• Skip to main content
  • Skip to primary sidebar
  • Skip to footer
  • Core Principles of Responsible AI
    • Accountability
    • Fairness
    • Privacy
    • Safety and Security
    • Sustainability
    • Transparency
  • Special Topics
    • AI in Industry
    • Ethical Implications
    • Human-Centered Design
    • Regulatory Landscape
    • Technical Methods
  • Living Dictionary
  • State of AI Ethics
  • AI Ethics Brief
  • 🇫🇷
Montreal AI Ethics Institute

Montreal AI Ethics Institute

Democratizing AI ethics literacy

A collection of principles for guiding and evaluating large language models

February 5, 2024

🔬 Research Summary by Matthias Samwald, an associate professor at the Medical University of Vienna and works on making powerful AI systems driving biomedical progress more trustworthy.

[Original paper by Konstantin Hebenstreit, Robert Praas, and Matthias Samwald]

NOTE: Published at Socially Responsible Language Modelling Research (SoLaR) workshop, NeurIPS 2023


Overview: This paper addresses the challenges in assessing and guiding the behavior of large language models (LLMs). It proposes a set of core principles based on an extensive review of literature from a wide variety of disciplines, including explainability, AI system safety, human critical thinking, and ethics.


Introduction

Large language models (LLMs) like GPT-4 show great promise. But what are the principles that should govern how they reason and behave? 

This paper tries to provide answers to these questions by assembling ideas from a wide variety of sources, such as work on machine learning explainability, AI system safety, human critical thinking, and ethical guidelines (Figure 1). The study’s main objective is to compile a set of core principles for steering and evaluating LLMs’ reasoning, using an extensive literature review and a small-scale expert survey. These principles can prove useful to monitor and improve model behavior during training and inference and can guide human evaluation of model reasoning.

Key Insights

Compilation and curation of principles

The authors set out the study by conducting an extensive literature review across several domains, leading to the extraction of 220 principles. These were categorized into seven areas: assumptions and perspectives, information and evidence, robustness and security, utility, reasoning, ethics, and implications. A consolidated set of 37 core principles was distilled, considering the relevance to current LLMs and adaptability to various reasoning tasks and modalities.

Expert Survey and Validation

For external validation, domain experts were surveyed to rate the importance of each principle. This helped refine and validate the core principles, ensuring they align with expert opinions and are relevant to practical applications.

Results and Implications

The principles are meant to guide the reasoning of LLMs. They are envisioned to serve in monitoring and steering models, improving behavior during training, and aiding in human evaluation.

Between the lines

The principles are biased towards current language-based applications of LLMs. Expansions will likely be necessary soon to address emerging concerns, as AI systems evolve from LLMs into more autonomous agents.

The work is preliminary and the set of principles is far from exhaustive. The authors call for further community engagement to refine and expand these principles. They suggest methods like Delphi studies or online platforms for iterative principle development and emphasize the need for input from a broader range of stakeholders.

There is ample opportunity for empirical research testing the practical utility of these (and other) principles. As a simple example, prompts derived from these principles can be tested on state-of-the-art models to evaluate their impact on performance, explainability, and ethical alignment.

Want quick summaries of the latest research & reporting in AI ethics delivered to your inbox? Subscribe to the AI Ethics Brief. We publish bi-weekly.

Primary Sidebar

🔍 SEARCH

Spotlight

AI Policy Corner: Frontier AI Safety Commitments, AI Seoul Summit 2024

AI Policy Corner: The Colorado State Deepfakes Act

Special Edition: Honouring the Legacy of Abhishek Gupta (1992–2024)

AI Policy Corner: The Turkish Artificial Intelligence Law Proposal

From Funding Crisis to AI Misuse: Critical Digital Rights Challenges from RightsCon 2025

related posts

  • The Political Power of Platforms: How Current Attempts to Regulate Misinformation Amplify Opinion Po...

    The Political Power of Platforms: How Current Attempts to Regulate Misinformation Amplify Opinion Po...

  • Choices, Risks, and Reward Reports: Charting Public Policy for Reinforcement Learning Systems

    Choices, Risks, and Reward Reports: Charting Public Policy for Reinforcement Learning Systems

  • Mapping the Ethics of Generative AI: A Comprehensive Scoping Review

    Mapping the Ethics of Generative AI: A Comprehensive Scoping Review

  • From Pretraining Data to Language Models to Downstream Tasks: Tracking the Trails of Political Biase...

    From Pretraining Data to Language Models to Downstream Tasks: Tracking the Trails of Political Biase...

  • The Two Faces of AI in Green Mobile Computing: A Literature Review

    The Two Faces of AI in Green Mobile Computing: A Literature Review

  • Target specification bias, counterfactual prediction, and algorithmic fairness in healthcare

    Target specification bias, counterfactual prediction, and algorithmic fairness in healthcare

  • Artificial Intelligence: the global landscape of ethics guidelines

    Artificial Intelligence: the global landscape of ethics guidelines

  • Towards Environmentally Equitable AI via Geographical Load Balancing

    Towards Environmentally Equitable AI via Geographical Load Balancing

  • On the Generation of Unsafe Images and Hateful Memes From Text-To-Image Models

    On the Generation of Unsafe Images and Hateful Memes From Text-To-Image Models

  • Re-imagining Algorithmic Fairness in India and Beyond (Research Summary)

    Re-imagining Algorithmic Fairness in India and Beyond (Research Summary)

Partners

  •  
    U.S. Artificial Intelligence Safety Institute Consortium (AISIC) at NIST

  • Partnership on AI

  • The LF AI & Data Foundation

  • The AI Alliance

Footer

Categories


• Blog
• Research Summaries
• Columns
• Core Principles of Responsible AI
• Special Topics

Signature Content


• The State Of AI Ethics

• The Living Dictionary

• The AI Ethics Brief

Learn More


• About

• Open Access Policy

• Contributions Policy

• Editorial Stance on AI Tools

• Press

• Donate

• Contact

The AI Ethics Brief (bi-weekly newsletter)

About Us


Founded in 2018, the Montreal AI Ethics Institute (MAIEI) is an international non-profit organization equipping citizens concerned about artificial intelligence and its impact on society to take action.


Archive

  • © MONTREAL AI ETHICS INSTITUTE. All rights reserved 2024.
  • This work is licensed under a Creative Commons Attribution 4.0 International License.
  • Learn more about our open access policy here.
  • Creative Commons License

    Save hours of work and stay on top of Responsible AI research and reporting with our bi-weekly email newsletter.