You think you are controlling AI, but AI is actually domesticating you! The latest research warns: ChatGPT is changing the way it expresses itself in English, quietly implanting its own preferences. Is it time to re-examine whether we are expressing ourselves or the “repeater” of AI?
Recently, more and more media have discovered that the way they speak is becoming more and more like ChatGPT!
This is not an imitation, but an “invasion”.
Subtly, humans are “assimilated” by AI:
AI not only invades your pen,
It also quietly occupies your brain.
Eventually, pour out of your mouth!
Be careful if you speak English! ChatGPT is brainwashing you
The latest arxiv preprint gives a surprising conclusion: ChatGPT may be quietly changing our word habits.
Whether you find it convenient or uneasy, ChatGPT has quietly become a regular guest in writing, searching, and even everyday conversations, reshaping our language and way of thinking little by little.
How can product managers do a good job in B-end digitalization?
All walks of life have taken advantage of the ride-hail of digital transformation and achieved the rapid development of the industry. Since B-end products are products that provide services for enterprises, how should enterprises ride the digital ride?
View details >
At the same time, there was a heated discussion on the Internet:
Can you see through the content written by AI at a glance through details such as “dashes”? The answer may be changing.
However, the new study gives an unexpected answer: “AI traces” are becoming more and more difficult to recognize – because human language styles are becoming more and more like ChatGPT.
Researchers found that in just 18 months since the launch of ChatGPT, the so-called “GPT words” used by people in daily communication have skyrocketed.
This is the latest research from the Max Planck Institute for Human Development (MPIB) in Germany.
Paper link: https://arxiv.org/abs/2409.01754
Studies have pointed out that ChatGPT will affect the written expression of humans, but this time, researchers want to know if the rapid popularity of AI is also quietly changing our “way of speaking”.
For this study, millions of pages of emails, articles, academic papers, and news reports were uploaded to ChatGPT for polishing.
The researchers then analyzed the AI’s preferred words, such as “delve,” “realm,” and “meticulous.” They call these GPT vocabulary.
Subsequently, in more than 360,000 YouTube videos and 770,000 podcast clips, they tracked the changes in the frequency of use of these words and found an amazing upward curve!
Figure 2: Quantifying ChatGPT’s vocabulary preferences and its impact on human communication
(A) Time trend graph of the use of the word “delve” in academic YouTube reports, showing a significant increase in the use of the word after the release of ChatGPT (p=0.010 compared to the synthetic control group)
(B) ChatGPT vocabulary preference measurement method: By comparing the word frequency distribution between humanly written text and the modified version of ChatGPT, the calculated logarithmic ratio (i.e., GPT score) can quantify the degree to which specific words are preferred by ChatGPT
(C) ChatGPT’s characteristic vocabulary (GPT vocabulary) measured by GPT scores, and the strong preference for “delve” is consistent between different versions of ChatGPT
(D) Synthetic control groups with similar pre-release patterns were constructed from convex assemblages of donor vocabulary for inferring causal effects
This latest study tells us an interesting fact: these “AI accents” may be becoming more and more difficult to spot – because the way humans speak is gradually becoming more like ChatGPT.
Even excluding the influence of synonyms and scripted content, the study found that the usage of GPT vocabulary in spoken English did increase significantly.
This study is thought-provoking, but it also has some notable limitations.
First, the researchers analyzed data from specific GPT models: GPT-4, GPT-3.5-turbo, GPT-4-turbo, and GPT-4o. This limits the research to these specific versions of ChatGPT.
In the coming months and years, OpenAI will undoubtedly introduce new models that may reveal new forms of language use and vocabulary preferences.
Therefore, this study may become obsolete quickly.
Additionally, it is unclear whether ChatGPT does have an impact on more casual oral expressions, especially given the extensive data researchers have obtained from academic sources.
Moreover, the use of language and vocabulary will evolve over time due to many factors; While ChatGPT may have contributed to the change in the vocabulary we use to some extent, it is important to point out that many other sources in society and culture have also contributed to language change.
The rise of “pseudo-people”
Co-author Levin Brinkmann explained to the media: “The language patterns stored by AI technology are feeding back human thinking. Humans are born to imitate each other, but mainly for those who are knowledgeable or highly respected. We are more inclined to replicate their verbal behavior.”
Levin Brinkmann: Research Fellow at the Max Planck Institute for Human Development, focusing on the deepening integration of machines and human society. He uses complex systems theory, machine learning and experimental social science methods to explore the evolution of human-machine hybrid systems. He received a PhD in Computational Social Sciences and a Master’s degree in Physics from the University of Göttingen
In other words, a cultural feedback loop is forming between humans and AI: we train AI using written text, AI sends us statistically reconstructed text, and when we capture these patterns, we begin to unconsciously imitate them.
The study pointed out that more and more people regard AI as a cultural authority, which allows “machines originally trained on human data to quantitatively reshape human culture after forming their own cultural characteristics.”
