Capability without control is not progress.
Guardrails, evaluation and human oversight are built into our research from its first experiment.
Safety is not a separate team that reviews finished work. It is part of every experiment: what a system is allowed to do, how it is tested, who can stop it and how its decisions are recorded. Our governance is aligned with ISO/IEC 42001.
Questions we are researching
- Which evaluations reliably reveal unsafe or biased behaviour before deployment?
- How do guardrails hold up against prompt injection and misuse, in Arabic and English?
- What does meaningful human oversight look like for systems that act, not just answer?
How we approach it
- Guardrails
- Input and output controls, redaction of personal data and policy enforcement on every system.
- Evaluation and red-teaming
- Systems are tested for accuracy, bias and misuse before and after release.
- Human oversight
- People can review, override and stop consequential decisions.
- Privacy
- Personal data is minimised and protected under the UAE PDPL and the DIFC Data Protection Law.