Capability Lens turns Amartya Sen's capability approach into a working method, applies it to generative and agentic AI, and tests it on 44 real-world cases, from a farmer checking prices in her own dialect to an agent built to manipulate a lonely teenager. It ends in 42 recommendations and a theory of change, grounded in an evidence base of 872 screened interventions and 35 national AI strategies,, compared with risk-based AI regulation worldwide.
The scope is global. Where the study refers to AI law, it uses the EU AI Act's risk tiers as a reference model, because many other jurisdictions have drawn on them, not because the EU is the target.
A 16-step interactive lesson that explains every symbol "like you're 5", then in math, then with the nuances.
The general, domain-agnostic formalization: primitives, stages, axioms, five propositions, remaining assumptions and open modelling decisions.
Where generative and agentic AI enter the formula, and whose options grow or shrink.
44 cases scored on eight dimensions, with verdicts and regulatory risk tiers.
All 872 screened interventions, from the UN Activities on AI report, OECD.AI national policies, the practice scan and partner programmes, with the integrated analysis across every source.
thirteen findings and 42 recommendations, with who should act and when.
33 programmes, products and laws already widening capabilities or protecting agency, matched to the recommendations.
How the ten action areas turn into wider real freedom, with assumptions, evidence, loops, indicators and phasing.
35 strategies, policies and action plans from five regions, read through the capability formula.
An internal audit aligned terms, numbers and categories across all tabs (see the crosswalk in the Formal model). Claims carry confidence labels: established (external studies), indicated (patterns in the evidence base), illustrative (the 44 atlas cases, built to explain the framework) and hypothesis (the theory of change). Open work: calibrating prominence scores, verified coding of non-English strategies, harm databases, missing regions, and fielding the survey.