An emerging hallmark across many human diseases - such as cancer, autoimmune and neurodegenerative disorders – is the aberrant transcription of typically silenced repetitive elements. Once transcribed they can mimic pathogen-associated molecular patterns and bind pattern recognition receptors, thereby engaging the innate immune system and triggering inflammation in a process known as “viral mimicry”. Yet how to quantify pathogen mimicry, and the degree to which it is shaped by natural selection, remains a gap in our understanding of both genome evolution and the immunological basis of disease. Here we propose a theoretical framework that combines recent biological observations with statistical physics and population genetics to quantify the selective forces on virus-like features generated by repeats and integrate these forces into predictive evolutionary models. We establish that many repeat families have evolutionarily maintained specific classes of viral mimicry. We show that for HSATII and intact LINE-1 selective forces maintain CpG motifs, while for a set of SINE and LINE elements the formation of long double-stranded RNA is more prevalent than expected from a neutral evolutionary model. We validate our models by showing predicted immunostimulatory inverted SINE elements bind the MDA5 receptor under conditions of epigenetic dysregulation and that they are disproportionately present during intron retention when RNA splicing is pharmacologically inhibited. We conclude viral mimicry is a general evolutionary mechanism whereby genomes co-opt features generated by repetitive sequences to trigger the immune system, acting as a quality control system to flag genome dysregulation. We demonstrate these evolutionary principles can be learned and applied to predictive models. Our work therefore serves as a resource to identify repeats with candidate immunostimulatory features and leverage them therapeutically.