Scaling AI Incident Analysis: A Promising Step Forward

← Alle Beiträge
Künstliche Intelligenz · EN

Scaling AI Incident Analysis: A Promising Step Forward

A collaboration between Simon Mylius (Winter Fellow at the Centre for the Governance of AI) and the AI Incident Database (AIID) has resulted in a promising proof-of-concept: automated, scalable AI incident classification using LLMs.

How it works:

An LLM-based tool processes raw incident reports, classifying each according to two major frameworks:

1) The MIT AI Risk Causal and Domain Taxonomies, and
2) A Harm Severity Scale based on the CSET Taxonomy of AI Harm.

The tool summarizes each incident, identifies the type and severity of reported harms, assigns domain and causal risk classifications, and provides a confidence rating based on information quality. Importantly, every classification decision includes a short, machine-generated justification to improve transparency and enable validation.

Why it matters:

With AI incidents likely to rise as deployment grows, scalable methods like this are crucial. Automation enables: Faster analysis of increasing volumes of incident reports, application of updated taxonomies to historical data, and integration of datasets into consistent structures for pattern and trend analysis.

Methodology highlights:

The system uses two LLM calls per incident: one for summarization/classification, another for independent severity rating.
The harm severity scale covers 10 categories, including physical harm, financial loss, privacy infringement, and harm to democratic norms.
Incidents are also categorized according to EU AI Act risk levels (unacceptable, high-risk, limited risk, minimal risk).

Limitations and next steps:

Although early results are promising, spot-checking suggests approximately 20% of incidents have a minor mismatch between the automated severity rating and human evaluation. Qualitative categories show more variance than quantitative ones, and some discrepancies stem from sparse or inconsistent original reports.

Interactive dashboards are now available on the AI Incident Tracker site, allowing you to explore trends, groupings, and individual cases in detail.

Source I recommend: https://lnkd.in/eA5fEaCd

Zu diesem Beitrag gibt es auf LinkedIn ein Bild. Auf LinkedIn ansehen →

Auf LinkedIn diskutieren