Tengteng Ma

I am an assistant professor in the School of Information Systems, Muma College of Business, University of South Florida. I earned my Ph.D. degree in Management Information Systems from University of Illinois Chicago.

I study how generative AI changes user behavior on digital platforms, and I build interpretable AI models, from deep learning architectures to large language models, that assess whether AI systems are treating users appropriately.

CV  /  Google Scholar

Projects

As more people turn to AI companions like Replika for friendship and even romance, a pressing question is who keeps those conversations safe. We built the first public collection of real companion chats labeled for safety risks and tested whether 20 leading AI models can serve as watchdogs. They catch obvious harm well, but they miss subtle manipulation and often mistake harmless role-play for danger.

Ask a chatbot a question and you might get an answer that sounds right but misses what you actually needed. We design a 'sensible' generative AI system that first figures out what the user is really asking, then shapes its response to match, and we show how it works in online education and healthcare.

With endless videos competing for attention, platforms are betting that a quick AI-written summary will help viewers decide what to watch. Studying Bilibili's rollout of these summaries, we find they actually get people talking more, especially on longer videos and on videos people watch to learn something rather than just for fun.

When YouTube hid dislike counts, the goal was to spare creators from pile-ons and encourage them to keep posting. We find it backfired: with the signal gone, viewers engaged less with videos, and creators, getting less back from their audience, posted less too.

On video platforms, creators fold sponsorships into their own videos, and the audience talks back in real time through live comments scrolling across the screen. We use those live reactions to study what makes a self-inserted ad land or fall flat, from whether the creator shows their face and shares personal experience to whether they admit a product's flaws and where in the video the pitch arrives.

A sudden crowd sounds like good news for a live streamer, but chat gets noisier and harder to follow. Using Twitch raids as a natural experiment, we show that bigger audiences can crowd out real interaction, the regulars start to fall silent. and that the right kind of moderation, bot or human, helps hold the conversation together.

For brands on social media, guessing which post will spark a like, comment, or share can feel like trial and error. We design a new artificial intelligence model aims to replace that guesswork with sharper prediction and clearer insight into why people engage.

We systematically evaluate evidence from controlled studies of interventions using virtual humans on their effectiveness in health-related outcomes. The design and implementation characteristics of these systems are also examined.