Education Research Current About VU Amsterdam NL
Login as
Prospective student Student Employee
Bachelor Master VU for Professionals
Exchange programme VU Amsterdam Summer School Honours programme VU-NT2 Semester in Amsterdam
PhD at VU Amsterdam Research highlights Prizes and distinctions
Research institutes Our scientists Research Impact Support Portal Creating impact
News Events calendar Healthy living at VU Amsterdam
Israël and Palestinian regions Culture on campus
Practical matters Mission and core values Entrepreneurship on VU Campus
Governance Partnerships Alumni University Library Working at VU Amsterdam
Sorry! De informatie die je zoekt, is enkel beschikbaar in het Engels.
This programme is saved in My Study Choice.
Something went wrong with processing the request.
Something went wrong with processing the request.

ALP Guide: Spotting AI writing

With the rise of AI tools like ChatGPT, Claude and Gemini, we are all reading more texts with doubt about whether a human or an AI writing tool created the text. Interestingly, like human writers, these tools have common tendencies in their writing. And so they have their own recognisable voices. Here, we look at some of the tendencies of AI writing to help you recognise if a text has been fully generated, or very heavily edited, an AI tool.

These tips are not a policy on detecting AI writing, and it is important to follow university and programme policy on the use and detection of AI. These tips are less about pattern recognition or catching students, and more about considering what you value in good writing and what kind of writers you want your students to become.

We are also working on a Canvas course with activities to practise spotting AI writing. Send us an email at alp.sgw@vu.nl if you would like to be added to that course when it is ready.

We will explore what current research says about common tendencies of AI-generated writing. But first, some disclaimers:

  1. Many of these tendencies aren’t inherently bad and are often employed by skilful writers. AI tools have been trained on a large amount of data, and much of it represents good writing. Some of the tendencies are, however, more problematic.
  2. These tendencies aren’t a means to detect inappropriate AI use and fraud. Human writing can also include these tendencies, and many courses allow AI tools for editing, which, just like text generation, can produce some of these tendencies. Ultimately, it is important to know and follow the university and programme policies on inappropriate AI usage. But recognising AI text tendencies can give you confidence in your judgements about writing and in your ability to comment on the quality of the writing.
  3. The technology is constantly evolving. Any description of the tendencies of AI writing tools is just a snapshot of the current situation. These tendencies will continue to change. Furthermore, as AI-generated texts become more common, these trends may influence human writing. For example, writers across social media, news, and blogs have noted that AI tools often use the em dash—a long dash that marks a break or adds emphasis. As a result, some writers now use it deliberately, while others avoid it because they worry that their writing will look AI-generated.
  4. These are simply tendencies. Not all texts produced by AI writing tools have these features, and AI humaniser tools and prompting can remove some of them.

With those disclaimers out of the way, let’s examine what the some of the current research says about the tendencies of AI generated text. We’ll outline the linguistic, structural and content clues that have been identified through statistical linguistic analysis. We’ll also look at clues identified through more qualitative research relying on expert and instructor opinion. Click on the headings below to read about the clues in each area. 

Back to homepage

Academic Language Programme

About the ALP

The ALP is part of the VU Language Policy

Spotting AI writing

  • Linguistic clues

    • Repetitive sentence structures: AI-generated writing tends to use less varied sentence structures compared to human writing. Variation in sentence length is popularly known as burstiness, meaning that the writing contains a mix of short, medium-length, and longer sentences. AI writing has been shown to use more simplistic sentence structures compared to human writing (Bao et al., 2025) and to use shorter sentences with a lower number of characters per word (Durak et al., 2024).
    • Repetitive vocabulary: AI-generated texts tend to repeat the same words more often. ​Durak et al. (2025)​ found that human-written texts had a substantially higher number of unique words, about four times more than the AI-generated texts.
    • AI-favoured vocabulary: AI-generated content often includes certain specific AI-favoured word choices. For example, Bao et al. (2025) identify language more frequently used by ChatGPT, listing 100 adjectives (e.g. intricate, valuable, exceptional, and pivotal) and 100 adverbs (e.g. primarily, thoroughly, subsequently, and particularly).
    • A more confident and less cautious tone: AI-generated texts tend to sound more confident, and perhaps overconfident (Johansson, 2023). This is because they tend to use more words and phrases that make statements and claims sound more certain. These are called boosters (e.g. certainly, undoubtedly, essential and by far). At the same time, they use fewer words that makes writing more cautious or careful, which are called hedges (e.g. almost, in general, probable and appear to be).
    • No, or very few, errors: AI writing tools are capable of generating large amounts of error-free text very quickly, much more so than a human ever would be even with the aid of spelling and grammar checkers. So, if a text has no visible errors, that could be a sign that it might have been generated by an AI tool.  
  • Structural clues 

