Kao used a technique known as natural language processing, or NLP, to scan more than 22 million comments submitted to the FCC’s website. He found that more than 17 million were duplicates or close parallels. But many of those were, he writes, “legitimate public mailing campaigns,” which provide boilerplate text for real people to submit.
Intriguingly, the comments that Kao ultimately concluded were ‘fake’ were actually quite diverse in their specific phrasing – but that variation was only superficial. As an example, Kao highlights the anti-net neutrality phrase “Individual citizens, as opposed to Washington Bureaucrats, should be able to select whichever services they desire.” The system used to generate the fake comments swapped out words in such phrases again and again – for instance, switching “people like me” for “individual citizens” and “products” for “services” – to produce 1.3 million superficially distinct variations on the same basic block of text.