Introduction
Artificial intelligence has become a bigger part of our daily conversations as it develops. We use it across many fields, such as language learning, academics, or even for very simple tasks like generating emails. As a result, some people have been noticing that humans have also started to sound like AI. This article examines whether AI is actually leaking into human language, drawing on evidence from five different studies.
AI’s Distinctive Style
To ask whether AI language spreads, we first need to know if it’s different from human language in the first place: Mieczkowski et al. (2021) studied smart replies, the short reply suggestions offered in messaging apps. They found that 93% of smart replies were positive and only 6% were negative. The smart replies also contained more positive emotion words than human messages did (Mieczkowski et al., 2021). Smaglii et al. (2026) compared 20 ChatGPT texts with 20 texts from English coursebooks. The ChatGPT texts had slightly more lexical diversity (variety of unique words), longer sentences, and more linking words such as “moreover” and “however.” They also followed a predictable paragraph pattern with a generic tone. The coursebook texts, on the other hand, were more varied and closer to real communication (Smaglii et al., 2026). These findings reveal that AI has a generic and predictable structure of writing with a systematic skew toward positivity.
Short Conversations: Mixed Signs
Next, researchers tested what happens when people actually use AI suggestions. In Mieczkowski et al. (2021), 34 pairs of university students did a picture-matching task over a chat app. One partner could see smart replies and was told to use them. Messages containing smart replies were more positive than the messages in the control group. However, when the AI-written parts were removed, the users’ own words were no more positive than the control group’s. Their partners did not become more positive either, and only two of 34 matchers said their partners sounded even slightly machine-like (Mieczkowski et al., 2021). Hohenstein et al. (2023) ran two online experiments in which pairs chatted about a policy issue. In the first, when one partner used more smart replies, the other partner’s messages became more positive, even when the smart replies themselves were left out of the analysis. But in the second one, language changes came mainly from people using the suggestions themselves, and with smart replies removed, the authors found minimal differences between conditions. People who were only suspected of using smart replies were rated as less cooperative and less close, yet people who actually used them were rated as more cooperative and closer (Hohenstein et al., 2023).
AI’s Effect on Language Learning and the Risk of Over-reliance
Almutairi (2025) analyzed interaction data and interviews with ESL students and educators, including about 120 learners of English as a second language and 10 instructors. Among the learners, 85% said their confidence in writing increased and 68% reported better vocabulary and grammar, but 40% worried about relying too much on AI. The paper’s background section also notes, citing earlier research, that heavy reliance on AI-generated text can reduce linguistic diversity (Almutairi, 2025). The results are self-reported, which the paper acknowledges may introduce bias, and the reported results do not show whether learners’ actual speech or writing changed (Almutairi, 2025). Smaglii et al. (2026) add a related worry: if learners mainly use AI texts as models, they may get used to language that is correct but overly general.
Evidence from Real Speech
Yakura et al. (2024) offer the most direct evidence. First, they identified words that ChatGPT prefers when editing text, such as “delve,” “boast,” “meticulous,” and “showcase.” Then they analyzed 737,083 hours of unscripted conversation from 824,634 podcast episodes. They chose spontaneous speech because, unlike writing, it cannot involve copying and pasting from a chatbot (Yakura et al., 2024). Using a method that estimates what would have happened without ChatGPT’s release, the authors found that “delve” was used about 44% more than expected in science and technology podcasts 13 to 18 months after release. Education and business podcasts showed increases of about 32% and 31%, while a mixed sample rose only about 9% and sports podcasts dropped slightly. Across 3,535 words, the more strongly ChatGPT favored a word, the more its spoken use tended to rise. Among the top 1% of ChatGPT-favored words (36 words), 28 increased and 8 decreased (Yakura et al., 2024). The authors also ran an experiment with 496 online participants, who played a picture-guessing game with a chatbot that secretly used particular words from synonym pairs, such as “mug” instead of “cup.” During the game, participants used the chatbot’s word 61% of the time, compared with 17% for the alternative. After a three-minute distraction task, when describing new images with no chatbot present, they still used the chatbot’s word 58% of the time, compared with 23%. In a forced choice between two words, they picked the chatbot’s word 63% of the time. Only 15 of 496 participants (3%) noticed the chatbot’s word pattern (Yakura et al., 2024). The effect was not permanent for every word. Use of “delve” peaked around mid-2024 and then fell below its earlier baseline, while “boast,” “meticulous,” and “showcase” stayed elevated. The authors suggest that people may avoid words once they become known as signs of AI (Yakura et al., 2024).
Conclusion
Taken together, the five studies suggest that AI language has a recognizable style and that this style can spread into human language. How strongly it spreads depends on what is measured. The two smart-reply studies found little or mixed evidence. Mieczkowski et al. (2021) found that people did not pick up the AI’s positive tone in their own words, and Hohenstein et al. (2023) found signs of it in one experiment but not clearly in the other. Yakura et al. (2024) found clearer evidence for individual words in both real speech and an experiment. One possible reason is that the smart-reply studies looked at the emotional tone, while Yakura et al.’s (2024) study looked at specific word choices. There are also hints that people may push back. Mieczkowski et al. (2021) suggest that senders may have avoided copying the AI’s positivity so as not to sound like the AI, and Yakura et al. (2024) suggest that people avoid words once they are seen as AI-related. Hohenstein et al. (2023) also found that suspected AI use lowered ratings of cooperation. Together, these hint that social pressure could limit how far AI language spreads, although none of the studies mentioned here tested this directly. In short, AI is seeping into human language in specific, measurable ways, especially word choice, but the effect is uneven and may fade for some words. Longer studies across more languages and everyday settings are needed to show how far it goes, and the ethical implications of the impact of such AI usage should not be disregarded.
Kaynakça
- References
- Hohenstein, J., Kizilcec, R. F., DiFranzo, D., Aghajari, Z., Mieczkowski, H., Levy, K., Naaman, M., Hancock, J., & Jung, M. F. (2023). Artificial intelligence in communication impacts language and social relationships. Scientific Reports, 13(1), 5487. https://doi.org/10.1038/s41598-023-30938-9
- Yakura, H., Lopez-Lopez, E., Brinkmann, L., Kirfel, L., Gupta, P., Soraperra, I., Eisenmann, T. F., Wulff, D. U., & Rahwan, I. (2024). Empirical evidence of Large Language Model's influence on human spoken communication. https://arxiv.org/abs/2409.01754
- Mieczkowski, H., Hancock, J. T., Naaman, M., Jung, M., & Hohenstein, J. (2021). AI-mediated communication: Language use and interpersonal effects in a referential communication task. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), Article 17. https://doi.org/10.1145/3449091
- Smaglii, V., Korolova, T., Yukhymets, S., & Kivenko, I. (2026). Human vs AI-generated texts in language learning: A linguistic comparison. Arab World English Journal (AWEJ) Special Issue on CALL, 17(12), 278-287. https://dx.doi.org/10.24093/awej/call12.16
- Almutairi, F. (2025). Language and AI interaction: Studying how AI language models are affecting human communication and language learning [Conference paper presentation]. World Conference on Emerging Science, Innovation and Policy 2025. Futurity Research Publishing. https://doi.org/10.5281/zenodo.15852254