An AI-powered diagnostic tool deployed in rural clinics across Sub-Saharan Africa has helped identify tuberculosis cases that human clinicians missed, potentially saving thousands of lives. The same type of AI technology, deployed in a different context, has been used by authoritarian governments to conduct mass surveillance of their citizens. A single AI system for optimizing energy grids has reduced carbon emissions at one power company by 15%, while the data centers training the next generation of AI models consume as much electricity as some small countries. AI is not inherently good or bad. Its impact depends on who builds it, how it is deployed, and who benefits. A comprehensive impact assessment looks at the full picture: direct and indirect effects, short-term and long-term consequences, intended uses and foreseeable misuses.