New UB research addresses privacy leaks in digital conversations

As the use of large language model agents, such as virtual assistants or chatbots, increases at work, at home and in schools, users may be sharing more private information than they realize, according to new research from the School of Management.

The researchers’ paper, accepted by the leading data science and AI conference KDD 2026, examines how privacy-detection models can be developed for real-world AI interactions. Seemingly harmless prompts and casual questions can reveal such personal details as intentions, travel, opinions and personal circumstances — especially when the interactions are tied to a logged-in account.

“Proactively alerting users to potential privacy exposure during interactions with LLMs has become an urgent and practical need,” says study co-author Shaojie Tang, professor and chair of the Department of Management Science and Systems, and faculty director of the Center for AI Business Innovation in the School of Management. “Because these tools are tied to user accounts, every prompt is connected to a real person.”

To explore whether users could be warned before unintentionally sharing private details about themselves, researchers constructed the first large-scale multilingual dataset of real AI conversations using nearly 250,000 user queries and more than 150,000 annotated privacy phrases and then used an advanced AI system to analyze the data step-by-step.

First, the system determined whether a message revealed anything private about the user, such as plans, preferences or work details. Next, it identified the exact words in the message that caused the privacy risk and generated a brief explanation of what those words revealed. The final dataset was used to test smaller, privacy-friendly AI tools that could one day warn users before they hit send that their prompt might share more than they intend.

“There is a need for AI systems that can warn users in real time before the information is shared,” says Tang. “With the right training, smaller AI models running on personal devices can detect privacy risks more effectively than much larger cloud-based systems, offering users more control over how much they are sharing.”

The findings provide a foundation for future research into practical privacy detection on the devices people use daily to better safeguard personal information.