International Scientific Report on the Safety of Advanced AI

 

Here I share the report of the AI Safety

 

international_scientific_report_on_the_safety_of_advanced_ai_interim_report.pdf

4.01MB

 

https://www.gov.uk/government/publications/international-scientific-report-on-the-safety-of-advanced-ai

International Scientific Report on the Safety of Advanced AI

 

I will write the contents below 

 

0. Title : International Scietific Report on the Safety of Advanced AI - Interim Report 

 

1. Contributors 

Chair : Prof. Yoshua Bengio

Expert Advisory Panel

- Prof Bronwyn Fox, The Commonwealth Scientific and Industrial Research Organization (CSIRO) (Australia) 

- Andre Carlos Ponce de Leon Ferreira de Carvalho, Institute of Mathematics and Computer Sciences, University of Sao Paulo (Brazil) 

- Dr. Mona Nemer, Cheif Science Advisor of Canada (Canada) 

- Raquel Pezoa Rivera, Federico Santa Maria Technical University (chile) 

- Dr. Yi Zeng, Institute of Automation, Chinese Academy of Science (China) 

- Juha Keikkila, DG Connect (European Union) 

- Guillaume Avrin, General Directorate of Enterprises (France) 

- Prof. Antonio Kruger, German Research Center for Artificial Intelligence (Germany)

- Prof. Balaraman Ravindran, Indian Institute of Technology, Madras (India) 

- Prof. Hammam Riza, KORIKA (Indonesia)

- Dr. Ciaran Seoighe, Science Foundation Ireland (Ireland)

- Dr. Ziv Katzir, Israel Innovation Autority (Israel) 

- Dr. Andrea Monti, University of Chieti-Pescara (Italy) 

- [Interim] Mary Kerema, Ministry of Information Communications Technology and Digital Economy (Kenya)

- Dr. Jose Ramon Lopez Portillo, Q Element (Mexico)

- Prof. Haroon Sheikh, Netherlands' Scientific Council for Government Policy (Netherlands) 

- Dr. Gill Jolly, Ministry of Business, Innovation and Employment (New Zealand)

- Dr. Olubunmi Ajala, Innovation and Digital Economy (Nigeria) 

- Dominic Ligot, CirroLytix (Phillippines)

- Prof. Kyung Mu Lee, Department of Electrical and Computer Engineering, Seoul National University (Repulic of Korea)

- Ahmet Halit Hatip, Turkish Ministry of Industry and Technology (Republic of Turkey) 

- Crystal Rugege, National Center for AI and Innovation Policy (Rwanda) 

- Dr. Fahad Albalawi, Saudi Authority for Data and Artificial Intelligence (Kingdom of Saudi Arabia) 

- Denise Wong, Data Innovation and Protection Group, Inforcomm Media Development Autority (IMDA) (Singapore)

- Dr. Nuria Oliver, ELLIS Alicante (Spain) 

- Dr. Christian Busch, Federal Department of Economic Affairs, Education and Research (Switzerland)

- Oleksii Molchanovskyi, Expert Committee on the Development of Artificial Intelligence in Ukraine (Ukraine)

- Marwan Alserkal, Ministry of Cabinet Affairs, Prime Minister's Office (United Arab Emirates)

- Saif M. Khan, U.S. Department of Commerce (United States) 

- Dame Angela McLean, Government Chief Scientific Adviser (United Kingdom) 

- Amandeep Gill, UN Tech Envoy (United Nations) 

 

Writing Group 

- Daniel Privitera (Lead Writer), KIRA Center

- Tamay Besiroglu, Epoch AI 

- Rishi Bommasani, Stanford University

- Stephen Casper, Massachusetts Institute of Technology 

- Yejin Choi, University of Washington / A12

- Danielle Goldfarb, Mila - Quebec AI Institute

- Hoda Heidari, Carnegie Mellon University

- Leila Khalatbari, Hong Kong University of Science and Technology

- Shayne Longpre, Massachusetts Institute of Technology 

- Vasilios Marvroudis, Alan Turing Institute

- Mantas Mazeika, University of Illinois at Urbana-Champaign

- Kwan Yee Ng, Concordia AI 

- Chinasa T. Okolo, Ph.D The Brookings Institution

- Deborah Raji, Mozilla

- Theodora Skeadas, Humane Intelligence

- Florian Tramer, ETH Zurich 

 

SCIENTIFIC COORDINATOR

 

 

List

Forewords

This report is the beginning of a journey on AI Safety 

 

AI Safety is a shared global issue

 

A critical step forward and a Call to Action on AI Safety 

The rapid advancement of AI stands poised to reshape our world in ways both profound and unforeseen. From revolutionising healthcare and transportation to autoating complex tasks and unlocking scientific breakthroughs, AI's potential for positive impact is undeniable. 

