Â鶹ÒùÔº

In the discussion around artificial intelligence (AI) and automation, Sanmay Das has found one constant: While many people believe that some jobs can be replaced by AI, they believe that their own job is far too nuanced and complex to be handed over to the machines.

That dynamic is at the center of his research at the Sanghani Center for Artificial Intelligence and Data Analytics, where he is the associate director of AI for social impact. He and his students study both the potential and the pitfalls of delegating decision-making in the social services field, particularly around homelessness.

Caseworkers are often overworked and must balance a set of competing objectives. They frequently need to decide who might need more intensive interventions, such as counseling and certain health resources, and who might be able to succeed with less intensive resources.

There's also the push and pull of addressing immediate needs while ultimately trying to set people on a path to more stable, long-term housing, all on tight, fixed municipal budgets.

"The question we've been thinking about for a long time is, "How do I decide how to do this kind of prioritization? Who gets allocated which resources?'" said Das. "And then, 'What are the implications of this, both in terms of what it means from a distributive point of view, but also in terms of efficiency? Are we doing the best job in getting people out of homelessness?'"

To that end, Das has conducted several studies that have yielded interesting results. Several of his initial projects looked at how well AI could predict the ways different types of households would respond to different interventions and how that information could be used to optimize allocations. His recent work, which he will talk about at Tech on Tap, looks at just how much value caseworkers bring to that intricate decision-making.

AI vs. human decision-making

In the , the researchers presented intervention scenario options for people facing homelessness to non-experts. Given limited resources, some people focused on addressing the most vulnerable, while others prioritized those they believed would experience the best outcomes.

Interestingly, when people were also shown an AI prediction on how well their subjects would do given the intervention they chose, they tended to shift toward allocating resources to those they believed would experience the best outcomes.

In a , Das and his team looked at real data from caseworkers' decisions, attempting to understand the outcomes when they made discretionary judgments about homeless service allocation in certain cases.

He found that despite heavy workloads and all those priorities to juggle—a strategic allocation of limited resources for timely interventions and a complex mix of federal guidelines, community, and local agency policies—their discretionary judgments improved outcomes on both ends of the spectrum.

"Interestingly, the data seems to reveal that human caseworkers are able to identify something that is hard to pick out in general," said Das.

"They are pretty consistent in their decision-making, and when they are making their discretionary decisions, they're making them in a way that helps the households that are getting allocated more intensive resources, but doesn't hurt too much the households that are getting less intensive resources than you'd expect."

In the , which appears on the preprint server arXiv, Das and the team asked a group of large language models (LLMs) to make these same kinds of judgments about homeless intervention services.

What they found was that not only did the decisions look much more like the non-expert, lay-person choices from the first experiment, but that they were also extremely variable, with virtually no correlation with existing scoring systems used in prioritization that assess vulnerability and medical frailty.

Taken together, the work that Das and his students and colleagues are doing paints a picture of a technology with potential, but one that is not ready to simply replace human judgment, especially for such an important task.

"It's not clear that these LLMs are ready to just be 'plugged in' as replacements for the kinds of jobs that humans are doing," said Das.

That's at the core of Das's work in this emergent moment of generative AI.

The value of the human touch

"What Sanmay does in a lot of his projects is to really think about people and impacts of automated decisions on their life or work, and what happens if you get them wrong," said Naren Ramakrishnan, director of the Sanghani Center. "That's the most crucial aspect."

One of the required courses for a graduate certificate in data analytics at Virginia Tech's Institute for Advanced Computing is called Ethics and Professionalism in Data Science. It includes examples of experiments that have gone wrong, including areas such as predictive policing, which can be based on inequities of historical data and create a reinforcing loop. Breaking down the ways in which AI technology can have those blind spots and where human touch is still required is a crucial focus.

"There needs to be more public understanding of technology," said Ramakrishnan. "You don't do anyone any favors by making it look magical or mysterious."

That said, Das is currently using machine-learning techniques to try to study the effective decision-making that the social workers made in his second experiment to try to better quantify their motivations and analysis. He hopes technology can be a tool to help lighten their work, while allowing their expertise to continue to guide intricate and complex choices.

"AI is super useful and it can do a bunch of different things, but how you structure that interaction has all kinds of implications on the kinds of judgments and decisions you're going to get at the end of the day," he said. "And that can have significant impact on people's lives."