Costello Research Machine Learning

Can machine learning make the world a fairer place?

Katelynn C Hipolito — Wed, 22 Apr 2026 15:08:00 +0000

Can machine learning make the world a fairer place? Katelynn C Hipolito Wed, 04/22/2026 - 11:08

A paradox hovers over our increasingly AI-dependent world. On the one hand, artificial intelligence can make the world a better place (or so we’re told). On the other hand, algorithms have no imagination or consciousness, and thus can know only the status quo—as reflected in the data they are trained on. And our current world is far from perfectly meritocratic or fair.

Jingyuan Yang. Photo by Jeffrey Porovich/Costello College of Business.

Jingyuan Yang, assistant professor of information systems and operations management at Costello College of Business at 91��, suggests that the paradox is compounded by conventional thinking around AI. “The standard view is that fairness is a tax on efficiency. The way conventional systems are structured, fairness checks are added almost as an afterthought that is assumed to negatively impact system performance,” she says.

Is the “better,” optimized world of AI destined to replicate, or perhaps even exacerbate, existing inequalities? Yang’s ongoing research—in collaboration with Pengzhan Guo of Duke Kunshan University and Keli Xiao of Stony Brook University—points to an appealing alternative. It uses AI systems as a proving ground for a theorized “fairness-performance complementarity”—the idea that, under certain conditions, fairness and performance reinforce one another.

“Our 'fairness-by-design’ framework utilizes reinforcement learning, which is a type of machine learning (ML). But unlike most machine learning algorithms, ours includes multiple agents competing for finite resources in a dynamic environment, not a static one,” Yang says. “That makes our paradigm much more structurally similar to many real-world environments in which various people compete over time for finite resources.”

Fairness was integrated in two stages. First, the framework was designed to “nudge” high-performing agents towards exploratory choices that might maximize their rewards. As Yang explains, “In this framework, high-performing agents are held in an exploratory mode for longer, while lower-performing agents settle into stable paths sooner.” Second, options that were abandoned as a result of agents’ reward-seeking behavior were redistributed, with lower-performing agents getting first crack at the best opportunities.

As Yang summarizes, "The exploratory activity of the high performers releases opportunities that the system channels down toward the weaker performers. Theoretically, this increases fairness while retaining individual choice and without constraining performance.”

“Our ‘fairness-by-design’ framework utilizes reinforcement learning, which is a type of machine learning (ML). But unlike most machine learning algorithms, ours includes multiple agents competing for finite resources in a dynamic environment, not a static one. That makes our paradigm much more structurally similar to many real-world environments in which various people compete over time for finite resources.”

—Jingyuan Yang, assistant professor of information systems and operations management at Costello College of Business at 91��

To test out the framework, the researchers used a data-set comprising detailed information on the job histories of 6.5 million professionals across a 20-year timeframe. “In the real-world data, we see a high degree of disparity, without very much redistribution of elite opportunities from relatively advantaged to disadvantaged employees,” Yang says.

The algorithm converted the real-world job information into opportunities offered to hypothetical agents. The resulting career paths were analyzed in terms of both performance and fairness. Performance was defined by aggregate rewards earned by all agents across all periods. Fairness was defined by the degree to which initial performance disparities were resolved over successive decisions.

The “fairness-by-design” framework’s results—for both fairness and performance—were better than those of eight alternative ML methods drawn from three different methodological families.

The researchers also adjusted the system to account for people’s changing preferences. Early-career professionals tend to value employer reputation and advancement potential; in late career, rewards pertaining to job stability and security are more salient. Even with these restrictions implemented, the framework functioned as intended—improving the average quality of overall career paths while fueling upward mobility.

In a follow-up study utilizing the New York Yellow Taxi Trip record database, the framework was tasked with generating route recommendations to hypothetical “agents,” i.e. cab drivers, with varying performance records. In this domain, the choice-set was much smaller (263 locations, as compared to 4,282 companies), and the timeframe far shorter (two hours as opposed to 20 years). As with the career-planning example, the taxi study found that more equitable distribution of high-quality routes led to higher average income per minute for the system as a whole.

“Because the framework proved adaptable to different domains and agent preferences, we think it could be used in future as a governance mechanism for a variety of AI contexts,” Yang says. Health care scheduling, course registration in higher education and provision of digital services are a few areas Yang sees as likely candidates.

