Manifesto
Why we called it Guardian
A model built to protect, not to maximise engagement — the principles behind the name, and what they cost us to keep.
Research
Research
Studies, engineering notes, and the principles behind our models: how a European AI company thinks about safety, honesty, and privacy. Releases live on Models; announcements on News.
Honesty
Honest by default: teaching a model to say “I don’t know”
Why a calibrated “I’m not sure” beats a confident wrong answer — and how we treat honesty as a design principle, not an afterthought.
EU AI Act
Designing for the EU AI Act: transparency and explainability
How we think about transparency, documentation, and explainability in regulated AI environments — built in from the start, not retrofitted.
Privacy
Privacy by architecture: minimal data, by design
Minimal retention, transparent processing, and encryption in transit and at rest — how privacy is built into the system instead of bolted on.
Multimodal
Understanding speech across 113 languages
What our speech work taught us about accents, code-switching and staying robust where it is hardest.