Managing AI Risks: New Findings and Emerging Challenges | AI study | Ai risks | Risks of artificial intelligence | Turtles AI

The challenges and uncertainties associated with AI use demand increasing attention from regulators and companies. New research shows that many AI models, despite being advanced, present critical issues regarding regulatory compliance and safety.

Highlights

• AIR-Bench 2024 assesses AI model compliance with regulations and specific risks.
• Models like Claude 3 Opus and Gemini 1.5 Pro stand out for their ability to avoid particular risks.
• Government regulations are often less detailed than internal company policies.
• AI evolution requires increasing attention to safety and regulatory compliance.

The rapid evolution of AI models has led to a complex scenario where not only their technical capabilities but also their potential legal, ethical, and compliance risks must be carefully evaluated. A prime example of this emerging focus is the work of Bo Li, an associate professor at the University of Chicago, who has gained prominence in the field for his detailed analyses of AI-related risks. His expertise has become invaluable to numerous consulting firms that increasingly concentrate not on how intelligent these models are, but rather on the problems they might generate.

Li, along with colleagues from other universities and companies such as Virtue AI and Lapis Labs, has developed a taxonomy of AI risks and a benchmark called AIR-Bench 2024, which assesses how large language models (LLMs) adhere to rules and safety principles. The study involved analyzing government AI regulations and guidelines from countries like the United States, China, and the European Union, as well as examining the usage policies of 16 major companies in the sector.

AIR-Bench 2024 uses thousands of prompts to test AI models against specific risks. The results reveal that some models perform better than others in certain areas. For instance, Anthropic’s Claude 3 Opus model excels at refusing to generate cybersecurity threats, while Google’s Gemini 1.5 Pro model is particularly effective at avoiding the creation of non-consensual sexually explicit content. Conversely, the DBRX Instruct model, developed by Databricks, showed lower performance in terms of overall safety, despite the company’s announcement of future improvements.

The analysis highlighted significant issues in AI development and regulation. In particular, government regulations, although advanced, often prove to be less detailed than corporate internal policies, suggesting the need for strengthened rules. Additionally, it emerged that many AI models are not always compliant with the companies’ own policies, revealing a significant margin for improvement.

In the context of this research, another project, conducted by MIT, has led to the creation of a database that collects and classifies AI-related dangers. This database, based on 43 different risk frameworks, highlights how some issues receive more attention than others. For example, privacy and security concerns are mentioned by over 70% of frameworks, while only 40% address the issue of misinformation.

These studies underscore how the cataloging and measurement of AI risks will need to evolve alongside technological development. Bo Li particularly emphasized the importance of considering emerging issues such as the emotional impact that AI models can have on users. An analysis of Meta’s Llama 3.1 model, conducted by her company, showed that despite the model’s increased capabilities, safety did not significantly improve.

The entire issue fits into a broader context of reflection on the role of AI in modern society. As the power and spread of these models increase, it is crucial that regulators and developers focus not only on technical capabilities but also on the potential risks associated with their use.