Alt Text, Captions, and Transcripts: Accessibility Work AI Can Actually Finish

Why the Deadline Pressure Is Real for US Nonprofits

Accessibility is first a matter of serving people well. Donors, volunteers, clients, and community members who are blind, Deaf, hard of hearing, or who have cognitive or motor disabilities cannot engage with content they cannot perceive. That reason alone justifies the work. Still, the legal context has shifted in ways that affect many nonprofits directly, and it is worth understanding which rules might apply to you. What follows is general orientation, not legal advice, and an organization with real exposure should talk to counsel.

ADA Title III and the general duty of effective communication

Title III of the Americans with Disabilities Act covers places of public accommodation, which can include nonprofits that operate programs, venues, or services open to the public. Title III has no specific technical rule for websites. However, the Department of Justice's guidance on web accessibility and the ADA states that businesses open to the public must make their online offerings accessible, and it specifically lists text alternatives for images and captions for videos among the practices to consider. Courts have reached different conclusions about how Title III applies to websites, which is part of why legal advice matters, but the practical expectation is clear enough: if people use your website to learn about or access your services, it should be usable by people with disabilities.

ADA Title II and the 2024 web rule

Title II covers state and local governments, not private nonprofits. Yet in April 2024 the Department of Justice finalized a rule requiring public entities to make their web content and mobile apps conform to WCAG 2.1 Level AA, and that rule reaches content provided on a public entity's behalf through contracts, licenses, or other arrangements. If your nonprofit runs a program, publishes materials, or hosts content for a city, county, school district, or state agency, you should expect that government partner to ask whether your deliverables conform. In April 2026, the DOJ issued an interim final rule that extended the compliance dates by one year, to April 26, 2027 for entities serving populations of 50,000 or more and April 26, 2028 for smaller entities and special districts. The technical standard did not change. Government partners are now actively inventorying content, and contract language is following.

Section 504, HHS funding, and Section 508

Section 504 of the Rehabilitation Act prohibits disability discrimination by recipients of federal financial assistance, which includes many nonprofits with federal grants. In 2024, the Department of Health and Human Services updated its Section 504 rule to adopt WCAG 2.1 AA for the websites and apps of its recipients. HHS later extended those compliance dates to May 11, 2027 for recipients with 15 or more employees and May 10, 2028 for smaller recipients, again without changing the substance. Health and human services nonprofits receiving HHS funds, directly or as subrecipients, should treat this as a concrete requirement.

Section 508 is often mentioned in the same breath, and it is worth being precise. Section 508 applies to the information and communication technology of federal agencies themselves. Nonprofits usually encounter it indirectly: when they build content or deliver products for a federal agency under contract, when a grant agreement references Section 508 standards, or when a state has adopted 508-based requirements for its own vendors and grantees. If your grant or contract documents mention 508, read the clause carefully, because it may obligate deliverables such as reports, training videos, and toolkits to meet that standard.

Captions: From Auto-Generated to Actually Accurate

Nearly every video platform a nonprofit uses now generates automatic captions. YouTube, Vimeo, Zoom, Microsoft Teams, Instagram, TikTok, and most video editors offer them, and dedicated tools can produce caption files from any recording. This is a genuine change. A decade ago, captioning a library of videos meant paying a service per minute or typing captions by hand. Today, the first draft is free and nearly instant. The catch is that automatic captions are a draft, and a draft with errors in exactly the places that matter most.

Where automatic captions go wrong

Speech recognition handles common words spoken clearly by a single person in a quiet room very well. It struggles with proper nouns, including your organization's name, program names, partner names, and the names of the people speaking. It stumbles over sector jargon and acronyms, so a mention of WIOA, SNAP, or a local coalition becomes a string of plausible but wrong words. Accents, overlapping speakers, background music, outdoor wind, and poor microphones all reduce accuracy. Automatic captions also tend to omit punctuation, speaker identification, and sound information such as [applause] or [music], which captions are supposed to include for viewers who cannot hear the audio. University accessibility offices like Purdue's publish step-by-step guides for correcting these issues inside YouTube Studio, and the process is similar on other platforms.

