Cambridge AI Safety Hub · Reading Course

AI Verification

A structured reading course on the technical and political infrastructure for verifying claims about AI development. The course moves from motivation, to the research agendas now forming, to concrete mechanisms, and then into project scoping.

Use it as a weekly reading sequence, an independent introduction, or the recommended pre-work for the Hardware Assurance Programme. Progress is saved locally in your browser, so you can move through the units without making an account.

Open a unit, mark readings as done, and come back where you left off.

1 Why this matters 4 readings · motivation and historical anchors

Computing power as a governance targetCore argument International AI Safety Report 2025Risk consensus Why alignment is hardAssurance limits Nuclear verification lessonsHistorical case study

2 What people are pursuing 7 readings · agendas and field map

Technical AI governance open problemsField overview Hardware-enabled mechanismsHardware taxonomy Frontier AI auditingAuditing agenda Agreement design and compute governanceFour further readings

3 How we could do this 23 readings · mechanism families

Hardware-rooted approachesSecure chips, flexHEG, memory Inference verificationStatistical and cryptographic checks Telemetry and detectionNetworks, timing, training rules Attestation and auditTEEs, certification, cloud chokepoints Cross-cutting componentsTraining data, erasure, surveys

4 Project ideation prompts · move from reading to doing

Theory of changeWhere would the result plug in? Scoping a projectSmallest useful experiment Where to take it furtherProgramme, RFDs, supervisors, feedback