— August 21, 2024

Have you ever read an article and thought, “An AI chatbot must have written this”?

If so, you’re likely picking up on subtle linguistic cues that betray the text’s machine-generated origins. Large Language Models (LLMs) still have a long way to go.

One of the most persistent and revealing cues is the overuse of generic participial phrases.

These phrases often begin with a comma and a present participle.

Examples include everything you can attach an “-ing”-ending to, such as:

“, fostering…”

“, revealing…”

“, hindering…”

“, creating…”

… and so on.

This has become the verbal tick of LLM-generated text.

“Verbal ticks”

Verbal ticks are expressions such as “um“, “like“, or “you know. Those are filler words and phrases that we often unconsciously sprinkle into our speech.

They often surface when we’re nervous, searching for the right word, or simply buying time to formulate our thoughts.

While occasional ticks are natural, their overuse can become distracting, undermine our message, and create a perceived lack of confidence.

Most of us have witnessed this type of speech in one setting or another:

— “So, I was at the store, you know, and I saw this amazing dress, you know? It was like, the perfect color, you know, and it fit me so well, you know? But then, you know, I saw the price tag, and it was, like, way too expensive, you know? So, I had to put it back, you know, but I’m still thinking about it, you know? I might go back and get it, you know, if it goes on sale, you know?”

It is often annoying.

In the realm of AI-generated text, we have noted that generic participial phrases are chatbot versions of verbal ticks.

These phrases, frequently appearing as end clauses, signal a lack of genuine human expression or understanding

The Grammar and Linguistics of the Problem

So, what exactly are these generic participial phrases that are so prevalent in LLM-generated text?

In the context of sentence structure, an “end clause” refers to a clause that appears at the end of a sentence. It can be a dependent clause (requiring the main clause for complete meaning) or an independent clause (capable of standing alone as a sentence).

In our discussion, we’re specifically referring to participial phrases that function as dependent clauses and are positioned at the end of sentences. These phrases, often beginning with a comma and a present participle, modify the main clause and provide additional information.

For example, in the sentence “The cat slept peacefully, dreaming of chasing mice,” the end clause is “, dreaming of chasing mice.”

Grammatically, these phrases function as adverbial clauses, modifying the main verb of the sentence. They provide additional information about the action or state being described, often expressing purpose, result, or accompanying circumstances.

Linguistically, their appeal to LLMs likely stems from their versatility and perceived neutrality. They can be easily inserted into a variety of contexts without requiring significant syntactic or semantic adjustments.

Moreover, they often convey a sense of generality and abstraction, which may align with the LLMs’ tendency to avoid specificity and commitment. After all, they’re not really grasping the “conversation”, just guessing with often brilliant results.

The present participle (that “-ing” verb form), is a grammatical multitasker. It acts as an adjective to describe nouns (“the singing bird”), or teams up with helping verbs to create continuous tenses (“I am writing”). Its adaptability makes it a cornerstone of expressive language.

However, this very versatility and neutrality can also be their downfall. Overused, these phrases become predictable and formulaic, stripping the text of its vibrancy and individuality.

The challenge, then, is to strike a balance. Participial phrases, when used judiciously and purposefully, can enrich and enliven our writing. But their overuse are currently (August 21, 2024) impossible for users to remove from LLM-generated text. Open AI, Google and others haven’t been able to fix this setting (or don’t care to).

While the following example of an LLM overusing generic participial phrases is slightly exaggerated, the structure remains the norm:

— “The company implemented new strategies, aiming to increase productivity and boost employee morale. This initiative is expected to yield positive results, focusing on clear communication and collaboration, fostering a more engaged and motivated workforce. Additionally, the company will invest in professional development opportunities, empowering employees to enhance their skills and knowledge, driving innovation and growth.”

Regardless of your programming skills and computer literacy, how many hours you spend “debating” with or coaxing the AI assistant.. or if you swear, try refining prompts and rephrasing requests… and regardless of how many fake apologies and empty promises you get — the chatbot is programmed with this flaw.