Researchers and industry developers in artificial intelligence (AI) and natural language processing (NLP) have uniformly adopted a Rawlsian definition of fairness. On this definition, a technology is fair if performance is maximized for the least advantaged. We argue this definition has considerable loopholes, which can be used to legitimize common practices in AI/NLP research that actively contributes to social and economic inequalities. Such practices include what we shall refer to as Subgroup Test Ballooning and Snapshot-Representative Evaluation. Subgroup Test Ballooning refers to the practice of initially tailoring a technology to a specific target group of technology-ready early adopters to collect feedback faster. Snapshot-Representative Evaluation refers to the practice of evaluating a technology on a representative sample of current end users. Both strategies may contribute to social and economic inequalities but are commonly justified using arguments familiar from political economics and grounded in Rawlsian fairness. We discuss an egalitarian alternative to Rawlsian fairness, as well as, more generally, the roadblocks on the path toward globally and socially fair AI/NLP research and development.