> ## Content Index > Fetch the complete content index at: https://www.aigl.blog/llms.txt > Use this file to discover other available public pages before exploring further. # Summary of Adding Structure to AI Harm - URL: https://www.aigl.blog/summary-of-adding-structure-to-ai-harm/ - Published: 2026-09-18T19:01:43.000Z - Updated: 2026-09-18T19:01:43.000Z - Description: This one-page summary describes the CSET AI Harm Framework, a modular structure for defining and categorising harms linked to AI systems. It outlines harm classifications, customisation options, and an example taxonomy for the AI Incident Database. - Author: Jakub Szarmach - Tags: Framework, AI Governance, Risk Management, Responsible AI, Model Evaluation, #aigl-library ## ⚡ Quick Summary Published by the Center for Security and Emerging Technology (CSET), this one-page summary introduces the CSET AI Harm Framework. It presents the framework as a way to identify the key elements of AI harm, their relational structure, and definitions, without requiring one fixed interpretation of what AI harm means. The stated purpose is to support the monitoring and examination of varied and widespread harms from artificial intelligence systems, helping risk-mitigation efforts through improved understanding, earlier detection of emerging harms, and better direction of prevention resources. The framework defines AI harm as harm, or potential harm, experienced by an entity and directly linked to an AI system’s behaviour. It separates harms into tangible and intangible forms, permits users to add categories within both forms, and distinguishes harm that has occurred from harm that may occur. The summary also says that the accompanying paper explains how users can adapt the framework into a customised taxonomy, illustrates this through CSET’s AI Harm Taxonomy for the AI Incident Database, and outlines future additions to the framework. ## 🧩 What’s Covered The summary presents the framework’s stated rationale, core structure, and the additional subjects addressed in the accompanying paper. - **Purpose of AI-harm analysis:** The document links monitoring and examining AI harms to risk mitigation. It says analysis can improve understanding of how AI systems cause harm, support earlier detection of emerging harm types, and direct prevention resources to where they are most needed. - **Framework approach:** The CSET AI Harm Framework is described as identifying key elements of AI harm, their relational structure, and definitions. Its approach is explicitly not to impose a single interpretation of AI harm. - **Definition of AI harm:** AI harm is defined as a situation in which an entity experiences harm, or potential for harm, that is directly linked to the behaviour of an AI system. This definition includes potential harms as well as experienced ones. - **Tangible and intangible harm:** The framework groups harm from AI into tangible or intangible harm. Tangible harm is described as observable, verifiable, and definitive; intangible harm cannot be directly observed or does not have a material or physical effect. - **Expandable categories:** The document says that users can define additional categories of tangible and intangible harm. The framework supplies some common categories while allowing new ones to be included. - **Occurred and possible harm:** A further distinction separates harm that actually occurred from harm that may occur. The stated effect is to enable tracking of definitively experienced harm while also supporting research and analysis of harm risk and vulnerability. - **Customisation and example use:** The accompanying paper discusses how users can create a customised taxonomy by modifying the base modular structure and providing definitions. It illustrates this process with CSET’s AI Harm Taxonomy for the AI Incident Database and notes that future additions to the framework are detailed. ## 💡 Why it matters? For AI governance and risk teams, the framework offers a shared structure for recording both observed harms and potential harms associated with AI-system behaviour. That distinction can support separate work on incident tracking, harm-risk analysis, and vulnerability research, rather than treating all reported concerns as the same type of evidence. Its modular design is also practically relevant where an organisation needs categories that fit a particular context. The summary states that users may modify the base structure, provide their own definitions, and add tangible or intangible categories. This makes the framework relevant to teams developing taxonomies for analysing AI-related incidents or risks. ## ❓ What’s Missing This document is a one-page summary rather than the full report it links to. It does not list the framework’s key elements, relational structure, or the common harm categories it says the framework provides. It also does not show the customised taxonomy used for the AI Incident Database, explain the customisation process step by step, or specify the future additions it mentions. No examples of tangible or intangible harms are supplied, and the summary does not state implementation procedures, data-collection methods, governance responsibilities, validation criteria, or limitations of the framework. It provides no publication date, version number, or named authors. ## 👥 Best For This summary is best suited to AI governance, risk, and research teams seeking a concise account of CSET’s approach to classifying AI harms. It is particularly useful for people considering an incident or risk taxonomy that needs to distinguish observed from potential harm and allow tailored harm categories. ## 📄 Source Details *Summary of Adding Structure to AI Harm* is a one-page English document published under the CSET, Center for Security and Emerging Technology, name. No individual authors, publication year, edition, or version are printed. The document identifies a linked report at [https://cset.georgetown.edu/publications/adding-structure-to-ai-harm](https://cset.georgetown.edu/publications/adding-structure-to-ai-harm?ref=aigl.blog). The complete one-page PDF was available for analysis.