Towards Safer Generative Language Models: A Survey on Safety Risks, Evaluations, and Improvements
Provides a structured taxonomy of LLM safety risks that APS governance and risk teams can use when assessing generative AI deployments.
Key points
- MIT AI Risk Repository summarises a survey identifying seven core safety risks in generative language models.
- Risk categories include toxic content, hallucination, privacy leakage, and malicious use - directly relevant to APS AI governance frameworks.
- Survey is from 2023 (arXiv:2302.09270); useful as a taxonomy reference but not cutting-edge given rapid field evolution.
Implications for Australian agencies
- Consider Governance and risk teams could assess whether this taxonomy aligns with or usefully supplements existing agency AI risk frameworks and assessment templates.
- Monitor Policy teams may want to monitor the MIT AI Risk Repository more broadly as a curated source of risk frameworks relevant to responsible AI work.
Implications are AI-generated. Starting points, not advice — see methodology for how they're framed.
View original source
Copied.
"Towards Safer Generative Language Models: A Survey on Safety Risks, Evaluations, and Improvements"
Source: MIT AI Risk Repository – Blog
Published: 18 September 2024
URL: https://airisk.mit.edu/blog/towards-safer-generative-language-models-a-survey-on-safety-risks-evaluations-and-improvements
MIT's AI Risk Repository blog summarises a 2023 academic survey cataloguing seven safety risk categories for large language models: toxic content, bias and discrimination, ethics and morality, controversial opinions, hallucination and misleading outputs, privacy and data leakage, and malicious use. The paper also reviews evaluation methodologies and improvement strategies across the model development lifecycle. While the underlying research predates many current frontier model developments, the risk taxonomy remains a useful structured reference for agencies developing AI risk assessments or evaluating generative AI use cases.
Implications for Australian agencies:
- [Consider] Governance and risk teams could assess whether this taxonomy aligns with or usefully supplements existing agency AI risk frameworks and assessment templates.
- [Monitor] Policy teams may want to monitor the MIT AI Risk Repository more broadly as a curated source of risk frameworks relevant to responsible AI work.
Retrieved from SIMS, 18 July 2026.