Brinkmann emphasized:
“delve” is just the tip of the iceberg, and other high-frequency GPT words include:
“underscore”, “comprehend”, “bolster”, “boast”, “swift”, “inquiry”, “groundbreak”……
Figure 3(a): Linear regression analysis of the frequency of high-frequency GPT vocabulary occurrence and synthetic control. Dots indicate the monthly aggregation frequency, and trend lines indicate turning points marked with dotted lines. Words such as “comprehend,” “boast,” and “swift,” all of which are similar to “delve,” have shown significant growth
Figure 3(b): Correlation between the GPT score of a vocabulary (x-axis) and the magnitude of its usage trend change (y-axis). The histogram represents the 95% high-density interval of the posterior distribution, and the shaded region in the inset is the 95% confidence interval of Gaussian process regression. Focusing on the top 20 GPT words (far right of the illustration), a group of words with an annual growth rate of about 25%-50% can be observed
Although the study mainly focuses on the vocabulary level, many researchers have noticed that the influence of AI is gradually penetrating into the tone –
For example, speech becomes longer and more organized, and emotional expression is weakened.
Cornell University research found that using smart replies in chat can lead to a sense of cooperation and intimacy, as users tend to choose more positive and emotional language.
However, once people notice that the other party may be using AI, they will feel that the other party is not sincere enough, and their attitude becomes tougher. The key is not that the other party actually used AI, but that feeling of “doubt”.
Paper link: https://www.nature.com/articles/s41598-023-30938-9
As study co-author Malte Jung, associate professor of information science at Cornell University, said, we judge others based on their linguistic characteristics, and what really influences our impressions is the nature of those languages themselves.
Malte Jung’s research aims to explore the dynamics of human interaction in groups and teams, and how these dynamics are influenced by machines
Mor Naaman, a professor of information science at Cornell Institute of Technology, pointed out that this paradoxical phenomenon of “AI both improves communication and triggers suspicion” reflects a deeper crisis of trust.
Mor Naaman is the Associate Dean for Provost and the Don and Mibs Follett Professor of Information Science at Cornell University’s Science and Technology Campus, the Jacobs Institute for Technology and the Cornell Ann Powers School of Computing and Information Science. He focuses on interdisciplinary topics such as technology and media. He received his PhD in computer science from Stanford University’s InfoLab. In addition, he played for the basketball team, competing as a professional basketball player
He proposed the triple signal of loss of humanity in AI communication:
- Layer 1 signal: human nature. Those clumsy, stuttering tones are “I’m really talking”.
- Layer 2 signal: Attentive. Instead of copying and pasting, take the time to type every word.
- Layer 3 signals: capabilities. True personality is to show humor, emotion and depth in real language.
In what you say, who you are; Once AI endorses us, we may no longer be “ourselves”.
Naaman believes that the future of AI-mediated communication lies in reconstructing and reinforcing these signals, as AI not only changes language but also reshapes thinking.
On dating apps, what is the point of “humor” in your profile when AI can show you humor?
His biggest concern is the loss of subjectivity from language to thinking: “We no longer express our true thoughts, but express what AI helps us organize… We will be more and more convinced by it.”
He warned that without these signals, humans would only trust face-to-face communication — or even video calls.
AI prefers English “London accent”?
When AI quietly defines what constitutes “appropriate” language, the crisis of trust becomes more complex.
University of California, Berkeley research found that AI prefers “standard English”:
When users try to express themselves in Singaporean English, Black English, etc., the AI’s feedback is often distorted, exaggerated, and even uncomfortable.
It’s not just about language, it’s about respect for culture.
Paper link: https://arxiv.org/abs/2406.08726
A Singaporean English-speaking netizen commented: “The exaggerated Singaporean English in the reply is embarrassing.”
Research shows that AI not only favors standard American English, but also devalues users by suppressing other dialects.
This mechanism not only solidifies the wrong perception of ethnic groups, but also distorts the standard of “correct” English.
So it’s not just about preserving linguistic diversity, it’s about protecting the imperfect qualities that truly build trust.
When everyone around us starts speaking “correct” English, we lose those stuttering expressions, local slang, and unconventional phrases – the very beacons of vulnerability, authenticity, and character.
We are at a fork in the road, full of contradictions:
Academic resistance to words such as “delve” and people deliberately avoiding AI accent and other backlash phenomena show that humans may self-regulate against homogenization;
AI systems themselves may be more expressive and personalized, thereby alleviating the current “AI cavity” problem;
The deepest risk is not language convergence, but the loss of human autonomy in thinking and expression.
AI is invading our subconscious. It is secretly rewriting our language genes and manipulating the most instinctive communication code of human beings! Where will this silent spiritual colonization take humanity?
The horror is that we don’t even realize we’re being transformed. Linguists issue an urgent warning: AI’s all-round cultural penetration of human society must be monitored immediately, otherwise the consequences will be unimaginable!
Resources:
https://gizmodo.com/humans-are-starting-to-talk-more-like-chatgpt-study-claims-2000628916
https://www.theverge.com/openai/686748/chatgpt-linguistic-impact-common-word-usage
https://www.scientificamerican.com/article/chatgpt-is-changing-the-words-we-use-in-conversation/
