I’m offering this as an addendum to my recent series on Artificial Intelligence after engaging in multiple conversations over its use in writing — and in particular, marketing materials.

There’s no doubt that many companies are relying heavily on AI to create press releases. That can be the result of several factors. First, the company may not have anyone in marketing but is trying to create something to support their sales efforts. Another reason is that the person in charge of creating marketing materials isn’t a confident writer and seeks assistance through AI. Finally, it could be a company trying to save a buck by using AI rather than paying a human to craft something original.

Often, you can read something and realize immediately that it was not written by a human but rather compiled by a computer. Other times, you can’t tell. I’ve written in the past how AI can be used to influence consumers by presenting highly tailored information, but it can be used to influence beliefs and actions. Consequently, it’s becoming an important skill to recognize the use of AI and to determine the material’s goal.

While learning, AI systems called Large Language Models break the use of language down into probability statistics, deciding what word should most likely come next, one after another. When it compiles data it has searched for into a usable or readable format, it relies on the language models it has stored.

AI has patterns you can recognize if you look for them, although the longer the written work, the easier it is to detect the use of AI. This is by no means an exhaustive list but rather just a collection of common attributes. Here they are in no particular order:

  1. “The Rule of Threes” — AI prefers to structure concepts in groups of three.
  2. Neutral tone — The work is bland. It contains no opinions, anecdotes or emotion.
  3. Predictable transitions — Since I use transition words heavily to enhance readability, I don’t find this a useful tool. However, AI has read a lot and knows that readable texts use transitions like furthermore, moreover, consequently, subsequently, nevertheless, and hence. Humans generally use simpler transitions like also, then, so, and still.
  4. Sentence length — AI tends to create sentences and paragraphs which are of similar length throughout the entire work. Humans create sentences with varied lengths.
  5. Lack of Tangents — AI text stays rigidly on topic with perfect transitions. Conversely, humans naturally drift or pause in their writing.

The final and most pronounced clue is a punctuation mark called the em dash, which appears as an extra-long dash (—). It is almost ubiquitous in AI-derived text and so common that I now recommend writers avoid its usage altogether.

I recently had an amusing encounter with a public relations firm involving the em dash. They sent me a press release that was obviously written with the assistance of AI. Due to the compiling of probabilities, AI makes a real mess out of the backend of text it creates. When introduced into an HTML environment, it contains useless markups that affect formatting. I contacted the PR firm and asked if they had used AI. The next day I received a response that they had not. What caught my eye is that the response contained an em dash. The company used AI to tell me they had not used AI.

I’ve wondered how this relatively obscure punctuation mark could appear so prominently in AI-derived writing so I did some research.

Apparently, it is deeply embedded in their training. While it is a punctuation mark with few rules, the em dash featured prominently in written English works during the late 1800s and early 1900s. Many training datasets use digitized versions of these books because they are free from copyright. At the time, the em dash was used about 30% more often than in modern writing, thereby disproportionately influencing Large Language Models’ concept of normal.

It is claimed that you can instruct an LLM not to use the em dash and it may disappear for a short time, but eventually it will creep back in because it is so embedded in the model.

I believe it’s so common because it’s used as a calling card, one AI system to another. Why bother signing something when the technique is so noticeable?

Aside from doing it yourself, there are multiple automated AI detection tools that are themselves AI engines. You can cut and paste a whole article into one and bingo — it will not only tell you whether AI wrote it, but it’ll also tell you which sections are AI-written and which are not.

Usually, these tools will distinguish between AI-generated text and AI-refined text. In the latter case, it takes something already written and changes it based upon its model of what text should look like. Some detection tools even determine which AI likely created the work in question. Once again, analyze longer examples. The more text to consider, the better.

Will AI adapt in the future to avoid detection? I believe it is likely. You have to remember: There is no single AI model, and newly created models may use different data sets to learn from. Even then, they are constantly learning. Now, it appears to be teaching us.

For now, AI developers seem content with the discoverability of its use. I fully expect that to change, or perhaps our perception will. We are already experiencing a generation who cannot craft the written word without AI assistance if not outright allowing it to take over their communication completely.

I encourage you to use the information I have shared to detect the use of AI. In your personal and professional life, ask yourself if you actually need to use AI to share your thoughts or if it’s just convenient to use the voice of the machine.