Aamu
An independent European lab for understanding how advanced AI systems fail — and turning those failures into tests the entire ecosystem can use.
Approach
We treat AI incidents as evidence, not anecdotes. Each one becomes hypotheses about the underlying failure mechanism, tested under controlled conditions and turned into evaluations anyone can run.
- Incident
- Hypothesis
- Experiment
- Evaluation
- Evidence
Work
- Incident-derived evaluations
- Open, reproducible tests built from real failures, run across as many models as we can reach.
- Technical incident reports
- Postmortems of significant AI failures, focused on causes rather than blame.
- A failure taxonomy
- A shared language for how AI systems fail, organized by mechanism rather than by harm.
Principles
- Independent
- Not an arm of any company, regulator or cause. Credibility comes from technical rigor.
- European, globally relevant
- Being European is a source of independence. The research questions are global.
- Technical first
- Policy input follows the evidence, never the other way round.