A useful quality benchmark comes from television. The Federal Communications Commission's caption quality standards, adopted in 2014 for TV programming, define four dimensions: accuracy, synchronicity, completeness, and placement. Those rules do not apply to a nonprofit's YouTube channel, but they are a sensible checklist. Accurate captions match what was said, including speaker names and meaningful sounds. Synchronous captions appear when the words are spoken and stay on screen long enough to read. Complete captions run from the first second to the last. Well-placed captions do not cover faces, on-screen names, or important text.

A practical editing workflow

The fastest way to edit captions is to fix the vocabulary before you fix anything else. Many transcription tools let you add a custom vocabulary or glossary, so feed them your organization's name, program names, staff and board names, partner organizations, local place names, and common acronyms. Then generate captions. Next, export the transcript text and ask a general AI assistant to flag likely errors, giving it your glossary and a short description of the video. It will often catch "Hope House" rendered as "hopes house" or an acronym spelled out phonetically. Finally, a person watches the video at normal or slightly faster speed with captions on, correcting what is left, adding speaker labels and sound tags, and checking timing. For a ten-minute video with a single clear speaker, this can take less time than the video itself. Panel discussions and field recordings take considerably longer.

Sidecar files versus burned-in captions

Captions come in two forms, and nonprofits usually need both. Closed captions are delivered as a separate sidecar file, typically SRT or WebVTT, that the viewer can turn on or off, resize, and restyle. They are searchable, they can be translated, and they are what accessibility standards generally expect on your website and YouTube channel. Open or burned-in captions are rendered into the video image itself. They cannot be turned off or adjusted, but they display everywhere, including social feeds where videos autoplay without sound and where many viewers never enable captions.

A sensible rule is to keep a corrected sidecar file as the master for every video, upload it wherever the platform supports it, and burn in captions for short social clips where silent autoplay is the norm. If you burn in captions, edit them first, because a burned-in error cannot be fixed without re-exporting the video. Keep burned-in text large, high contrast, and clear of the bottom area where platform buttons and descriptions overlay the frame. Our guide to video content on a budget covers more of the production side.

Do not forget audio description

Captions serve people who cannot hear the audio. People who cannot see the video need something different: a description of important visual information that the soundtrack does not convey. A fundraising video with on-screen statistics and no narration, or a program video showing a process without explaining it aloud, fails blind viewers even if it is perfectly captioned. WCAG 2.1 AA requires audio description for prerecorded video in these cases. The simplest fix is prevention: write scripts where the narrator says what matters, so that the video works with eyes closed. For existing videos, AI can help draft description text by working through a transcript and a shot list, and some tools can generate synthetic voice descriptions, but a person should verify that every description is accurate and well timed.

PDFs and Documents: Remediate, Replace, or Retire

Documents are usually the largest and most intimidating part of the backlog. Annual reports, program flyers, application forms, policy documents, research briefs, and board materials accumulate over years, and many were exported from design software with no thought to accessibility. An accessible PDF has tags that identify headings, paragraphs, lists, and tables, a logical reading order, alt text for images, a document title and language setting, and properly labeled form fields. An inaccessible one may be a flat image of text, or text in a jumbled order that a screen reader reads as nonsense.

Triage before you remediate

Not every document deserves the same effort. Start by sorting your files into three groups. Some documents should be retired: outdated flyers, superseded policies, and duplicates that nobody needs. Remove them or move them to a clearly labeled archive. Some documents should be replaced: content that changes often, short documents that are mostly text, and anything people need to read on a phone works better as a regular web page, which is easier to make accessible than a PDF. The rest should be remediated: documents that must remain PDFs, such as signed policies, formal reports, and files partners require in that format.

