Taxonomy of Risks Posed by Language Models
A structured LLM risk taxonomy offers APS agencies a reusable reference for scoping AI risk registers and governance frameworks—though the source is not new.
Key points
- MIT AI Risk Repository spotlights a 2022 Google DeepMind taxonomy covering six domains and 20 subdomains of LLM risk.
- The taxonomy distinguishes observed versus anticipated risks - a useful framing for APS risk assessments and AI governance documentation.
- The underlying paper is from 2022; the blog post adds no new analysis, limiting its immediate signal value.
Summary
The MIT AI Risk Repository has highlighted a 2022 FAccT paper by Weidinger et al. (Google DeepMind) that provides a structured taxonomy of ethical and social risks from large language models. The taxonomy organises risks across six domains: discrimination and exclusion, information hazards, misinformation, malicious uses, human-computer interaction harms, and environmental and socioeconomic harms. It distinguishes between observed and anticipated risks and focuses on risks from operating raw language models rather than specific applications. The blog post itself adds no new analysis beyond summarising the paper.
Implications for Australian agencies
- Consider Agencies developing or updating AI risk registers may want to cross-reference the Weidinger et al. taxonomy to check coverage of LLM-specific risk categories.
- Monitor Teams tracking the MIT AI Risk Repository may want to monitor subsequent spotlighted frameworks for more recent or operationally focused risk taxonomies.
Implications are AI-generated. Starting points, not advice.
"Taxonomy of Risks Posed by Language Models" Source: MIT AI Risk Repository – Blog Published: 12 May 2025 URL: https://airisk.mit.edu/blog/taxonomy-of-risks-posed-by-language-models The MIT AI Risk Repository has highlighted a 2022 FAccT paper by Weidinger et al. (Google DeepMind) that provides a structured taxonomy of ethical and social risks from large language models. The taxonomy organises risks across six domains: discrimination and exclusion, information hazards, misinformation, malicious uses, human-computer interaction harms, and environmental and socioeconomic harms. It distinguishes between observed and anticipated risks and focuses on risks from operating raw language models rather than specific applications. The blog post itself adds no new analysis beyond summarising the paper. Implications for Australian agencies: - [Consider] Agencies developing or updating AI risk registers may want to cross-reference the Weidinger et al. taxonomy to check coverage of LLM-specific risk categories. - [Monitor] Teams tracking the MIT AI Risk Repository may want to monitor subsequent spotlighted frameworks for more recent or operationally focused risk taxonomies. Retrieved from SIMS, 18 May 2026.