Incident Report Translation: The incident database is currently collected from and presented to the English reading world, but intelligent systems are deployed across languages, cultures, and locations. This project is working towards support of all machine-translatable languages so that incident reports can be sourced from any language and presented back to languages supporting 7+ billion potential users.
Best Practice Resources: One motivation behind the AI Incident Database is to make incidents shareable with product and engineering teams so they will work to prevent or mitigate the incident before deployment. Currently, the database only presents the problems (i.e., incidents), but we are working towards selectively displaying resources that are most likely to prevent and mitigate the problem.
Post-Mortem Reporting: Hosted services have a practice of publishing "post-mortems" after service outages where the company explains what happened, who was affected, for how long, and what the company will be doing to prevent or mitigate such outages in the future. This is a practice the AI Incident Database is looking to support explicitly by giving special placement on the incident citation pages to the AI incident post-mortems.
Incident Monitoring: As more intelligent systems are deployed within the real world, it will become increasingly difficult to monitor, collect, and categorize incidents. A scraper that monitors content sources for AI incidents and facilitates the easy ingestion thereof would ensure the AIID continues to be relevant for the most recently developed intelligent systems. A student team is currently working to develop the core API for this project.
Potential Future Projects
Technical Failure Taxonomy: the AIID lacks a rigorous technical taxonomy linking experienced harms to their technical elements. We would like to close this gap with a team of machine learning engineers by applying their knowledge of how systems are built to classify incidents.
Incident Data Collection: Many incident types produce incident data that is analyzed by companies and used to avoid its recurrence internally. A research associate could be tasked with promoting the voluntary disclosure of incident data, which could then be associated with incident records. In the future, this type of disclosure is likely to be mandatory in some jurisdictions and the AIID should be there to support it.
Your Project Here: The AI Incident Database is governed by the Responsible AI Collaborative (RAIC). You should consider this an open invitation to collaborate on projects that match the impact thesis of the AI Incident Database.
Initial Incident Collection: The initial dataset was collected in 2019 by combining incident listings from Roman Yampolskiy, Catherine Olsson, and Sam Yoon. These incidents are associated with more than 1k total incident reports.
Incident Discovery Application: Subsequent to initial dataset collection, Sean McGregor developed a Discover application that supported the indexing and cleaning of the more than 1k reports assembled in the initial incident report collection.
Taxonomies: Intelligent systems are in development for every market segment and function of government. Developing the research quality of the data product requires systems for regularizing and classifying incidents according to incident types, scale, technologies involved, and impacted parties. These and other classifications are a point of contention in multi-stakeholder systems. To advance the research utility of the database while not requiring full consensus on every classification, it is possible to apply taxonomies independently of a central taxonomic authority. Each "taxonomic scope" is managed by an editor that develops the taxonomy, and is responsible for applying the taxonomy across incidents. Contact us for more information.