AI Studies
Most Commonly Used ChatGPT Words & Phrases
Are humans as good at detecting AI content as they think? We analyzed 10 million ChatGPT terms to find out! Discover ChatGPT's most common words and phrases.
With the release of ChatGPT in November 2022, many content marketers and experts have come up with ways to easily identify AI content. Most recently, numerous online discussions and blogs have cropped up, claiming to have captured its most commonly used words. But can humans accurately identify ChatGPT-generated content? Or do we just think we can?
To see how helpful and accurate ‘ChatGPT’s Most Common Words/Phrases’ roundups are, we collected articles, blogs, and Reddit posts that listed ChatGPT’s most common outputs and compared them to our extensive dataset of AI-generated text.
Multiple studies have shown that humans are not as good at detecting AI content as they think they are.
A number of reviews, blog posts, and social posts have claimed to be able to easily identify ChatGPT-generated content through certain words and phrases — i.e. many suggest that the word “delve” is used in almost every ChatGPT-generated text.
But can we really use these lists to help detect AI-generated text?
While you may be tempted to answer ‘yes,’ there are multiple data points that suggest otherwise. For example, a study by the University of Washington and this similar one, shows us that humans are likely overconfident in their AI-detecting abilities. In the first study, researchers found “that [human] evaluators were unable to distinguish between GPT-3 and human-authored text across… domains”.
Ironically, the human evaluators’ confidence, measured in the percentage of “Definitely” responses, remained high even as their accuracy scores plummeted across various testing conditions. Humans reached opposite conclusions for the same observations — proving that humans detect AI subjectively, with personal bias.
Above is an excerpt of two evaluators’ reasonings for why the text is human-generated (left) or GPT-generated (right). Humans reached opposite conclusions for the same observations — proving that humans aren’t as good as they think they are at identifying AI-written content.
To have a comprehensive dataset of the words and phrases ChatGPT is actually using, we collected and cleaned seven datasets of AI-generated text, then tallied up the most commonly used words and phrases in those outputs. Here is where our compiled dataset was gathered:
Below is a list of all the most commonly used ChatGPT words, from highest (most used) to lowest (least used).
Below is a list of all the most commonly used ChatGPT phrases, from highest (most used) to lowest (least used).
There aren’t many words that stand out or are unique when compared to human-generated content. For example, some of the most common words are:
These aren’t very unique or easily identifiable — most of these words are actually commonly used in human-written content.
When we look at the most common phrases, we see:
Again, very common phrases that are not easily identified as ChatGPT text.
Moving down the list (to the less common words), we can see some more unique words, such as ‘designed,’ ‘reflect,’ ‘confident,’ and ‘healthier.’ We also see more unique phrases such as “perform operations,” “benefits and drawbacks,” and “nature of reality.” However, these were not found in our datasets very often and therefore aren’t nearly as common in ChatGPT outputs.
At first glance, this dataset doesn’t show us any major differences between the words ChatGPT uses and the words used in human writing. In other words, ChatGPT does not use many words that can easily be spotted by humans.
This theory is backed up by both our dataset and other studies such as Elizabeth Clark’s study which looks at how likely it is that humans can detect AI-generated content. This study determined that even after training, humans were about 55% accurate in detecting AI content — basically a coin flip. When looking at the common words used by ChatGPT in our dataset, we can see that this theory stands up. ChatGPT doesn’t tend to use unique or easily identifiable words in its outputs, which makes it difficult to identify.
Are there any words or phrases that are unique to ChatGPT like so many blogs and experts claim?
According to several blogs and social posts, ChatGPT-produced text is easy to identify because it commonly uses unusual words. To see if this is indeed true, we gathered all these words and phrases from publicly available articles and posts online. Here is a list of the so-called commonly used words and phrases that experts claim to be easily identifiable:
To verify whether or not they were actually being used in ChatGPT outputs, we compared them to our extensive dataset.
Here are the results:
The findings from this table are interesting — there was a lot of overlap between the online lists we found and our dataset. In fact, almost every word was used in our dataset. Here’s the caveat — most words were NOT frequently used by ChatGPT.
The most popular word — “delve” — which according to multiple sources is the most easily identifiable culprit behind ChatGPT-generated text, was only used 146 times. This is minuscule compared to the tens of millions of words and phrases we analyzed in our dataset.
In other words, although there are some unique identifying words that might be used commonly by ChatGPT, from this list we can see that many aren’t — and even those that are can be difficult to identify and distinguish from human-written text.
Here are the results:
Similar to the words dataset, our research on phrases highlights that what experts claim are common ChatGPT-generated phrases — aren’t actually that common at all!
The datasets demonstrate that there aren’t any words or phrases that help humans easily identify ChatGPT-generated content.
Our initial findings show that in contrast to their confidence levels, humans are actually NOT good at identifying AI-generated content.
Humans tend to be overconfident regarding their ability to correctly spot AI-generated content and lists containing ‘ChatGPT’s Top Words and Phrases’ are more or less useless.
AI detectors like Originality.ai — which are significantly more accurate than humans at identifying AI content — should be used to help verify when and if AI is being used.
In short, no. Surprisingly, ChatGPT rarely uses words from the ‘most-common’ roundups. To the contrary, its most frequently used words and phrases — according to our extensive dataset — are generic and typical of human writing.
While it is true that many of the words from the blog roundups showed up in our dataset, they did not appear nearly as many times as you would expect, given how often they were mentioned across multiple articles and blogs.
No. According to the University of Washington study — a study of how likely humans are to detect AI-generated content — even after training, people were only able to guess the origins of a text 55% of the time — basically a coin flip. Without training, human accuracy scores were even lower — and this was for ChatGPT3, the less-sophisticated counterpart to ChatGPT4.
Not really. This is because they are not backed by reliable data — they usually represent the author's personal experience using ChatGPT. The ‘most-common’ words depend on:
The short answer is no. Humans use many of these words too! If however, you see an unusual word that humans rarely use, like “tapestry,” it is more likely your text was produced by ChatGPT. Since you can never be sure, we recommend you scan the text with a trusted AI detector, such as Originality.ai.
