A broad goal in natural language processing (NLP) is to develop a system that\nhas the capacity to process any natural language. Most systems, however, are\ndeveloped using data from just one language such as English. The SIGMORPHON\n2020 shared task on morphological reinflection aims to investigate systems'\nability to generalize across typologically distinct languages, many of which\nare low resource. Systems were developed using data from 45 languages and just\n5 language families, fine-tuned with data from an additional 45 languages and\n10 language families (13 in total), and evaluated on all 90 languages. A total\nof 22 systems (19 neural) from 10 teams were submitted to the task. All four\nwinning systems were neural (two monolingual transformers and two massively\nmultilingual RNN-based models with gated attention). Most teams demonstrate\nutility of data hallucination and augmentation, ensembles, and multilingual\ntraining for low-resource languages. Non-neural learners and manually designed\ngrammars showed competitive and even superior performance on some languages\n(such as Ingrian, Tajik, Tagalog, Zarma, Lingala), especially with very limited\ndata. Some language families (Afro-Asiatic, Niger-Congo, Turkic) were\nrelatively easy for most systems and achieved over 90% mean accuracy while\nothers were more challenging.\n