Text produced by large artificial intelligence language models is now ubiquitous, and people are eager to learn how to recognize them. A study by marketing firm Graphite found that in the past, artificial intelligence often used dashes or made suggestions to “explore deeper” – these markups are now a thing of the past, but chatbots retain some speech bias.
Image credit: Igor Omilaev / unsplash.com
Today, each artificial intelligence model has its own unique characteristics. Graphite experts discovered 13,000 phrases that appear in AI-generated content at least twice as often as in human-written text — phrases the researchers define as tokens. It is worth noting that the Anthropic Claude model increasingly produces human-like text, while the OpenAI GPT system, in contrast, is moving away from the natural distribution of human words.
As a control group, the researchers collected 10,000 articles that had been published before ChatGPT was released—and were apparently written by people. The authors of this project instructed various artificial intelligence models to rewrite these articles based on brief annotations to eliminate the influence of the original text. After receiving comparable text samples from people and each model, Graphite compared the frequency of use of certain words and phrases and began to identify general patterns in sentence structure.
A characteristic mark of the Anthropic Claude Opus 5.5 model is the word “reliable” – it appears 23 times more frequently in text written by humans. She has begun to avoid the “This is not X, this is Y” structure but still uses variations of “It’s not just X, this is Y”. The phrase “this is important” occurs 116 times more frequently in humans, while the phrase “why X is important” occurs 92 times more frequently.
Photo credit: Milad Fakurian / unsplash.com
OpenAI GPT-6 Astra has its own set of unique features. This model likes to describe “another dimension” to the topic being discussed, noting that an action “might provide” or “could provide” some kind of benefit. The researchers believe that the most striking sign of this model is “corrective representation”: when an object is defined as “not just X” or presented as an alternative – “rather than relying on X”. Similar expressions appear more than a hundred times more frequently in Astra texts than in humans.
Previously, a characteristic of text written by artificial intelligence was a large number of dashes. The developers took this into account, and in the text samples studied by Graphite, the Claude Opus 5.5 model used this type of punctuation 99% less often than Opus 5. OpenAI GPT-6 Astra uses it 88% less often than humans, while Google Gemini 3.1 Pro abandons long dashes almost entirely. Overall, the number of speech tokens is not decreasing — some tokens are being replaced by others, “each version of the model has its own,” the researchers said.
Developers are really trying to achieve as close to a human writing style in AI as possible – Anthropic made this statement with the release of Opus 5.5, and OpenAI made this statement with the launch of GPT-6 Sol and Luna. But according to Graphite, AI labs are unlikely to completely eliminate this voice preference. It’s worth noting that if AI models strive to communicate more humanly, human language will become increasingly similar to chatbot-generated text — something representatives of multiple industries have complained about recently.
If you find an error, select it with your mouse and press CTRL+ENTER.
