Can AI help prosecutors prove human trafficking?

Human trafficking cases often fail not because the crime did not occur, but because it cannot be proven.

Trafficking signs are frequently subtle. Victims may not identify themselves as victims. Evidence is fragmented across conversations, images, and behaviors that, taken alone, appear insignificant. By the time a case reaches court, law enforcement observations from the field can be difficult to reconstruct, explain, or defend.

That gap between suspicion and proof is now the focus of a collaboration between researchers at the University of Virginia School of Data Science and industry partner AINA Tech, who are working together to develop artificial intelligence systems designed not just to detect trafficking, but to support cases that can hold up under legal scrutiny.

“High-stakes AI is useless if its logic can’t survive a legal cross-examination,” said Kimberly Adams, Co-Founder and Chair, who leads the strategic architecture of AINA Tech. “From the beginning, our question wasn’t how do we build a smarter model. It was how do we build a system that can withstand interrogation.”

The approach centers on what researchers call “defensibility,” a framework that requires every AI-generated signal to be traceable, explainable, and verifiable. In practice, that means documenting how data was collected, how it was processed, and how a model arrived at a conclusion — creating a chain of reasoning that can be examined in court.

This is a departure from most current AI systems, which prioritize speed and pattern recognition but often cannot explain their outputs in a way that meets legal standards.

“What I mean by defensibility is that we can trace the output,” Adams said. “From the original question being posed … to why they got that answer … so that I as an end user can defend my position.”

Human trafficking presents a uniquely difficult challenge for such systems. The crime is frequently mischaracterized or overlooked, in part because it overlaps with other offenses and lacks clear visual or behavioral markers. In the United States, cases are often prosecuted under charges such as prostitution or labor violations, even when trafficking is the underlying crime.

“There is no stereotype that you can rely on,” Adams said. Victims may appear to be living ordinary lives, making it difficult for observers to recognize exploitation when it occurs.

The consequences of those misconceptions can be significant. Misclassification can lead to missed opportunities to intervene, and weak or incomplete evidence can result in cases being dismissed or downgraded.

“It becomes extremely difficult to prove in a court of law,” Adams said, noting that victims themselves may not recognize or report their own exploitation.

The AI system under development is designed to address this problem by identifying patterns across disparate data sources — including text, images, and contextual information — and flagging combinations of signals that may indicate trafficking. But detection alone is not the goal.

“If you can’t defend that signal, then it’s just information,” Adams said.

Instead, the system is being built to provide a detailed explanation of how each signal was generated, allowing investigators and prosecutors to understand and justify their decisions. This includes maintaining a record of the original inputs, the transformations applied to them, and the reasoning behind the model’s output.

For law enforcement, that could mean the difference between acting on a vague suspicion and building a case that meets the threshold for probable cause.

The system is also designed to support, rather than replace, human judgment. Researchers emphasize that AI should not make legal determinations on its own, particularly in high-stakes contexts where errors can have serious consequences.

“We cannot have a computer decide reasonable suspicion,” Adams said.

Instead, the goal is to surface signals that might otherwise be overlooked and provide the context needed to evaluate them. For example, a conversation that appears benign on its own may take on new significance when analyzed alongside other data points, such as patterns in language or repeated associations between individuals.

Building such a system requires careful attention to the data used to train it. According to Shweta Jain, AINA Tech’s co-founder and technical architect, ensuring that data is representative and accurately labeled is essential to capturing the nuances of trafficking cases.

“In order to train a model to recognize these nuances, we always need a human in the loop,” Jain said. “The data should be representative, it should be diverse … and it should have an accurate provenance.”

That emphasis on human oversight reflects broader concerns about bias and reliability in AI systems, particularly those used in law enforcement. By involving subject matter experts at each stage, the team aims to ensure that the system’s outputs are both accurate and grounded in real-world expertise.

The work comes as policymakers and regulators increasingly focus on the risks associated with AI in high-stakes applications. Requirements for transparency and accountability are becoming more stringent, especially in areas involving public safety and civil liberties.

“We’re starting to see the regulations,” Adams said. “Defensibility can’t just be an added feature. The liability is too high.”

Researchers expect that demand for explainable, defensible AI will continue to grow as institutions seek systems that can support decisions under scrutiny.

“The future is going to be technologies that are making more responsible use of AI,” Jain said. “We may be moving toward more explainability for every single case where AI is used.”

For human trafficking cases, where the difference between suspicion and proof can determine whether victims are identified and perpetrators are held accountable, that shift could be consequential. AI systems that not only detect hidden patterns but also explain them may help bridge a gap that has long limited the ability to pursue justice.