Accessibility Deadlines Turned Alt Text Into a Machine Workload

Accessibility requirements for digital products — websites, apps, e-books, ticketing, banking, e-commerce — used to be an aspiration with a standards document behind it. In several major markets they now carry enforcement dates, complaint mechanisms and penalties, and they reach private companies rather than only public bodies. The scale of the backlog is the problem: an established retailer might have hundreds of thousands of product images without descriptions, years of video without captions, and a document archive that a screen reader cannot navigate. Nobody is going to write those descriptions by hand, and models that describe images and transcribe speech are the obvious tool. In practice the useful pattern has proved to be generation plus review rather than generation alone. A model will describe a photograph accurately and yet miss the point of it — the caption on a product image needs to say what a sighted shopper would use it for, not that it contains a grey object on a white background — and errors in captions for spoken content have a way of inverting meaning exactly where precision matters. The organisations doing this well have built triage: automated for the bulk, human review for the pages that carry legal or purchase decisions, and a feedback loop from users who rely on the result. There is a second-order effect that is easy to miss. Structured, well-described content is exactly what search engines, comparison services and, increasingly, AI shopping agents consume. Work that was budgeted as compliance turns out to improve how the catalogue is understood by every machine reader, not only the assistive ones, and the firms that treated the deadline as a cost are finding it was also a data-quality project they had been putting off. The caution is the same as for any compliance-driven automation: the regulation asks for usable accessibility, not for the presence of an alt attribute, and a description generated once and never checked can fail the person it was meant to help while passing the audit. Generate at scale, by all means; sample and read what was generated, always.