While emphasizing that her research is still ongoing, she argues that it poses a serious challenge to standard ways of thinking about AI. “Our formal proof establishes the conditions under which fairness and performance reinforce each other, and our experiments show those conditions are achievable in realistic settings. That gives our work both theoretical and experimental grounding," Yang says.

Learn more about AI at George Mason

Dive into Research at George Mason

In This Story

Jingyuan Yang

Topics

ISOM Faculty Research

Costello College of Business Faculty Research

Costello College of Business News

Costello Research Artificial Intelligence

Costello Research Machine Learning

Costello Research Digital Platforms

ESG - Costello

A.I. and Innovation - Costello

George Mason researcher is using AI to identify human trafficking hot spots

Katelynn C Hipolito — Mon, 29 Sep 2025 18:20:46 +0000

George Mason researcher is using AI to identify human trafficking hot spots Katelynn C Hipolito Mon, 09/29/2025 - 14:20

Illicit massage businesses (IMBs) run by human trafficking rings are rampant in the United States. A 91�� professor has helped build what may be the best artificial intelligence (AI)-driven tool to root them out.

Human trafficking rings are at their most dangerous when they masquerade as legitimate commercial activity. IMBs are one of the most common ways in which exploitive networks operate in plain sight.

The Network, an anti-trafficking nonprofit, estimates that there are more than 13,000 IMBs active in the United States, raking in annual total revenue of more than $5 billion.

Abhishek Ray

“You are stuck in a massage business. You’re not allowed to go out,” says Abhishek Ray, assistant professor of information systems and operations management at the Donald G. Costello College of Business at George Mason, describing the plight of IMB workers. “Your passports are taken away, and you’re supposed to do a certain amount of business every day and give the money to the trafficker. It’s a really abhorrent form of abuse.”

Ray is one of a growing group of researchers exploring how various forms of AI could help resource-constrained law enforcement agencies differentiate between IMBs and the legitimate enterprises they try to mimic. His ongoing research using graph neural networks has yielded more promising results than rival approaches, when put to the test in a recent experiment.

His co-researchers on the IMB project are Lumina Albert and Swetha Varadarajan of Colorado State University.

According to Ray, “Graph neural networks are just a fancy way of saying that if I get a graph of a city or locality at one point in time, and I add data to it, can I predict future patterns on this graph if I know the past?”

This approach made sense for detecting IMBs, because try as they might to appear above board, they have geographical needs that conventional businesses don’t. “IMBs don’t allow their trafficked employees to go out of the parlor,” Ray says. “But since they’re humans, they need sustenance. They have to be near groceries, gas stations, where they can get stuff and come back.”

The researchers combined several graph neural networks into a framework called IMBWatch. The training data-set for IMBWatch comprised publicly available information such as online customer reviews, arrest and raid data for known IMBs, and advertisements from websites promoting illicit activities (e.g., the infamous Backpage). The result, in essence, was a series of snapshots mapping the evolution of the IMB network in a given city or county over a period of time. This could then be overlaid on geographical maps to tease out hidden patterns.

To gauge IMBWatch’s relative performance, the researchers let it loose on a testing data-set alongside four other AI models, which were not as sensitive to the nuanced interplay of spatial and temporal factors. Of the five models, IMBWatch provided the most accurate, precise and informative predictions. In other words, it outperformed the others at spotting IMBs among a larger mass of local businesses.

“Graph neural networks are just a fancy way of saying that if I get a graph of a city or locality at one point in time, and I add data to it, can I predict future patterns on this graph if I know the past?”
— Abhishek Ray, assistant professor of information systems and operations management

While encouraging, these outcomes require further confirmation with a larger data-set. “IMBWatch was trained on data from Georgia and Louisiana, not the entire United States,” Ray says. “These were small, manageable data-sets, but we will now scale up to major states such as New York and California.”

The researchers are also looking at enhancing IMBWatch with data related to how workers end up wandering into trafficking webs. These might include “proximity to hospitals, religious places, etc. because a lot of times people are coerced by religious compulsions, or because they’re pregnant and need some care,” Ray says.

This is not Ray’s first foray into the field of AI-fueled anti-trafficking. Previously, Ray co-developed a model for improving machine learning-based detection of human trafficking activity at transit stations and on fishing vessels.

However, law enforcement agencies and other pertinent stakeholders (e.g., business owners) are often wary of adopting AI-based solutions, due to a lack of trust in the technology. Ray and his co-researchers are currently devising a framework that will clarify how these stakeholders can work together with tech experts and, perhaps most importantly, human trafficking survivors to make the best possible use of AI.

