AI alignment and safety
Insight Lab works on AI alignment and safety from the deployment side. We study how AI objectives drift from human intent inside real organisations, whether human control holds up in daily operation, and what assurance can actually establish about a system.
Five questions we work on. Each draws on direct experience of building, running and assessing AI systems in large, regulated organisations.
How do fixed objectives fail once a system meets real organisational incentives?
Deployed systems are usually given a measurable target. We document where optimising that target diverges from what the organisation intended, and which design choices keep the gap small.
Does the ability to correct or stop a system survive day-to-day operation?
Oversight that works at launch can erode through automation bias, workload and tight integration. We treat correctability as a property of the whole organisation, and study what keeps it intact.
Which principals does an enterprise AI system actually serve?
Boards, operators, users and affected third parties want different things. We map how their preferences are represented, traded off or left out when a system is specified.
What can an audit establish about a system's alignment, and what can it not?
Behavioural tests and certification show how a system acted under test conditions. We examine the gap between that evidence and the claims organisations make on the strength of it.
What happens when thousands of organisations depend on the same few models?
Concentration on a small number of foundation models means one failure mode can appear everywhere at once. We study how that correlated risk can be measured, and who ends up bearing it.
Our team has built and run AI in production and answered for risk at board and ministerial level. That puts us close to where alignment problems first become visible.
We work with researchers, with organisations that deploy AI, and with the people who govern both.
Field evidence from deployed systems for academic and lab researchers, shared under appropriate confidentiality.
An assessment of a deployed system for objective drift, strength of human control and exposure to correlated failure.
Short, evidence-based briefings on alignment and safety risk for boards, regulators and policymakers.
Contact
We welcome conversations with researchers, labs and policymakers working on alignment and safety.