ERC Grant to Bring Privacy to LLMs on Your Phone

Reconciling privacy and functionality when using Large Language Models (LLMs) such as ChatGPT on personal devices – this is what Professor Dan Alistarh, PhD student Eugenia Iofinova and colleagues are working on at the Institute of Science and Technology Austria (ISTA). To help move their technology towards commercialization, the researchers have just been awarded an ERC Proof of Concept grant worth 150,000 Euro.

The grant, announced today, supports their project titled “PersonalAI: Efficient and Private Text Personalization with Large Language Models.” It aims to enable people to use LLMs on their own devices without sharing their personal data.

Dan Alistarh explains why this is needed in a world where more and more are turning to ChatGPT, Gemini, Copilot, Claude and other LLM-powered chatbots: “To process the user’s prompts, data is sent to the cloud. This has created a critical tension between functionality and privacy,” he says. “While users increasingly rely on these services for personal and business communications, such cloud-based systems require surrendering control of sensitive data to third-party companies, often outside EU jurisdiction.”

That’s exactly the problem Alistarh and his team are tackling. “I am happy to be able to lead this project with a talented team by my side, including Eugenia Iofinova, who designed a tool to train a system on a user’s writing style, and two expert advisors from MIT,” Alistarh says. “By turning scientific discoveries into a product people can use, we hope that PersonalAI will be setting a new standard where users can keep their data secure and still enjoy top-quality LLMs.”

Four steps combine privacy & performance

Eugenia Iofinova explains how a four-step system will bring about solutions. “First, the system figures out if a task is simple or complex. It handles simple tasks locally on the device using models tailored to the user. If the task is complex and can be cleaned of personal data, the system does so before sending the request to the cloud, then personalizes the returned text to fit the user’s needs and style,” Iofinovasays. “This makes PersonalAI a great fit for the European AI market, where keeping data private and following data protection laws like GDPR is very important. It’s the first solution that offers state-of-the-art personalization but with full data privacy.”

Alistarh’s team now has 18 months to use the 150,000 Euro and work towards these goals for their project.The European Research Council’s Proof of Concepts (PoC) grants are designed to help grantees bridge the gap between the results of their pioneering research and the early phases of its commercialization or societal application. They are available only to researchers who currently hold, or have previously been awarded, ERC frontier research grants. Alistarh received an ERC Starting Grant in 2018, as well as another PoC grant in 2024.

2025: Two out of eight Austrian ERC PoC recipients based at ISTA

The ERC’s 2025 work program included two rounds of PoC calls. PersonalAI is one of five ERC PoC grants awarded to Austria-based researchers as part of the second call announced today.

The winners of the first call, announced six months ago in July 2025, also included a project from ISTA. Professor Johannes Fink and his team obtained funding for their project “Integrated optical coupling for low-loss electro-optic interconnects” (CoupledEOT) to move towards commercialization of a key technology for a future optics-based quantum internet. At the time, CoupledEOT was one of three Austrian projects to receive grants.