Workshop on Datasets and Evaluators of AI Safety

Key information (Call for Papers)


Background

Advanced AI systems have the potential to drive economic growth and productivity, boost health and well-being, improve public services, and increase security. However, AI models can also cause societal harms and can be misused. This workshop focuses on evaluating the safety of AI models and in particular on LLMs. We are especially interested in work on improving datasets and benchmarks, as well as devising methods for evaluating the safety of AI models through the development of evaluators.


The goals of this full-day workshop, organised in collaboration with Kaggle, King's College London and the Open Data Institute are to bring together academic and industrial researchers working on datasets and evaluators for AI safety. 


Concerns regarding the safety of AI emerge from the potential harmful uses or consequences of AI outputs, which can result from inaccuracies, irresponsibility, or inappropriate applications.  As AI becomes increasingly integrated into critical systems and everyday activities, addressing AI safety issues is imperative. The misuse of AI technologies for generating misinformation, conducting sophisticated cyberattacks, developing weapons, or providing harmful advice presents grave concerns. 


AI can cause societal harm, encouraging radicalisation and promoting biased or skewed views. AI-generated fake, yet highly realistic content could reduce public trust in information and government bodies. Moreover, long-term existential risks associated with the development of superintelligent AI systems cannot be ignored.


A significant portion of these safety concerns can be attributed to data-related problems at various stages of the AI lifecycle. The growing adoption of frontier foundation models in mainstream applications has amplified these concerns. Specifically, the lack of transparency regarding the data used to pre-train these models and the data approaches to fine-tuning these models for custom applications can lead to unintended consequences. 


Characteristics of AI systems that need to be evaluated to ensure their safety include, but are not limited to, alignment, robustness to adversarial attacks, fairness, trustworthiness, deception capabilities, AI drift, explainability, privacy preservation, and reliability. Evaluation of these characteristics is challenging, not less due to the lack of benchmarks that are able to certify the level of safety of a given AI system.


This workshop will explore the role of data in AI safety, with a particular emphasis on data-centric AI approaches and their current limitations across the AI lifecycle.


Topics 

This workshop welcomes contributions towards creation and improvement of datasets and benchmarks used for evaluation of safety of AI systems (in particular LLMs) according to the characteristics listed above. The work can be theoretical (new approaches to creating datasets or proving a dataset's adequacy for the task) or practical (suggesting new datasets, checking the validity of existing datasets, improving existing datasets). We also welcome contributions that use existing benchmarks to demonstrate safety or AI systems or find unsafe behaviours.


Format of Workshop


Attendance 


Organizing Committee


Program Committee




Registration

Please register via AAAI: https://aaai.org/conference/aaai/aaai-25/


Call for papers

Please find the call for papers here: Call for Papers