Why accessibility text needs a person
The Web Content Accessibility Guidelines, published by the W3C and cited in this page's sources, are the widely used standard for web content, and the site's writing for the web guide summarises what they ask of writers: images need alternative text that says what the image conveys, or are marked decorative when they convey nothing. The guidelines also cover time-based media: video with sound needs captions, and audio and video need text alternatives so the information is available to people who cannot hear or see it. Conformance requirements vary by jurisdiction and sector, so treat the guidelines as the shared reference and check what applies locally.
The key word in that summary is conveys. Good alt text does not describe everything in an image; it carries the information the image adds to the page. The same photograph of a person at a desk needs different alt text on a product page for the desk, on a team page introducing the person and in a blog post about working from home. A model looking at the image alone cannot know which of those pages it is on, unless it is told, and even then it does not know what the reader needs from it.
That is why generated accessibility text is a draft. A model can describe what is visible faster than any person, and it can transcribe speech at a scale no team could match by hand. What it cannot reliably supply is purpose, names, exact figures and the judgement about what matters, and those are exactly the details that a person who depends on the text is missing. The people who rely on alt text and captions are also the people least able to notice when they are wrong.
Alt text: what to describe and what to skip
Start with the question the image answers on this page. A product image answers "what does it look like": colour, material, shape, the variant shown, the angle. A chart answers "what does the data show": the trend or comparison the chart was included to make, with the key figures. A photograph in an article answers whatever the paragraph beside it is about. A decorative flourish answers nothing and takes an empty alt attribute, so that screen readers skip it.
The image SEO guide sets out the same principles from the search side: describe the subject, what is distinctive about it and the context if it matters; leave decorative images empty; never use alt text as a caption or a keyword field. Google's image guidance, cited in this page's sources, says it uses alt text together with the page's content to understand what an image shows, and warns against filling alt attributes with keywords. Alt text written for a person is the alt text that serves SEOImproving visibility in unpaid search results. as well.
| Image type | What the alt text should carry | What to skip |
|---|---|---|
| Product photo | Product, variant, colour, material, the view shown | "Image of", brand slogans, keywords |
| Chart or diagram | The point the figure makes and its key values, or a pointer to a data table | Every data point read out in sequence |
| Image containing text | The text itself, exactly as shown | A description of the font and layout |
| Person in a team or author photo | The person's name and role, as the page uses them | Guesses about age, ethnicity or mood |
| Illustrative photo in an article | What the image shows that relates to the paragraph | Background detail with no bearing on the text |
| Decorative image | Nothing: an empty alt attribute | Any description at all |
Two cases need extra care. Text inside an image, such as a banner headline, a price, a date or an offer, is invisible to screen readers unless it appears in the alt text or on the page, and a model reading it can misread a figure. And charts carry the most information of any image type, which is why the image SEO guide recommends a text equivalent: alt text summarising the point, a caption, or the data in a table on the page.
Captions and transcripts
Captions are synchronised text of what is said in a video, with speaker identification where it is not obvious and the sounds that matter, such as music that sets a mood, a laugh, a doorbell that the plot turns on. Transcripts are the full text of an audio or video item, readable without the media, and for video they work best when they also note the visual information a listener would miss. Both depend on accuracy in the places speech recognition is weakest.
The YouTube video publishing checklist asks for a caption file edited by a person rather than accepted from automatic captions, and lists the reason among its common mistakes: product names, technical terms and accents come out wrong, and the captions are what many viewers read with the sound off. The same applies to captions drafted by any AI tool. Generated captions also tend to miss speaker changes, run sentences together across cuts and leave sounds undescribed, because speech recognition transcribes words rather than meaning.
Transcripts add one more gap. A transcript produced from the audio track alone contains nothing that was only shown: the product on screen, the chart a presenter points at, the text on a slide, the demonstration that replaces an explanation. Where the video carries information visually, the transcript needs that information written in, and that is a job for the person who knows the content.
Review rules for generated accessibility text
A small set of rules covers most of what goes wrong. They work best as a short checklist the reviewer runs on every item, and as instructions given to the model before it drafts.
- Give the model the context. Supply the page text or the product record with the image, and the script, product names and speaker names with the video. A model that knows the purpose describes the right things.
- Check every name. People, products, places and brands are the most common errors and the most damaging ones. Names come from the page or the record, never from the model's reading of a face or a logo.
