SiRelations is an AI safety and research organization working to build reliable methods for human oversight of AI systems. We believe that as AI grows more capable, maintaining meaningful human agency becomes both more important and more difficult.
Contact usWe treat human oversight as a systematic science, not an assumption. Our work develops empirical methods to measure whether humans can actually exercise oversight — not just whether they're nominally "in the loop."
We translate foundational research into tools that organizations can use. Regulatory frameworks mandate oversight, but provide no methods to assess it. We're building those methods.
We work with policymakers to communicate what our research reveals about the gap between oversight requirements and oversight capabilities.
We're preparing for a world where humans regularly collaborate with systems that exceed human capabilities. The measurement methods we develop today will become essential infrastructure.
We focus on three interconnected areas that together address the oversight challenge.
Foundational investigation into how humans can maintain agency when working with advanced AI. We study discernment — the capacity to remain clear, responsible, and sovereign in AI collaboration. Our research identifies the cognitive and contextual factors that enable or undermine human judgment.
Developing empirical methods to assess oversight capacity. Current regulatory frameworks assume humans can "correctly interpret" AI output and decide when to override — but provide no way to verify this. We build the assessment tools that make oversight measurable.
Translating research findings into actionable guidance for organizations and policymakers. The EU AI Act's Article 14 requires human oversight. Impact assessment frameworks identify risks. Our work addresses what these frameworks leave open: whether designated overseers can actually do the job.
We report what we find, even when the findings are uncomfortable. If oversight isn't working, we say so.
Claims about human oversight should be empirically testable. We build the methods that make them so.
We deliver specific, measurable products — not open-ended consulting. Constraints create precision.
We're building for a future where human-AI collaboration is the norm, not the exception.
The challenge of human oversight affects everyone. Our research contributes to the broader ecosystem.
Develop the science and methods for meaningful human oversight of AI systems — ensuring that as AI grows more capable, humans retain genuine agency in the collaboration.