The updated tracker applies five severity levels across 13 categories, including psychological harm, damage to knowledge and information, and child sexual exploitation and abuse
The MIT AI Risk Initiative’s tracker classifies reported AI harms across 13 categories and five severity levels. Conceptual illustration.
The AI Risk Initiative at the Massachusetts Institute of Technology (MIT) has released an updated scale for assessing the severity of reported AI harms, using automated analysis of more than 1,500 incidents.
Its AI Incident Tracker draws on the AI Incident Database, maintained by the Responsible AI Collaborative. Arcola AI built and runs the large language model (LLM) pipeline that assesses incident reports and assigns ratings.
Ten of the harm categories are based on the Center for Security and Emerging Technology’s (CSET) AI Harm Taxonomy. The tracker includes three additional categories: psychological harm, epistemic harm and child sexual exploitation and abuse.
Epistemic harm covers damage to knowledge and information, ranging from fabricated citations and false claims to pollution of the scientific record and erosion of trust in expertise.
Peter Slattery, of the MIT AI Risk Initiative and MIT FutureTech, announced the update on LinkedIn and credited Simon Mylius with leading the work.
Reported harm to knowledge, privacy and wellbeing
The tracker’s displayed table uses a base of 1,785 incidents and assigns severity ratings separately for each type of harm.
In the epistemic harm category, 701 incidents are assigned Minor, 26 Substantial and 29 Severe. The category’s definitions cover increasingly extensive damage to information and knowledge, including flawed decisions, degraded scientific evidence and corruption of information systems.
For psychological harm, the table records 346 incidents in the Minor band, 17 in Substantial and two in Severe. Its definitions span temporary distress through more serious emotional harm and widespread psychological damage.
The privacy category includes 95 incidents assigned Substantial and 41 Severe. In the child sexual exploitation and abuse category, 61 incidents are assigned Substantial and six Severe.
These figures count incidents assigned to a particular category and severity band. The same incident can appear under several types of harm.
Slattery explained: “An incident can cause several types of harm, so counts overlap across categories, and Negligible includes incidents with no reported harm of that type.”
How the five severity levels should be read
Each category uses the same five labels: Negligible, Minor, Substantial, Severe and Catastrophic. Their definitions depend on the type of harm being assessed.
For direct financial losses, for example, the Minor band covers losses from $10,000 to $1 million. The Catastrophic band covers losses above $10 billion, with definitions extending to economic collapse.
Other categories use measures such as the number of people affected, the extent of service disruption or the scale and persistence of damage.
Alongside the three additional categories, the framework covers physical harm, infrastructure damage, property damage, financial loss, environmental damage, toxic or malicious content, differential treatment, human and civil rights, democratic norms and privacy.
A Negligible rating can include both the lowest-severity cases and incidents with no reported harm in that particular category. It therefore describes one dimension of an incident, rather than an overall judgment that the event was harmless.
The tracker also has a defined boundary around what it measures. Slattery said: “The scale rates individual incidents, so it does not capture diffuse harms that build up across many events, such as labor displacement.”
Automated ratings depend on the underlying reports
The classifications are produced from raw incident reports submitted to the AI Incident Database by members of the public and subject matter experts.
The tracker acknowledges that those reports vary in quality, reliability and detail. Voluntary reporting also introduces sampling bias, limiting what the collection can establish about wider patterns of AI harm.
The classification pipeline has been developed iteratively, with spot-checks used to identify misclassifications and provide feedback for improvements.
According to the tracker’s methodology statement, the tool’s agreement with human expert consensus is comparable to agreement between two independent human experts. The supporting article is described as forthcoming.
The statement also acknowledges that some automated classifications are likely to differ from expert consensus. It advises treating patterns and trends as indicative and validating them through further analysis.
Slattery said: “We're open to feedback; we plan to add new categories and improvements in future iterations.”
ETIH Innovation Awards 2027 open for entries
The ETIH Innovation Awards 2027 are now open for entries, with 26 categories recognising companies, institutions, projects and individuals making a measurable impact across education technology. The global awards span AI in education, digital learning, higher education, skills and workforce, inclusion, education operations, evidence and impact, with entries managed through ETIH’s new online awards platform.
Entries are open worldwide, with the early-bird deadline on Friday 18 December 2026 and the final entry deadline on Friday 26 February 2027 at 23:59 UK time.
Explore the 2027 categories and judging criteria, read the How to Enter guide or start an entry through the ETIH Innovation Awards 2027 platform.
