Dash Kindly

The "em-dash" has become the new "kindly."

Generally speaking, Americans do not use "kindly" in professional writing. It is, however, common in parts of the Commonwealth of Nations, especially those parts that many Americans associate with Internet fraud, such as India, Kenya and Nigeria. Accordingly, it's become the one thing that many unsophisticated people look for when attempting to sniff out a bogus investment scheme or employment offer, because it's allegedly a foolproof telltale that allows a correspondence to be immediately written off. Some people are willing to go so far as to claim that the use of "kindly" is a deliberate marker of fraud, and that people who continue to correspond after having seen it are showing themselves to be rubes, ripe for the plucking. This, in turn, has lead to a certain amount to ridicule being directed at anyone who shows uncertainty about a message with the word "kindly" in it.

The "em-dash" has become this marker for spotting "A.I. slop," as the Internet likes to call it. Because, supposedly, "real people" never use the em-dash, so any text in which is features prominently must have been created by generative automation. The fact that this forgets that large language models basically take their training data, chop it up into pieces, figure out the mathematical relationships between various parts and create their outputs based on that is surely of no real importance to the process.

The common thread here is the idea that there is a shortcut to discernment. Be on the lookout for "kindly" and you'll never be tricked by an Internet fraud. Stay alert to the "em-dash" and you'll always be able to tell when a passage was generated by an LLM. Utterly reliable and remarkably easy; who knew that the signs were so obvious?

That there is widespread anxiety about being deceived is understandable, at least in the case of internet frauds. While many schemes may only cost someone a small amount of money, individual losses can easily run into the thousands, or tens to hundreds of thousands. Avoiding that financial hit by spotting a supposedly ubiquitous telltale is clearly attractive. With text generated by LLMs, the case is a little less clear. Okay, so someone generated text with a Copilot or Perplexity prompt rather than writing it all out themselves. As long as it gets the point across, who cares? The intersection between generative automation and internet fraud is a concern, but the "em-dash" isn't something that is apparent in a faked celebrity pitch video for a fraudulent investment scheme.

It seems that a lot of what drives the hype (hysteria?) concerning both "kindly" and the "em-dash" is the need to appear discerning to other people. Despite the fact that there weren't any consistent glaringly obvious tells in GPT-2's text output, the idea that there's a silver bullet to detect generated text from any current LLM is strong and, if the enduring attachment to "kindly" is any indication, not going anywhere anytime soon. It's common for people to seek ways of ranking themselves against one another; it's been postulated that people find meaning in positive self-worth, and one way of attaining that is to seek out ways of showing that one is better than others. But that also means avoiding indicators that one might be worse than others. If one doesn't jump on the "em-dash" or "kindly" bandwagons and is subsequently tricked, the ego-bruising ridicule is likely to follow.

I spend more time than is likely healthy looking at dodgy material in my Junk Mail folder and, quite often, my inbox. And I've done my share of looking at things online. While there are certainly tells for a good amount of internet fraud, most of them come from the fact that the average would be perpetrator of internet fraud isn't targeting specific people, or is clearly attempting to pass themselves off as someone they aren't. An e-mail that claims to have found someone's résumé online but is addressed to "applicant" or the recipient's e-mail alias, or a LinkedIn profile that claims to be for a recruiter working in a hybrid model for a company in Texas but who supposedly lives in Liechtenstein. But these are really markers of poverty and poor education; the things that drive people to internet fraud for a living in the first place. It's how "kindly" first became a marker of fraud; people attempting to pass themselves off as Americans without having any idea of how Americans actually write. Pretty much no-one in the United States feels the need to remind another American that Atlanta is in Georgia, let alone the United States.

The tells from the uneducated are easy to spot, because they tend to recur over and over again. The graduates of "hustle kingdoms" in West Africa, for instance, go on to teach others the same things that they have learned, and this seemingly rote transfer of knowledge results in tells that are decades old still being common. And these are the sorts of fraudsters that the average person is likely to encounter. The truly sophisticated fraud artists aren't randomly e-mailing people, playing a numbers game.

With the "em-dash" the motivation, avoiding looking like a rube, is the same, but this time, it's the reader who is uneducated. Again, LLMs use the "em-dash" because it's common in the training data. To the degree that the "em-dash" is a marker of LLM-generated text, it's something of a self-fulfilling prophecy; people are starting to avoid the punctuation mark specifically because Internet paranoids have marked it as a surefire sign of inauthenticity. And because most of the people who have the time to be online a lot are young, and young people are reading less, they're not as exposed to how people older than themselves commonly write. But the makers of LLMs do not limit themselves to the young and online; they're vacuuming up all the data they can get. "Doesn't read like a human" is really "doesn't read like a young and affluent online Westerner."

The idea that there is a quick way to be "smart" about things is an attractive one, and difficult to debunk, so it's going to persist. With any luck, it's not too expensive of an idea to hang on to.