Large Language Models’ Trustworthiness in the Light of the EU AI Act
The emergence of large language models (LLMs) has transformed the landscape of artificial intelligence (AI), particularly in natural language processing. However, with their increasing deployment in high-stake domains, concerns regarding their trustworthiness have escalated. This study aims to systematically assess how LLMs adhere to the principles outlined in the EU AI Act, a pioneering legal framework introduced to ensure responsible AI development and deployment.
1. Introduction
The rapid advancements in deep learning, coupled with the availability of vast public datasets and powerful computational resources, have propelled the development of LLMs. Models like BERT and the GPT series have significantly improved machines’ ability to process and understand complex text, generating human-like responses. This progress has led to their adoption across various industries, including customer service, healthcare, education, and finance.
Despite their remarkable capabilities, the propensity of LLMs for hallucinations and inherent biases raises significant trust concerns. Establishing principles for responsible AI use is critical, and the EU Trustworthy AI framework outlines core principles such as fairness, transparency, accountability, and safety. The EU AI Act categorizes AI systems by risk levels and imposes stringent requirements on high-risk applications to safeguard human rights and safety.
2. Key Contributions
This study offers a systematic assessment of LLMs, focusing on:
- A structured analysis of the current state of LLMs concerning the trustworthiness aspects defined by the EU AI Act.
- Exploration of emerging trends in domain-specific LLM applications, highlighting existing gaps and underexplored areas.
- A comprehensive review of the methodologies applied in research on LLM trustworthiness, identifying types of research contributions.
3. Methodology
The study employs a systematic mapping process, structured into three classic phases: planning, conducting, and documentation. During the planning phase, research questions were established, and a search string was developed to identify relevant studies across scientific databases.
The conducting phase included study retrieval, selection, classification, and data extraction, ensuring comprehensive coverage and rigorous analysis of the selected literature. Finally, the documentation phase involved a thorough analysis of the extracted data, represented through various visualizations to address the research questions effectively.
4. Trustworthiness Dimensions
The study emphasizes several key dimensions of trustworthiness as per the EU AI Act:
- Human Oversight: Ensuring that LLMs are designed to allow human intervention, minimizing risks to safety and fundamental rights.
- Record-Keeping: Implementing logging capabilities to enhance accountability and traceability of LLM operations.
- Data Governance: Ensuring that the datasets used for LLM training are representative and well-documented to mitigate biases.
- Transparency: Providing clear instructions and explanations of LLM outputs to enhance user understanding.
- Accuracy: Striving for high levels of accuracy in LLM outputs while maintaining consistency throughout their lifecycle.
- Robustness: Designing LLMs to perform reliably under varying conditions.
- Cybersecurity: Implementing strong measures to protect against unauthorized access and manipulation.
5. Research Findings
The analysis of the literature reveals that a significant focus has been placed on models like GPT and BERT, with a noticeable gap in trustworthiness research for newer and niche models. Key findings include:
- Trustworthiness aspects such as accuracy and transparency dominate the current research landscape, while dimensions like cybersecurity and record-keeping require more attention.
- High-impact application domains like healthcare and education have been extensively studied, but critical areas such as cybersecurity, finance, and environment remain largely underexplored.
- Most studies emphasize solution proposals and evaluations, indicating a need for more conceptual and experience-based research approaches.
6. Conclusion
This systematic mapping study underscores the increasing importance of trustworthiness in LLMs and the need for a more balanced research approach that includes underrepresented domains. The findings suggest that while LLMs like GPT and BERT are the focus of much research, emerging models and high-risk application areas must also be examined to ensure the responsible deployment of AI technologies.
As the EU AI Act aims to take full effect, its principles will likely shape the future landscape of LLM development, emphasizing the need for transparency, accountability, and ethical AI use across various sectors.
