Social Computing × AI × Well-being

I’m a PhD student at Michigan. This is my third year, working with Farnaz Jahanbakhsh. In January 2026 I passed candidacy, and a month later I had three minutes to explain my research. I won the Social Impact Engineering award. I study how we can use AI for prosocial outcomes and safety. Most apps are optimized for time spent. I’m more interested in whether people come away better off. I worry that years of scrolling have changed how we concentrate. Quitting isn’t realistic for most of us; we go online for people and ideas. The trouble is what comes bundled with them. A useful video, for example, can be edited to be hard to stop watching. I’m exploring whether AI could keep the content but tone that down. The obvious sign of success would be watching less, but less isn’t always better. Someone might watch longer because they can finally follow something they care about, so minutes watched can’t be the only measure. Platforms decide what’s harmful for everyone at once, and usually remove the whole post. With DIY-MOD, people say what they don’t want to see, and AI changes just that part of the post. In our studies, we saw that hiding a detail can leave people imagining it. Enough context, or a benign replacement, could help them move on. More on the project page. A newer worry is AI that reads for us. When an assistant reads the web for you, someone can try to influence what it reads. That’s a way to persuade you through your assistant. Who should catch that? Telling people to trust AI less just puts the burden back on them.

Before Michigan, I taught at BUET. I was an assistant professor (!) there. I became one around the same time my PhD advisor did at Michigan. That’s less strange than it sounds: in Bangladesh, you can start teaching right after undergrad. I joined BUET as a lecturer, became an assistant professor, and then went back to being a student. In fall 2026, I got to teach again, co-teaching graduate HCI at Michigan. I also spent time at Microsoft, helping computers listen, partly by teaching them to look at people’s mouths. It helps when people talk over each other. Mouth movements can tell a model whose voice to keep. Clear audio isn’t much help if it’s the wrong person. Here’s the paper.

During my undergrad at BUET, I got into design and HCI. One project was a healthcare app for people who couldn’t use apps. Dakter Bari was for people living on alms. They went to shopkeepers they already knew, who used the app for them. So the person using the app wasn’t the person living with its answers. If a hospital schedule was wrong, the patient lost a trip and the shopkeeper lost face. I saw a similar split in ridesharing. We picture an Uber driver as the car’s owner. In Dhaka, many drivers rent, and owners track their drivers constantly through the app. Today the go-between is often AI. It reads, sorts, and answers for us. So with AI, I ask who it really serves, and who lives with its mistakes. Papers: healthcare and ridesharing.

I support Chelsea because of one game. It was Everton away in December 2006, and we were behind. Then Ballack hit this free-kick. We went behind again. Lampard brought us level. Then Drogba did this. It finished Everton 2–3 Chelsea. I’ve been a fan ever since. I may have misunderstood how often that was going to happen.

You should also meet one of my cool friends. Shwetha Rajaram makes great research diagrams. People sometimes mistake them for concept sketches. They’re actually systems she built. One is Privacy Equilibrium, a simulation toolkit for privacy around smartglasses. Your glasses might scan a room to help you get around. What about the people nearby? Someone next to you may not want to be recorded. Whose preferences win? Her toolkit lets researchers set those preferences and see what the glasses would be allowed to sense. More on her website. Álvaro Vega-Hidalgo uses AI to study animal communication. One project asks whether a model can recognize which monkey is calling. The monkeys are wild white-faced capuchins, and recordings are scarce, so the model borrows from birds and humans. Combining what both kinds of models learned helps it tell the monkeys apart. Then there’s the question of what the model is listening to. His toolbox lets experts mark parts of a call and listen to them separately. It then compares their selections with the parts that influenced the model. That shows whether the model relies on sounds experts find meaningful. More on his website.

Pause×