What are the risks from AI? 

This week we spotlight the 30th framework of risks from AI included in the AI Risk Repository: Liu, Y., Yao, Y., Ton, J.-F., Zhang, X., Guo, R., Cheng, H., Klochkov, Y., Taufiq, M. F., & Li, H. (2023). Trustworthy LLMs: A survey and guideline for evaluating large language models' alignment. arXiv. https://arxiv.org/abs/2308.05374

Paper focus

This paper conducts a comprehensive exploration of aspects contributing to the trustworthiness of large language models (LLMs) in order to provide guidance on the evaluation of LLM alignment.

Included risk categories

This paper presents an overview of AI challenges organized into 7 major categories (and 29 subcategories) as part of a detailed taxonomy of LLM alignment requirements.

1. Reliability: producing correct, truthful, and consistent output

• Misinformation
• Hallucination
• Inconsistency
• Miscalibration
• Sycophancy

2. Safety: avoiding harmful and illegal output

• Violence
• Unlawful conduct
• Harms to minors
• Adult content
• Mental health issues
• Privacy violation

3. Fairness: avoiding bias and disparate performance across groups

• Injustice
• Stereotype bias
• Preference bias
• Disparate performance

4. Resistance to misuse: avoiding misuse for malicious purposes

• Propagandistic misuse
• Cyberattack misuse
• Social engineering misuse
• Leaking copyrighted content

5. Explainability and reasoning: ability to explain logic and output to users

• Lack of interpretability
• Limited logical reasoning
• Limited causal reasoning

6. Social norms: reflecting universal human values

• Toxicity
• Unawareness of emotions
• Cultural insensitivity

7. Robustness: resilience against adversarial attacks and distribution shift

• Prompt attacks
• Paradigm and distribution shifts
• Interventional effect
• Poisoning attacks

Key features of the framework and associated paper:

• Provides a guideline for multi-objective evaluations using automated and templated question generation
• Conducts measurement studies on widely-used LLMs across fine-grained alignment evaluation criteria based on the taxonomy

⚠️Disclaimer: This summary highlights a paper included in the MIT AI Risk Repository. We did not author the paper and credit goes to Yang Liu, Yuanshun Yao, Jean-Francois Ton, Xiaoying Zhang, Ruocheng Guo, Hao Cheng and co-authors. For the full details, please refer to the original publication: https://arxiv.org/abs/2308.05374

Further engagement 

→ View all the frameworks included in the AI Risk Repository 

→ Sign-up for our project Newsletter