The DOJ Title II rule includes limited exceptions for archived web content and for certain documents posted before an entity's compliance date, but the details are narrow, the exceptions disappear for documents people use to apply for or participate in services, and they apply to public entities rather than directly to private nonprofits. Do not assume your old files are exempt. Application forms, eligibility guides, and anything a client needs in order to get help should be first in line regardless of age.

What AI tagging can and cannot do

Adobe Acrobat Pro and several other tools now use AI to auto-tag PDFs, identifying headings, lists, tables, and multi-column reading order with much better results than older rule-based tagging. Even Adobe's own developer documentation notes that its Auto-Tag API will not instantly make documents compliant on its own and works best combined with human review. In practice, auto-tagging gets simple text documents most of the way there. Complex tables, nested lists, forms, footnotes, and design-heavy layouts usually need manual correction in a tag editor.

The more durable fix is upstream. Most PDFs are exported from Word, Google Docs, InDesign, or Canva, and the accessibility of the PDF depends heavily on the source file. Using real heading styles instead of bold text, real lists instead of typed dashes, simple table structures with header rows, and alt text in the source document produces a far better PDF with little extra effort. AI can help here by reviewing a draft and pointing out where visual formatting is standing in for structure. If your forms are a major part of the backlog, our article on plain-language forms with AI pairs well with this work, since a clearer form is also easier to make accessible.

A Backlog Workflow for a Small Communications Team

The work becomes manageable when you stop treating it as one enormous project and start treating it as two streams: clearing the existing backlog in priority order, and making sure new content never joins it. The second stream matters more in the long run. A team that fixes 400 old images while uploading 50 new unlabeled ones each month is running in place.

Inventory and prioritize

Start with a simple spreadsheet. List your top pages by traffic, your videos by views, your podcast episodes, and your documents, along with whether each has alt text, captions, a transcript, or tags. Scanners and your analytics can fill in much of this, and an AI assistant can help organize the export into a working list. Then prioritize by impact. Content people need in order to get services, apply for programs, register for events, or donate comes first. High-traffic pages and recent videos come second. Deliverables tied to government contracts or federal grants deserve priority too, given the compliance dates above. Old content with little traffic comes last, and some of it can simply be retired.

Batch the work by type

Grouping similar tasks makes AI far more efficient. Spend one session sorting and drafting alt text for an entire section of the site, with the page context loaded into each prompt. Spend another loading your glossary into a transcription tool and generating caption drafts for ten videos, then review them over the following days. Handle documents in their own batches, starting with the retire and replace decisions, since those reduce the remediation pile before you touch a tag editor. Batching also helps you notice recurring AI errors, such as a program name it always gets wrong, and fix them at the source.

Build accessibility into publishing

For new content, make accessibility part of the definition of done. No image is uploaded without an alt text decision. No video is published without corrected captions. No podcast episode goes live without a transcript. No PDF is posted without a check. Written down as a short publishing checklist, these habits add minutes per item rather than weeks of backlog later. Assign a named owner for each content type, and if you work with volunteers or AI champions on your team, our piece on building AI champions discusses how to spread these skills beyond one person.

Knowing When to Bring in Outside Help

AI and an organized team can handle most of the routine backlog. Some work still benefits from specialists. Complex PDFs such as financial statements, long research reports with many tables, and fillable forms can consume days of staff time in a tag editor and may be more economical to send to a remediation vendor. Live captioning for public events, audio description for flagship videos, and translation of captions into other languages often call for professional services. If a government partner or federal funder asks you to certify conformance with WCAG 2.1 AA, an independent review is worth the cost.

When you do hire help, AI still changes the economics. Vendors can work from your cleaned transcripts and glossaries, and your team can handle the simple items so that paid hours go to the hard ones. Ask vendors how they use AI themselves and what human review they apply, since the same standards you hold your own work to should apply to theirs.