporated version can lag the latest, so confirm which version applies to your context.
• The standard attaches to the experience, not to how it was made: an AI-narrated video has exactly the same obligations as a human-narrated one, with no exemption and no lowered bar.
• The bright line: an AI-generated experience that fails WCAG 2.2 AA does not ship; accessibility is a gate, not a polish step, full stop, and "the AI made the captions" is not a defense.
• Retrofitting access at the end is structurally expensive because it forces re-rendering and rebuilding; building it in from the first draft costs a fraction because the script and visuals carry their access from the start.
• Captions, alt text, and transcripts each serve a distinct learner and each have an AI failure mode: auto-caption errors that reverse meaning, alt text that describes pixels instead of instructional meaning, and transcripts that silently omit visual-only information.
• Auto-caption accuracy is uneven and context-dependent, stumbling on jargon, proper nouns, acronyms, and accented speech; the error rate is a number to verify in your context, and a reversed safety instruction in a caption is the lockout/tagout failure of accessibility.
• Good alt text answers "what does this image teach here," which requires knowing the objective; AI drafts it, but a human decides the meaning and whether a decorative image should have empty alt text instead.
• Keep an accessibility conformance record (sometimes a VPAT or ACR), the access counterpart to the SME sign-off log, so an auditor can see what was verified and by whom, because the human owns the conformance.