The development and proliferation of generative artificial intelligence systems that can generate text, images, audio, and code on scale has spread into society in an unprecedented way, and has given rise to a heavily fragmented debate on the social implications of this technology. That speech is usually either a list of harms or a list of benefits; it doesn't often present a systematic method to evaluate particular risks in particular contexts. While this paper introduces a conceptual framework for social impact assessment of generative AI, it focuses on three areas where impacts are most significant and where they are most discussed: misinformation, labor, and equity. The framework defines the principal risks and the mechanisms that create them for each domain and a common set of cross-cutting dimensions (likelihood, severity, reversibility, distribution, timescale, and tractability) to which any candidate risk can be brought to evaluate it. It combines these with domain-specific indicators to anchor assessment in observable evidence, a matrix between the level of severity and likelihood to establish mitigation priority and a mapping of governance and mitigation levers to actors placed to act on them. The framework deliberately aims to be evaluative as opposed to alarmist or dismissive; it is meant to balance potential risks and benefits and to differentiate between potential and actual harms in the three domains, acknowledging vastly different empirical evidence from one to another. It brings a common language and a replicable process for researchers, developers and policy makers to get beyond the general worry or excitement about generative AI to specific evidence-based and comparable evaluations of the social risks of these tools.