RiskNet: A large-scale dataset of AI risk incidents from news with alignment and multi-dimensional annotations

A new dataset called RiskNet aims to provide a large-scale empirical foundation for tracking real-world AI failures, drawing from hundreds of millions of multilingual news records to organize incident reports into structured, analyzable records [1]. The resource, described in a paper submitted to arXiv on June 7, 2026, applies a multi-stage pipeline that identifies AI risk news, screens event-level reports, aligns related incidents, and classifies them across multiple dimensions [1]. The resulting collection organizes dispersed news reports into incident-centered records and provides benchmark datasets for event classification, incident alignment, and incident-level risk labeling [1]. An online platform also allows browsing and exploration of the data [1]. The creators argue that existing incident collections are often manually curated and too small for continuous, data-driven monitoring [1]. High-quality labeled training datasets for supervised machine learning are usually difficult and expensive to produce because of the large amount of time needed to label the data [4]. RiskNet is intended to support downstream research on AI safety, governance, risk analysis, and benchmarking, as well as longitudinal and cross-source analyses of AI-related harms [1]. Large language models, the technology behind modern chatbots, are typically based on transformer architectures and can generate, summarize, and translate text [3]. Biased or inaccurate training data can make an LLM’s output less reliable [3]. Benchmark evaluations for LLMs attempt to measure model reasoning, factual accuracy, alignment, and safety [3]. According to the Stanford HAI 2025 AI Index Report, on harder benchmarks introduced in 2023, model scores rose sharply within a single year; performance increased by 18.8, 48.9, and 67.3 percentage points on the MMMU, GPQA, and SWE-bench testing suites respectively by 2024 [3]. RiskNet’s release comes as AI incidents continue to draw scrutiny. DeepSeek, a Chinese AI company, launched its R1 model in January 2025 with training costs it claimed were US$6 million, far less than the US$100 million cost for OpenAI’s GPT-4 in 2023 [9]. Observers described the breakthrough as triggering a “Sputnik moment” for the United States in artificial intelligence [9]. The RiskNet paper is accessible through arXiv, which since November 2022 has collaborated with Hugging Face to integrate machine learning demos directly alongside papers [7]. Hugging Face’s paper pages also allow users to find models, datasets, and apps related to a paper [6].