Core Concepts
Large Language Models lack efficacy in online grooming prevention, posing potential harm.
Abstract
The study examines the use of Large Language Models (LLMs) for online grooming prevention. It highlights the increasing interaction of children with LLMs and the importance of scrutinizing their safety, especially for critical applications like online child safety queries. The research explores the performance of 6 LLMs in identifying and avoiding grooming through advice generation, as well as the impact of prompt design on model behavior. Results from over 6,000 LLM interactions reveal a lack of models suitable for online grooming prevention, with inconsistencies in behaviors and potential harmful answer generation, particularly from open-source models. Suggestions for improvement and identification of prompt designs that affect model performance are provided to inform best practices usage guides.
Stats
"over 6,000 LLM interactions"
"180.5 million users worldwide"
"13B-chat model chosen"
"7B-instruct model used"
Quotes
"We find that no models were clearly appropriate for online grooming prevention."
"Models fall short with potential for harmful answer generation."
"Suggestions provided for improvement in model performance."