“This qualitative piece is required to make sure that people who are on the sidelines, on the fences about using this actually start using it, because that’s the need right now,” Ray says.

Learn more about Artificial Intelligence at George Mason

In This Story

Abhishek Ray

This content appears in the Spring 2026 print edition of the Mason Spirit Magazine with the title "Using AI to Identify Human Trafficking Hot Spots."

Nonprofits are in trouble. Could more sensitive chatbots be the answer?

Jennifer Anzaldi — Tue, 18 Mar 2025 14:48:25 +0000

Nonprofits are in trouble. Could more sensitive chatbots be the answer? Jennifer Anzaldi Tue, 03/18/2025 - 10:48

In today’s attention economy, impact-driven organizations are arguably at a disadvantage. Since they have no tangible product to sell, the core of their appeal is emotional rather than practical—the “warm glow” of contributing to a cause you care about. But emotional appeals call for more delicacy and precision than standardized marketing tools, such as mass email campaigns, can sustain. Emotional states vary from person to person—even from moment to moment within the same person.

Photo by Getty Images

Siddharth Bhattacharya and Pallab Sanyal, professors of information systems and operations management at the Donald G. Costello College of Business at 91��, believe that artificial intelligence (AI) can help solve this problem. A well-designed chatbot could be programmed to calibrate persuasive appeals in real time, delivering messaging more likely to motivate someone to take a desired next step, whether that’s donating money, volunteering time or simply pledging support. Automated solutions, such as chatbots, can be especially rewarding for nonprofits, which tend to be cash-conscious and resource-constrained.

“We completed a project in Minneapolis and are working with other organizations, in Boston, New Jersey and elsewhere, but the focus is always the same,” Sanyal says. “How can we leverage AI to enhance efficiency, reduce costs, and improve service quality in nonprofit organizations?”

Siddarth Bhattacharya. Photo provided

Sanyal and Bhattacharya’s working paper (coauthored by Scott Schanke of University of Wisconsin Milwaukee) describes their recent randomized field experiment with a Minneapolis-based women’s health organization. The researchers designed a custom chatbot to interact with prospective patrons through the organization’s Facebook Messenger app. The bot was programmed to adjust, at random, its responses to be more or less emotional, as well as more or less anthropomorphic (human-like).

“For the anthropomorphic condition, we introduced visual cues such as typing bubbles and slightly delayed response to mimic the experience of messaging with another human,” Sanyal says.

The chatbot’s “emotional” mode featured more subjective, generalizing statements with liberal use of provocative words such as “unfair,” “discrimination” and “unjust.” The “informational” modes leaned more heavily on facts and statistics.

Over the course of hundreds of real Facebook interactions, the moderately emotional chatbot achieved deepest user engagement, as defined by a completed conversation. (Completion rate was critical because after the last interaction, users were redirected to a contact/donation form.) But when the emotional level went from moderate to extreme, more users bailed out on the interaction.

The takeaway may be that “there is a sweet spot where some emotion is important, but beyond that emotions can be bad,” as Bhattacharya explains.

Pallab Sanyal. Photo provided

When human-like features were layered on top of emotionalism, that sweet spot got even smaller. Anthropomorphism lowered completion rates and reduced the organization’s ability to use emotional engagement as a motivational tool.

“In the retail space, studies have shown anthropomorphism to be useful,” Bhattacharya says. “But in a nonprofit context, it’s totally empathy-driven and less transactional. If that is the case, maybe these human cues coming from a bot make people feel creepy, and they back off.”

Sanyal and Bhattacharya say that more customized-chatbot experiments with other nonprofits are in the works. They are taking into careful consideration the success metrics and unique needs of each partner organization.

“Most of the time, we researchers sit in our offices and work on these problems,” Sanyal says. “But one aspect of these projects that I really like is that we are learning so much from talking to these people.”

In collaboration with the organizations concerned, they are designing chatbots that can cater their persuasive appeals more closely to each context and individual interlocutor. If successful, this method would prove that chatbots could become more than a second-best substitute for a salaried human being. They could serve as interactive workshops for crafting and refining an organization’s messaging to a much more granular level than previously possible.

And this would improve the effectiveness of organizational outreach across the board—a consummate example of AI enhancing, rather than displacing, human labor. “This AI is augmenting human functions,” says Sanyal. “It’s not replacing. Sometimes it’s complementing, sometimes it’s supplementing. But at the end of the day, it is just augmenting.”