    • Headers: AI-generated essays sometimes use headers announcing the topic of the following paragraph(s) in order to organise the content of the paper ​(Goulart et al., 2024)​. AI texts also often use bolding or bullet points to emphasize certain words in the running text.
    • Uniform text patterns: There is some evidence that AI-generated essays follow uniform structures, with similar introductions and conclusions. Early sentences often begin by using key concepts to make general statements about the topic. Although this is a valid opening strategy, human writing often shows more variation ​(Herbold et al., 2023)​. 
    • Fewer different communicative purposes per text: In writing, different sentences serve different functions, or communicative purposes. For example, argumentative sections take a position and support it; explanatory sections provide clarification; narrative sections tell a story or describe events; and comparative sections highlight similarities and differences. Goulart et al. (2024) found that ChatGPT essays were primarily argumentative or explanatory and that students used narrative and comparative purposes in their writing, whereas the ChatGPT-generated texts did not. They also noted that human writers often combine different purposes within a text, but that ChatGPT employed a smaller range of different communicative purposes. The following examples from Goulart et al. (2024, p. 21) illustrate this difference in two short passages, with the functions labelled after each sentence.

      Excerpt 3 - Essay (Student)
      Accents are a distinctive way of pronouncing words in a specific language, and they are associated with specific nations, localities, or social classes (explanation). When you're in a classroom, applying your knowledge about linguistics, variation is very important because it is situated in the field of sociolinguistics (argumentation). (...) As a professional translator, I was familiar with the saying "language is constantly changing", so I had to put this in context (narrative).

      Excerpt 4 - Essay (GenAI)
      Language is a dynamic and multifaceted tool, constantly evolving and adapting to the diverse needs and contexts of its users. Language variation, the phenomenon encompassing differences in pronunciation, vocabulary, and grammar, among other linguistic elements, is a ubiquitous aspect of human communication (explanation).

    The student example shows a mix of functions, while the GenAI version only uses explanation.

  • Content Clues 

    • Hallucinated sources: Watson (2024) investigated suspected hallucinated sources in student essays and found that the majority were not real sources. But the citations often contained some genuine information such as authors and titles.
    • Unexpected sources: Texts created by AI writing tools tend to refer to unexpected sources more frequently. These can include obscure sources, or older sources. For example, ChatGPT was shown to cite older sources more frequently than students and used many sources published in the 90s (Goulart et al., 2024). In contrast, students used more recent information.
    • Lower citation frequency: Goulart et al. (2024) also noted that ChatGPT used fewer in-text citations than student writing, especially in essays.
    • Emphasis on information instead of argumentation: AI-generated texts were found to be more informational, less narrative and be less abstract (Berber Sardinha, 2024). In the case of ChatGPT, it shows a tendency to lack a clear main argument, instead presenting “a list of facts and information” (Goulart et al., 2024, p. 19).
  • Qualitative evidence 

    • Claims beyond likely student knowledge: ​Fisk (2025)​ notes a preference for information that the average undergraduate would be unlikely to know. For instance, one paper made a claim about how Psychology textbooks discuss placebos, but it is not likely that an undergraduate has read enough Psychology textbooks to make such a claim confidently.
    • Broad statements: AI text sometimes makes broad statements when specific information would be more suitable; for instance, it may refer to “researchers” or “authors” instead of giving names ​(Fisk, 2025)​. Furthermore, it seems that AI writing refers to more general topics and does not assume that the reader has existing specialist knowledge in a particular field, so it rarely refers to “niche information that its audience would know” ​(Garib & Coffelt, 2024, p. 7)​. 
    • Inaccurate or fake information: Inaccurate or false data may signal that the content was produced by an AI writing tool ​(Garib & Coffelt, 2024)​.
    • Not meeting assignment requirements: Some AI-generated papers ignore specific instructions in the assessment prompts ​(Fisk, 2025)​, and others do not meet the required length (Perkins et al., 2023).  
    • Omitting class content: Key information, content and texts discussed in the course were sometimes absent from AI-generated work (Perkins et al., 2023). Instead, AI texts may include information from outside of assigned texts ​(Fisk, 2025)​. 
    • Bland writing style: The styles of texts produced AI writing tools have been described as “bland, lifeless or robotic” ​(Fisk, 2025, p. 315)​ and “more formal, concise, and repetitive” ​(Garib & Coffelt, 2024, p. 5)​. Because AI tools are based upon probabilities and often produce the most likely option, they may produce writing that is unsurprising compared to human writing. The extent to which text is surprising is called perplexity. Garib and Coffelt (2024) also argue that human writing might be more nuanced and varied in emotional tone compared with AI-generated text, which may present inconsistent or extreme emotions or opinions. The lack of errors, or clear individual writing voice have been noted as contributing to this blandness, although careful prompting can include errors or mimic a style (​Perkins et al., 2024)​. 
    • Overly poetic language: In contrast, overly poetic language (“rich tapestry” and “symphony”) has also been noted as a possible indicator that a text was generated by an AI writing tool ​(Fisk, 2025, p. 315).​
  • References

