An AI that answers
privacy questions
in plain language.

An average privacy policy takes 15 minutes to read and is written in legal jargon — “data controller”, “legal basis for processing”. Instead of blindly accepting it, you ask a plain-language question — “Will I get marketing emails after I accept this?” — and the system returns the ranked passages from the actual policy and legislation that answer it. You can add your own sources, and escalate a hard question to a legal expert for a small fee.

The paper breaks the task into two: cutting a policy into clean sentences (sentence-boundary detection), then matching a question to the right ones (question-to-document matching). Across benchmarks, Stanza segmented legal text best, and — notably — a light, sparse BM25+ model matched or beat far heavier neural retrievers on both accuracy and speed. Simple, fast, and good enough to serve in real time.

Ask a general model to fill in “What does Migros do with my ___?” and it guesses “body” or “life”. A model tuned on privacy text answers “information” and “data”. The paper shows that training on a domain-specific corpus measurably improves comprehension — which is why the tool runs privacy-tuned models (BERT, SBERT, PrivBERT), served over a REST API with a vector database for real-time answers.

The work came out of “Profila Privacy Simplified reloaded”, a 30-month Innosuisse innovation project with HSLU – Lucerne University of Applied Sciences (CHF 654k Innosuisse contribution, CHF 1.14M total). It produced two peer-reviewed papers — the 2023 one won Best Paper at HCI International in Copenhagen — and a working demonstrator that was transferred to Profila.

This tab is the research. It produced a working assistant that answers privacy questions in plain language, transferred to Profila as a demonstrator.