Artificial intelligence often raises fears about replacing workers. However, researchers Osman Ozaltin and Maria Mayorga see it differently. Instead, they designed AI to support investigators fighting human trafficking. They worked alongside Margaret Tobey, an Operations Research Program Ph.D. alumna and Sherrie Bosisto, founder and Executive Director of the Global Emancipation Network. Together, the team focused on protecting both victims and the professionals who serve them.

First, Ozaltin explains why this job is so difficult. When investigators review online content, “they’re not just reading neutral data; they’re encountering descriptions of exploitation, coercion and abuse.” Even when the wording seems subtle, “the implications are disturbing.” As a result, repeated exposure takes a toll. Over time, “repeatedly analyzing this kind of content can create emotional fatigue and secondary trauma.” At the same time, analysts must stay focused and objective. Therefore, “that combination of cognitive intensity and emotional exposure makes the work uniquely challenging.”

Because of these pressures, the research team built a system that reduces unnecessary exposure. Instead of reviewing thousands of posts, investigators see only selected examples. Specifically, the AI uses active learning to choose the most informative content. In other words, the system learns from fewer smarter reviews. Consequently, investigators spend less time screening noise while keeping strong detection performance.

Moreover, Ozaltin says sustainability drove their work. “Labeling this type of data isn’t just time-consuming; it’s emotionally demanding,” he explains. Therefore, the system “strategically select[s] the most informative reviews.” By doing so, it reduces “unnecessary exposure to harmful content while maintaining high detection performance.” Over time, this approach “can lower burnout risk, improve productivity and make it more feasible for organizations with limited resources to continue this work effectively.”

In practice, this shift changes how experts use their time. If they “don’t have to sift through thousands of mostly irrelevant reviews,” Ozaltin says, “they can focus their attention on the small subset that truly matters.” Furthermore, the team showed that “smart selection of which reviews to label can significantly reduce manual effort while improving model performance.” As a result, analysts spend more time “building cases, identifying patterns and supporting investigations.”

Importantly, the researchers do not want to replace human judgment. Instead, they want to strengthen it. “This research is not about replacing human judgment; it’s about supporting it,” Ozaltin says. The system flags concerning material, however trained analysts make the final decisions. In addition, he notes that “advanced analytics… can both improve outcomes and protect the people doing difficult work.”

Meanwhile, Mayorga focused on investigator well-being from the start. “That was the main goal of this project,” she says, “reducing the amount of time and effort that investigators have invested while still obtaining good model performance.” In addition, she explains that models must adapt as language changes over time and across regions. However, she adds, “we don’t want to bog investigators down with the labeling task over and over.”

Overall, this research benefits society in clear ways. It helps investigators stay effective without burning out. At the same time, it keeps detection tools accurate as trafficking tactics evolve. Most importantly, it shows that AI works best when researchers, practitioners and advocates work together to protect people.

View the research paper at https://pubsonline.informs.org/doi/abs/10.1287/opre.2023.0625.

This post was originally published in the Edward P. Fitts Department of Industrial and Systems Engineering.