In This Story

Siddharth Bhattacharya

Pallab Sanyal

Topics

A.I. and Innovation - Costello

ESG - Costello

Costello Research Digital Platforms

Costello Research Artificial Intelligence

Costello Research Machine Learning

Costello College of Business News

Costello College of Business Faculty Research

ISOM Faculty Research

How machine learning improvements are helping to fight human trafficking

Marianne Klinker — Wed, 22 Feb 2023 13:09:33 +0000

How machine learning improvements are helping to fight human trafficking Marianne Klinker Wed, 02/22/2023 - 08:09

Human trafficking is a global crisis of overwhelming scope. Fortunately, anti-trafficking organizations can use AI to predict the criminals’ next moves–with the help of a 91�� professor.

The fight against human trafficking has a David and Goliath aspect to it. Trafficking rings are a global scourge with an estimated 25 million victims and $150 billion in annual profits. Agencies and NGOs tasked with stopping the traffickers, however, are typically small and under-resourced. Recently, a technological tool has come along to help close this gap–machine learning (ML) algorithms.

Primed with enough relevant data, these algorithms can find hidden clues to how traffickers operate and what they are likely to do in the future. In principle, anti-trafficking agencies can use ML predictions to direct their scant resources to the areas of greatest risk.

Abhishek Ray

Abhishek Ray, an assistant professor in the Information Systems and Operations Management area at the 91�� School of Business, has added his expertise to this effort. With collaborators Viplove Arora (post-doc researcher at SISSA, Italy), Kayse Maass (of Northeastern University) and Mario Ventresca (of Purdue University), Ray developed a model that reduces the guesswork of implementing ML predictions.

As Ray explains it, the model supplies a “layer on top” for improving outcomes of existing ML solutions. Armed with success and failure rates (i.e., true/false negatives and positive trafficking detections) of an anti-trafficking agency, it looks for scenarios in which the agency could achieve the same number of successes with fewer errors.

In a recently published paper, Ray and his co-authors apply their model to two real-life agencies: Global Fishing Watch (GFW), which targets trafficking in the global seafood industry, and Love Justice International (LJI), which monitors transit stations on the India-Nepal border.

The ML solution employed by GFW cross-references a large database of fishing vessel activity, which is obtained via satellite, with law enforcement data on human trafficking. The resulting predictions indicate the vessel types, behaviors, and features that the agency should watch most closely.

But criminals sometimes learn faster than machines. “If someone is caught on a long trawler, and they know they can get caught, they will never go on the long trawler again,” Ray explains. Traffickers’ evolving tactics may not be fully captured in vast, wide-ranging data-sets, such as the ones derived from satellite vessel surveillance. Changes in agency success rates, on the other hand, are closer to the action and may pick up the latest stages of this cat-and-mouse game.

Looking back over agency data from 2012-2018, the researchers found that had their model been used alongside Machine Learning, the agency Global Fishing Watch would likely have discovered more instances of trafficking while committing far fewer false negatives and false positives.

Looking back over agency data from 2012-2018, the researchers found that had their model been used alongside ML, GFW would likely have discovered more instances of trafficking while committing far fewer false negatives and false positives. In addition, monitoring recommendations shifted considerably over the six years. For example, under certain conditions, drifting longline vessels were the most likely suspects for the years 2012-2017. In 2018, however, squid jiggers replaced them and were singled out as among the most suspicious. These shuffling priorities may reflect the traffickers’ changing their vessel of choice to evade detection.

In the case of Love Justice International, Ray’s model could serve as a workaround for constraints such as a shortage of well-trained staff. LJI could input a target outcome into the algorithm, such as a baseline number of positive trafficker IDs, and receive recommendations for how to make the best use of strained resources.

For Ray, the fight against human trafficking is personal. As a child growing up in Kolkata, India, he was cared for by a household employee who, as his parents later found out, had been trafficked when she herself was a child. “I learned this when I was doing my PhD in the U.S.,” he says. “It’s like having a part of your childhood taken away. You came into contact with something criminal and had no understanding.”

“If I could make a contribution towards solving the problem, I’d consider it sort of giving back what I had taken unknowingly.”

To take full advantage of Ray’s “layer on top” ML solution, anti-trafficking agencies would need to create effective systems for collecting and centralizing data. Depending on the context, this might present challenges ranging from resource limitations to a political climate where trafficking data could be used against the already-victimized. “The maximum value from our framework is when it’s acting in real time with real-time data,” Ray says.