Investigating and detecting human trafficking recruitment on online platforms using a responsible innovation approach

Thesis

Human trafficking affects more than 40 million people in the world. In recent years, human traffickers have shifted from conventional physical methods and are increasingly using online platforms to identify and exploit victims for various forms of abuse. These harms can affect victims physically, emotionally, and psychologically. This thesis examines human trafficking recruitment on online platforms and explores the application of artificial intelligence to detect fraudulent job advertisements used for victim recruitment, while considering guidelines that mitigate potential unintended negative consequences. This research was carried out in four phases. The first phase was a qualitative study that explored the implications of developing and using anti-trafficking tools. Guided by the principles of responsible research and innovation, this phase identified challenges and the intended and unintended consequences of deploying anti-trafficking technology. It provided recommendations for addressing these challenges and proposed a framework to mitigate potential negative consequences, informing the subsequent phases of this research. The second phase focused on understanding human trafficking recruitment. Through the analysis and visualisation of two real-world datasets, this phase uncovered patterns, profiles of victims and perpetrators, recruitment indicators, processes, and recurring trends. The third phase was an ethnographic study that investigated how traffickers use social media to recruit victims using fraudulent job advertisements. It also provided insights into the indicators of such advertisements and highlighted the strategies anti-trafficking organisations employ to identify them. In the final phase, a labelled dataset consisting of job advertisements related to human trafficking was developed. This phase included benchmarking traditional machine learning algorithms, pre-trained neural language models, and large language models to classify fraudulent job advertisements. The research found that pre-trained neural language models were the most effective in identifying advertisements related to human trafficking, and these models were subsequently used to label 10,000 unlabelled data points. The research methods employed in these four phases were guided by principles of ethical responsibility, recent advancements in computer science, and a mixed-methods approach. These guiding principles ensured that the selected methods were both scientifically rigorous and socially responsible in addressing the goals of the research. The four phases collectively contributed to a deeper understanding of human trafficking recruitment and demonstrated how technology can be leveraged to help combat this global issue.

Authors: Moyo, T

• DOI: 10.5287/ora-zog2a2rk9
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: responsible research and innovation; technology-facillitated trafficking; human trafficking; artificial intelligence; machine learning