Protecting sensitive data at the point of use.

The questions we asked

  1. Warrant

    How reliably can AI identify sensitive information, and how should it communicate uncertainty?

  2. Legitimacy

    Who should control the handling of sensitive information, and when is automatic protection justified?

  3. Legality

    What does the GDPR require when personal data is processed locally or through cloud services?

  4. Practical context

    How can automated safeguards and informed user choice work together in everyday AI use?

Privacy protection that preserves human agency

An AI prompt can contain more than a question. Personal details, confidential information, or internal documents may reach external services before users fully recognise what they are sharing. Protecting such information requires safeguards at the moment of interaction.

Integrated into aihorizon R&D’s responsible AI learning space, the Security Module uses a locally running AI model to analyse prompts and files for potentially sensitive content. Users can choose between automatic and semi-automatic modes. In semi-automatic mode, the module provides an analysis and feedback, allowing users to decide whether their data should be processed by a local or cloud-based model.

The Security Module analysing a prompt in the learning space
The Security Module in the learning space

Grounded in our Responsible AI Framework, the project connects privacy protection with human agency, proportionality, and accountability. It also addresses the reliability and limitations of automated detection, alongside the data-protection requirements governing the processing and disclosure of personal information.

The aim is not to replace human judgment or suggest that local processing eliminates every risk. Instead, the module brings potential privacy risks into view before further processing, making data protection and informed choice part of everyday AI use.

Core design

Privacy and user control

Local analysis before the next processing decision

Automatic and semi-automatic