    Alexander, K., Savvidou, C., & Alexander, C. (2023). WHO WROTE THIS ESSAY? DETECTING AI-GENERATED WRITING in SECOND LANGUAGE EDUCATION in HIGHER EDUCATION. Teaching English with Technology, 23(2), 25–43. https://doi.org/10.56297/BUKA4060/XHLD5365

    Bao, T., Zhao, Y., Mao, J., & Zhang, C. (2025). Examining linguistic shifts in academic writing before and after the launch of ChatGPT: a study on preprint papers. Scientometrics. https://doi.org/10.1007/s11192-025-05341-y

    Berber Sardinha, T. (2024). AI-generated vs human-authored texts: A multidimensional comparison. Applied Corpus Linguistics, 4(1). https://doi.org/10.1016/j.acorp.2023.100083

    Doru, B., Maier, C., Busse, J. S., Lücke, T., Schönhoff, J., Enax-Krumova, E., Hessler, S., Berger, M., & Tokic, M. (2025). Detecting Artificial Intelligence–Generated Versus Human-Written Medical Student Essays: Semirandomized Controlled Study. JMIR Medical Education, 11. https://doi.org/10.2196/62779

    Durak, H. Y., Eğin, F., & Onan, A. (2025). A Comparison of Human-Written Versus AI-Generated Text in Discussions at Educational Settings: Investigating Features for ChatGPT, Gemini and BingAI. European Journal of Education, 60(1). https://doi.org/10.1111/ejed.70014

    Fisk, G. D. (2025). AI or Human? Finding and Responding to Artificial Intelligence in Student Work. Teaching of Psychology, 52(3), 314–318. https://doi.org/10.1177/00986283241251855

    Garib, A., & Coffelt, T. A. (2024). DETECTing the anomalies: Exploring implications of qualitative research in identifying AI-generated text for AI-assisted composition instruction. Computers and Composition, 73. https://doi.org/10.1016/j.compcom.2024.102869

    Georgiou, G. P. (2024). Differentiating between human-written and AI-generated texts using linguistic features automatically extracted from an online computational tool. https://doi.org/https://doi.org/10.48550/arXiv.2407.03646

    Goulart, L., Matte, M. L., Mendoza, A., Alvarado, L., & Veloso, I. (2024). AI or student writing? Analyzing the situational and linguistic characteristics of undergraduate student writing and AI-generated assignments. Journal of Second Language Writing, 66. https://doi.org/10.1016/j.jslw.2024.101160

    Herbold, S., Hautli-Janisz, A., Heuer, U., Kikteva, Z., & Trautsch, A. (2023). A large-scale comparison of human-written versus ChatGPT-generated essays. Scientific Reports, 13(1). https://doi.org/10.1038/s41598-023-45644-9

    Johansson, I.-R. (2023). A Tale of Two Texts, a Robot, and Authorship: a comparison between a human-written and a ChatGPT-generated text. Malmo University.

    Kreuz, R. (2025). Too many em dashes? Weird words like ‘delves’? Spotting text written by ChatGPT is still more art than science. The Conversation. https://doi.org/10.64628/AAI.PXEVMF9R5

    Liao, W., Liu, Z., Dai, H., Xu, S., Wu, Z., Zhang, Y., Huang, X., Zhu, D., Cai, H., Li, Q., Liu, T., & Li, X. (2023). Differentiating ChatGPT-Generated and Human-Written Medical Texts: Quantitative Study. JMIR Medical Education, 9(1). https://doi.org/10.2196/48904

    Perkins, M., Roe, J., Postma, D., McGaughran, J., & Hickerson, D. (2024). Detection of GPT-4 Generated Text in Higher Education: Combining Academic Judgement and Software to Identify Generative AI Tool Misuse. Journal of Academic Ethics, 22(1), 89–113. https://doi.org/10.1007/s10805-023-09492-6

    Watson, A. P. (2024). Hallucinated Citation Analysis: Delving into Student-Submitted AI-Generated Sources at the University of Mississippi. The Serials Librarian, 85(5–6), 172–180. https://doi.org/10.1080/0361526x.2024.2433640

Related content

Contact the Academic Language Programme

For VU staff: contact us about courses on communicative skills or questions related to language policy.

Dr. Gea Dreschler, ALP academic director

Dr. Nel de Jong, coordinator ALP for Dutch (NT1, NT2)

For students: for questions or information about a course, workshop or coaching, contact us at

Quick links

Homepage Culture on campus VU Sports Centre Dashboard

Study

Academic calendar Study guide Timetable Canvas

Featured

VUfonds VU Magazine Ad Valvas Digital accessibility

About VU Amsterdam

Contact us Working at VU Amsterdam Faculties Divisions
Privacy Disclaimer Safety Web Colophon Cookie Settings Web Archive

Copyright © 2026 - Vrije Universiteit Amsterdam