- Check every number and every word of on-screen text. Prices, dates, sizes, percentages and offers are checked against the source, not against the image.
- Remove guesses about people. Age, ethnicity, gender, disability and emotion are described only when the page needs them and the facts are known, never inferred from appearance.
- Cut what does not convey. Trim descriptions of backgrounds, lighting and composition unless they are the point of the image.
- Check captions against the audio. Listen through at normal speed with the captions on, and fix names, terms, speaker labels, missing sounds and timing.
- Add what was only shown. Write visual-only information into transcripts and, where needed, into the video's description or an audio description.
- Leave decorative images empty. A generated description of a divider or a background texture is noise for a screen-reader user.
The AI for creative guide covers the production and review of images and video themselves; this review is the accessibility layer on top of it, and it belongs in the same approval step so that an asset is not signed off without its text.
Workflow by asset type
The review effort should follow the risk and the reach of each asset type. The table sets out a workflow that a small team can run.
| Asset type | Draft with AI | Who reviews | Review rule | Where it is recorded |
|---|---|---|---|---|
| Product images | Alt text from image plus product record | Catalogue or ecommerce owner | Every product image; variant and colour against the record | Product record, so it regenerates with changes |
| Article and blog images | Alt text from image plus surrounding paragraph | Editor of the piece | Every image, as part of the edit | Content system's alt field |
| Charts and infographics | Summary of the point plus key values | Author of the data | Every figure against the data; table on page where needed | Alt field and page text |
| Social images | Alt text from image plus post copy | Person approving the post | Every post, as part of approval | Platform's alt text field |
| Short social video | Captions from audio | Person approving the post | Read with sound on before publishing | Burned in or uploaded caption file |
| Long-form video | Captions and transcript from audio, with a script or term list supplied | Producer or subject expert | Full listen-through; names and terms against the script | Caption file and transcript on the page |
| Podcasts and audio | Transcript from audio | Producer | Names, terms and figures checked; speaker labels added | Transcript page linked from the episode |
| Webinars and recorded events | Captions and transcript | Organiser with a speaker list | Names, slides and demonstrations written in | Recording page |
The social post publishing checklist includes alternative text for images and captions for video among its pre-publish checks; with AI drafting, the check becomes reviewing a draft rather than writing from nothing, which is what makes full coverage realistic. What a reviewed description or caption should contain on each kind of post, from carousels and quote cards to reposted customer photos, is set out in the guide to accessible social media posts. To see whether the work is holding, track a simple KPIA metric tied directly to an objective. per channel: the share of published images with reviewed alt text and the share of videos with corrected captions, read alongside any accessibility feedback from customers. The marketing accessibility programme playbook folds this review into each channel's publishing step and checks a sample of live work on a regular rhythm.
Clearing an image and video backlog at a ceramics studio
Corrastone Ceramics, an invented studio selling handmade tableware online, finds that 1,340 of its 2,100 product images have no alt text and none of its 46 workshop videos have edited captions. It drafts alt text for every product image from the image plus the product record, and captions for every video from the audio plus a list of 60 glaze, clay and technique names. Review of the alt text takes the catalogue owner about 9 hours across two weeks; 210 drafts are corrected, mostly glaze colours named wrongly and 2 prices read wrongly from a sale banner. Caption review finds the glaze names misheard in 38 of the 46 videos and no speaker labels in the 12 videos with two presenters. Transcripts for the 10 longest workshops get the demonstrated steps written in. Afterwards, every new product image and video passes through the same review before it is published.
Common mistakes
- Publishing generated alt text unreviewed. Wrong names, colours and figures go straight to the people least able to spot them.
- Describing the image instead of its purpose. Alt text carries what the image adds to the page, not an inventory of what is visible.
- Describing decorative images. They take an empty alt attribute so screen readers skip them.
- Keyword-stuffing alt text. It fails the reader, and Google's image guidance advises against it.
- Trusting automatic captions. Product names, technical terms and accents come out wrong, and speakers and sounds go unmarked.
- Transcribing audio only. Anything that was only shown is missing from the transcript until someone writes it in.
- Guessing about people. Appearance-based descriptions of identity or emotion are often wrong and rarely needed.
- Treating accessibility text as a separate project. It belongs in the same approval step as the asset, or the backlog grows back.
