ETS e-rater Automated Essay Scoring System Exhibited Gaming Vulnerabilities and Dialect Bias

In 2012, MIT researcher Les Perelman published groundbreaking research exposing critical flaws in Educational Testing Service's (ETS) e-rater automated essay scoring system, which was used to evaluate essays on the Graduate Record Examination (GRE) and Graduate Management Admission Test (GMAT). The e-rater system, developed by ETS and used alongside human graders, was designed to provide consistent and efficient scoring of written responses from hundreds of thousands of test takers annually. Perelman's research revealed that the e-rater system could be systematically gamed through verbose but semantically meaningless writing. He demonstrated that essays filled with complex vocabulary, long sentences, and sophisticated-sounding but nonsensical content consistently received higher scores than coherent, well-reasoned shorter essays. In one notable example, an essay that began with factually incorrect statements and contained logical contradictions throughout received a score of 5 out of 6 simply because it was long and used advanced vocabulary. The system appeared to prioritize surface-level linguistic features such as word count, sentence complexity, and vocabulary sophistication over actual content quality, logical reasoning, or factual accuracy. The research also uncovered systematic bias against writers using non-standard English dialects, particularly African American Vernacular English (AAVE) and other linguistic variations common among minority populations. Essays written in these dialects consistently received lower scores even when the content was substantively equivalent to essays written in Standard Academic English. This bias had profound implications for educational equity, as it potentially disadvantaged qualified candidates from diverse linguistic backgrounds in graduate school admissions processes. The technical root cause lay in e-rater's reliance on natural language processing algorithms that weighted heavily toward superficial textual features rather than semantic understanding or logical coherence. The system used statistical models trained primarily on essays written in Standard Academic English, creating an inherent bias toward that linguistic variety. ETS initially defended the system by noting that it was used in conjunction with human graders, but Perelman's work demonstrated that the algorithmic bias could influence overall scoring outcomes and perpetuate educational inequities. Following this research and subsequent criticism, ETS made modifications to the e-rater system and increased transparency about its limitations, though automated essay scoring remains a subject of ongoing debate in educational assessment.