I'm Not a Murderer! Issues with AI for OCR

Liam Downs-Tepper, University of Vienna

On June 21, 1901, a man by the name Albert Packard Curtis murdered Mamie Babcock. This was a busy year for Mr. Curtis, having just gotten married in April and having stolen $17,500 earlier in June. The above is transcribed from an early issue of The Detective, a journal for law enforcement which ran from the late 19th century into the early 20th. It consists primarily of wanted ads, notices of capture, and job postings for various public and private policing forces. Unfortunately, there are very few digitized issues of The Detective, and those which have been digitized have extremely poor OCR. To that end, this project aims to improve the digitisation of these issues, allowing for content-based analysis of the text as well. This project tests different pipelines for digitization and OCR, but in the context of this conference, narrows in more on the potential issues of AI transcription. Though the terminology used tends towards the academic – algorithmic bias, skewed training sets, and so on – the outcomes of this are more straightforward. This conference is intended to discuss complexity; AI is about the lowest common denominator. In this case, that means Mr. Curtis, who, on November 1, 1900, abandoned his wife for his mistress, Mamie Babock. He was not a murderer, did not steal thousands, and was married long enough before to have two children. But all of those details make it spicier and more exciting, and thus, according to the black box behind the AI transcription, are part of the original wanted advertisement. This is not an isolated instance; these issues are consistent across different providers while looking at crime material. This paper examines AI transcription issues for historical crime content, as well as offering other technical solutions to handle this material.

See extended abstract

 Presented in Session 164. Surveillance, Classification, and Bureaucratic Control