DONNA
AI safety
How DONNA configurations are tested, safeguarded and deployed responsibly.
Testing procedures
Testing to market standards
Every model version is systematically tested for safety before release and during ongoing operation.
Pre-release testing+
Every model version is tested for bias, toxicity, hallucinations and robustness before publication. Red-teaming procedures identify weaknesses before they reach production.
Ongoing monitoring+
In the gateway, requests are anonymised and continuously analysed for safety-relevant patterns — bias drift, distress signals, unexpected behaviour. Anomalies trigger internal reviews.
Benchmarks+
Standardised benchmarks measure performance and safety across model versions. Results are documented and feed into the further development of the DONNA platform and its configurations.
Data protection
Your data belongs to you
Clear rules for how your data is handled.
No model adaptation with customer data
Your data never feeds into our model adaptation. The data foundation and adaptation are based exclusively on licensed and public sources.
No storage
User data is not stored — except where a legal obligation exists. Your requests belong to you.
Inference in the EU
All requests are processed primarily in European infrastructure. Details on locations and fallback infrastructure can be found in the respective model documentation.
GDPR-compliant
Fully compliant with the General Data Protection Regulation. Transparent processing, clear legal bases.
Anonymised monitoring
Gateway monitoring exclusively on anonymised data. No conclusions can be drawn about individual users.
Data foundation
A responsible data basis
DONNA models are based on a curated data foundation drawn from proprietary, freely accessible and licensed third-party sources.
This data foundation is checked for personally identifiable information (PII) and copyrighted content. Identified PII is removed or anonymised before use.
No data behind paywalls, with explicit opt-out mechanisms or without appropriate authorisation enters the data foundation.