Two scholars will be extending their work in artificial intelligence at the MIT Initiative on the Digital Economy (IDE).
Mohsen Bahrami (pictured above) joined the IDE in April as a Research Scientist working under the supervision of IDE Director Sinan Aral. Previously, Mohsen was a Postdoctoral Researcher at the MIT Media Lab and a Research Scientist at the MIT Institute for Data, Systems and Society (IDSS).
Joseph Emmens joined the IDE’s AI, Quantum and Beyond research group in July as a Postdoctoral Associate, and he’ll be working with Neil Thompson.
Both Mohsen and Joseph spoke recently with IDE contributing writer Jennifer Zaino. This account of their conversations has been lightly edited.
Q: What’s the connection between your new projects and work you did previously at MIT?
Mohsen Bahrami: My previous research was more focused on computational social science and urban analytics. I studied human behavior and how people make decisions in different situations, which has implications for public decision makers, urban planners and business owners. For example, in a paper I recently co-wrote, NAICS-Aware Graph Neural Networks for Large-Scale POI Co-visitation Prediction: A Multi-Modal Dataset and Methodology, we introduced a graph neural network — an advanced machine learning technique — that tries to predict the co-visitation count, or patterns, between any two business brands in different states and cities across the United States.
For example, if you’re Walmart and you don’t have any business locations in one state, you can use this modeling approach to predict which brands are likely to be most co-visited with your brand. Then, based on that, you can decide on the location of your store in that state. This will help increase the patronage of your customers.
Of course, machine learning and the models we had been using in our work are all part of AI. The technology has been around for decades. But now, with the advances in Large Language Models (LLMs), you can give the user more flexibility and more options in decision intelligence systems.
I’m working mainly on integrating LLMs into decision intelligence systems.
Q: Can you explain more?
In a session I’ll be giving at the 2025 Informs Annual Meeting in October, I’ll discuss LLM-Enhanced Flexible and Dynamic Analyses for Model-based Decision Optimization. With the rise of LLMs, we have a new approach to the design of decision intelligence systems. It’s an LLM agent with a user interface, fine-tuned on relevant models, code examples and data documentation. The user can ask for different models, different types of result presentations, and the interpretation of the model and the results. The integrated AI part of the model does it all.
Q: What else are you looking ahead to?
I’m interested in applied analytics and how we can use these AI advances in real-world applications. How can we use LLMs to try to make the modeling I’ve done with prior research more applicable to real-world problems? Recently, I’ve been collaborating with Intelmatix, and specifically their research labs on related topics. Intelmatix is a deep tech AI company founded by a group of MIT technologists and alumni that provides decision intelligence platform services.
One of the most interesting topics is related to using multiple AI agents in different units of a supply chain to help in the decision-making process.
I’m also interested in understanding the human interactions with AI applications from a psychology perspective. Conversational AI is moving very fast. Gemini can see! Our interaction level with these systems and applications is increasing exponentially. I believe human psychology should be a center of attention, since we still don’t know the short- and long-term effects of this on our physical and mental health.
Q: Joseph, your prior research delved into teams, their contributors and the impact of those contributions on science and innovation. Where did AI enter the picture?
Joseph Emmens: I developed a number of projects, one in which I studied a firm’s choice to automate its production process as something of an umbrella for topics including AI. Still, the main focus of my work was on the research teams. I wrote a model to disentangle each of the team members’ contributions to a project, to help determine statements like, “If we break up a team, and then we rearrange each part, do we get better or worse research?” These papers asked: What is the impact of working in a well-established research area on the science and innovation those teams produce?
In computer science, for example, there are a lot of established paradigms you can build on, which can be extremely useful. But it’s also harder to be novel. Removing a team member who’s from the core of that field may allow you to release yourself from that paradigm. If you’re left with more inventors or scientists coming from upcoming fields, then maybe you’ll be able to do something a bit more novel, when can help create a breakthrough.
But we also found that if you are working on a really new idea — say, something in quantum computing — then actually adding somebody from an established field can be a benefit.
Q: How will you connect that research with AI?
I want to merge my first two works by taking a scientific production process and looking at what people can add to it. It’s going to be a study that focuses on how laboratories use resources, which could be scientists, data, machinery or experiments. We’ll see how those different components are combined to create science. While we can combine all those with computational power to create science, what if AI can substitute for some of those resources? What might happen there?
Q: Where do you think this could lead?
One thing I like about studying AI in science is that the potential for AI to be positive in this field is much more obvious than in other places. There’s a lot of fear around AI, which is kind of warranted.
But in science, people are genuinely quite positive about AI. They see so much potential for AI to help with searching, suggesting solutions and finding new problems. It’s the bright side of AI.
The point is seeing what AI can do. Because AI is good in some scientific fields, those fields might grow a lot. But because AI might be weaker in other fields, they might not grow. And what will that impact? If, say, there’s a boom in chemistry, but physics gets left behind, that maps into jobs, into employment and into our understanding of things.