Back in December, ProPublica published a fascinating look into the snake-oil industry around Scientific Content Analysis or SCAN, a so-called "law enforcement tool" that purports to help investigators determine whether their suspects lying. This highly-profitable yet totally-dubious training method works through a rigid grammar analysis that relies entirely on the assumption that human brains only ever work in one completely uniform, logical, rational, conscious, and deliberate manner:
With SCAN, Sapir encourages the asking of a simple, open question: What happened? After the person writes a statement, the SCAN investigator looks for signs of deception, analyzing, among other things, pronouns used, changes in vocabulary, what’s left out and how much of a statement is devoted to what happened before, during and after an event. Indications of truthfulness include use of the past tense, first-person singular (“I went to the store”); pronouns, such as “my,” which signal commitment; and direct denials, the best being: “I did not do it.” Signs of deception include lack of memory, spontaneous corrections and swapping one word in for another — for example, writing “kids” in one place and “children” in another.
Sapir likens SCAN to Sudoku, only with words, not numbers, sentences, not squares: “Everything must fit — left to right, and top to bottom.”
And of course, there's no consideration for the possibility that someone might be, idunno, nervous or anxious or god forbid under-educated and therefore might respond to this "test" in ways that seem arbitrarily "suspicious."
Yet there are still tons of cops who swear by it anyway — even though, as ProPublica reveals through a comprehensive analysis of SCAN test results, the system has about a 50 percent likelihood of accurately predicting whether a suspect is lying, which is … no better than a random guess. Read the rest