The MIT AI Risk Initiative has just released a new iteration of our harm severity scale and mapping. Our scale rates the severity of harm from the 1500+ incidents captured in the AI Incident Database… | Peter Slattery, PhD | 22 comments
MIT’s AI Risk Initiative released an updated harm-severity scale for more than 1,500 incidents in the AI Incident Database. It gives organizations a consistent way to assess and manage documented AI harms.
The framework covers 13 harm categories, including three additions: psychological harm, epistemic harm, and child sexual exploitation and abuse. Each category has five severity levels, from Negligible to Catastrophic, and one incident can count in multiple categories.
The scale evaluates individual incidents, so it does not measure cumulative harms such as labor displacement. “Negligible” also includes cases with no reported harm in a category, which commenters said could blur the difference between low harm and insufficient information.
MIT plans to revise the framework, add categories, and consider feedback.
Takeaways
The framework’s incident-level design makes category counts overlap and leaves cumulative effects, such as labor displacement, outside its measurement scope.
A key interpretive issue is that the Negligible level combines genuinely low harm with cases where no harm was reported, a distinction the team may address in future revisions.
Article timeline
Build this article’s timeline
Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.
Manipulation Score: LowHow much this article uses emotional framing vs. straight reporting
Loaded Language525%
Absolutist215%
Threat/Panic825%
Us vs Them015%
Engagement Bait1220%
Manipulation score look right?
Factual Claims (5)
lowtechnology
MIT’s AI Risk Initiative released an updated harm-severity scale for more than 1,500 incidents in the AI Incident Database.
The MIT AI Risk Initiative has just released a new iteration of our harm severity scale and mapping. Our scale rates the severity of harm from the 1500+ incidents captured in the AI Incident Database
mediumtechnology
The scale gives organizations a consistent way to assess and manage documented AI harms.
lowtechnology
The framework covers 13 harm categories, including psychological harm, epistemic harm, and child sexual exploitation and abuse.
It has 13 harm categories... Psychological Harm, Epistemic Harm and Child Sexual Exploitation and Abuse were added for this tracker.
lowtechnology
Each category has five severity levels from Negligible to Catastrophic, and one incident can count in multiple categories.
It has 13 harm categories, each with five levels from Negligible to Catastrophic... An incident can cause several types of harm, so counts overlap across categories
lowsocial
The scale evaluates individual incidents and therefore does not measure cumulative harms such as labor displacement.
The scale rates individual incidents, so it does not capture diffuse harms that build up across many events, such as labor displacement.
Comments
Login to view and post comments.