However, alongside these notable possibilities lie significant challenges that necessitate a forward-looking approach. Concerns range from unintended biases embedded in algorithms to the possibility of autonomous systems exceeding human control. These potential risks highlight the urgent need for a global conversation to ensure the safe, and responsible advancement of AI. 

In this context, the International AI Safety Report will provide vital groundwork for global collaboration. The report represents a convergence of knowledge from experts across 30 countries, the European Union, and the United Nations, providing a comprehensive analysis of AI Safety. By focusing on the early scientific understanding of capabilities and risks from general purpose AI and evaluating technical methods for assessing and mitigating them, the report will spark ongoing dialogue and collaboration among multi-stakeholders.

I hope that based on this report, experts from 30 countries, the EU, and the UN continue to engage in balanced discussions, achieving AI risk mitigation that is acceptable and tailored to the specific context of both developed and developing countries, thereby creating a future where innovation and responsible AI coexist harmoniously. 

Lee Jong-Ho, Minister of MSIT, 

Republic of Korea

 

Executive Summary

1 Introduction

2 Capabilities

2.1 How does General-Purpose AI gain its capabilities?

2.2. What current general-purpose AI systems are capable of

2.2.1 Capabilities by modality 

2.2.2 Capabilities and Limitations by skill

2.3. Recent trends in capabilities and their drivers

2.3.1. Recent trends in compute, data, and algorithms

2.3.2. Recent trends in capabilities

2.4. Capability progress in coming years

2.4.1. If resources continue to be scaled rapidly, would this lead to rapid advancements?

2.4.2. Will resources be scaled rapidly?

2.4.3. Will algorithmic progress lead to rapid advancements? 

3 Methodology to assess and understand general-purpose AI systems

3.1 General-purpose AI assessments serve to evaluate model capabilites and impacts

3.2. Approaches for model performance analysis

3.2.1 Case studies

3.2.2 Benchmarks

3.2.3. Red-teaming and adversarial attacks

3.2.4. Auditing 

3.3 Model transparency, explanations, and interpretations

3.4 Challenges with studying general-purpose AI systems

4 Risks

4.1 Malicious use risks

4.1.1 Harm to individuals through fake content

4.1.2 Disinformation and manipulation of public opinion

4.1.3 Cyber offence

4.1.4 Dual use science risks

4.2 Risks from malfunctions

4.2.1 Risks from product functionality issues

4.2.2 Risks from bias and underrepresentation

4.2.3 Loss of control

4.3 Systemic risks

4.3.1 Labour market risks

4.3.2 Global AI divide

4.3.3 Market concentration risks and single points of failure

4.3.4 Risks to the environment

4.3.5 Risks to privacy

4.3.6 Copyright infringement

4.4 Cross-cutting risk factors

4.4.1 Cross-cutting technical risk factors 

4.4.2 Cross-cuttin societal risk factors 

5 Technical approaches to mitigate risks

5.1 Risk management and safety engineering

5.1.1 Risk assessment

5.1.2 Risk management 

5.2 Training more trustworthy models

5.2.1. Aligning general-purpose AI systems with developer intentions

5.2.2 Reducing the hallucination of falsehoods

5.2.3 Improving robustness to failures 

5.2.4 Removing hazardous capabilities

5.2.5 Analysing and editing the inner workings of models

5.3 Monitoring and intervention

5.3.1 Detecting general-purpose AI-generated content

5.3.2 Detecting anomalies and attacks

5.3.3 Explaining model actions

5.3.4 Building safeguards into AI systems

5.4 Technical approaches to fairness and representation in general-purpose AI systems 

5.4.1 Minigation of bias and discrimination works throughout the stages of general-purpose AI development and deployment

5.4.2 Is fairness in general-purpose AI systems achievable? 

5.4.3 Challenges in achiving fair general-purpose AI systems

6 Conclusion

Chair's note on the interim report

Differing views

Glossary

References