Explainable Autofill for Trustworthy Surveys

Universities and research organisations rely on surveys and forms to collect information from

students, staff, and partners. Yet building good questionnaires takes time, and responses can be

incomplete or inconsistent, especially when people must search through documents or media to

answer accurately. At the same time, institutions must protect privacy and comply with strict data

protection rules.

PXL University College proposes a practical, privacy-first solution: a platform with two AI helpers

that improve both the creation of questionnaires and the quality of responses, while keeping users

in control.

• M-Chat helps administrators design better questionnaires, faster. It suggests clear,

consistent questions, applies style guidelines, adds helpful tags, and can express simple

evaluation rules, reducing manual work and increasing comparability across surveys.

• M-Autofill supports respondents with evidence-based suggestions. After a respondent

attaches relevant documents (e.g. text, audio, video, or webpages), the assistant identifies

the most relevant passages for each question, proposes a concise draft answer, shows

where the information came from (citations and highlights), and explains why those sources

were chosen. A lightweight review screen lets respondents accept, edit, or reject

suggestions at any time.

The platform is designed "privacy by default." It can run locally (on-premise) within an institution;

uses strict tenant isolation; minimises data processing; captures consent; keeps audit trails; and

avoids training AI models on personal data. If external AI services are used, this happens only for

well-defined purposes, with minimal data sharing and appropriate safeguards. Suggestions are

always advisory: people remain in charge.

Over six months (Jan-Jun 2026), we will build and pilot a working prototype at PXL (and, subject to

confirmation, with one Belnet-connected partner, e.g. UCLL). We will measure time saved for survey

authors and respondents, and quality indicators such as citation accuracy, completeness, and

edit/acceptance rates. Results will be shared through an online showcase and a live demonstration

at GÉANT's TNC conference. To encourage reuse across the GÉANT community, we will release core

components such as the M-Autofill module, deployment recipes, and reference integrations under a

restricted open-source licence, with clear documentation and APIs for integration into existing forms

or workflow tools.

This project aims to make data collection more reliable, transparent, and efficient for European

research and education. By combining explainable AI with robust privacy and security practices, it

offers immediate, practical benefits to researchers, administrators, IT teams, and data protection

officers, and lays the groundwork for wider adoption across GÉANT-connected institutions.

Requested funding: €50,000; PXL will also contribute approximately €28,470 of its own effort. The

final report and an ex-post follow-up will document outcomes and next steps for community uptake.