Rising Concerns Over AI-Generated Texts
In recent weeks, the discussion surrounding the identification of texts authored by artificial intelligence (AI) has gained significant traction across social media platforms. This debate has permeated various sectors, including media outlets, academia, and the literary community, as individuals ponder whether certain stylistic choices they commonly employ could render their work indistinguishable from machine-generated content.
A notable example of this discourse is an Instagram post that highlights several “indicators that an art-related text has been written by AI.” Among these indicators is the verb “to traverse,” which the post claims is a term frequently found in AI-generated narratives. However, the comments section reveals a counterpoint, with some users asserting that they employed such vocabulary long before the advent of tools like ChatGPT, while others point out that many of the identified phrases have been fixtures in academic, journalistic, or literary texts for years.
Pangram Labs, a startup dedicated to detecting AI-generated content and plagiarism, has endeavoured to pinpoint some of these patterns. Founded by Max Spero and Bradley Emi, both alumni of Stanford University with backgrounds in AI, the company has compiled guidelines that encompass indicators related to vocabulary, sentence structure, paragraph length, detail level, and textual organisation.
Common Vocabulary and Phrasing Trends
One of the most straightforward clues to identify AI-generated text is its vocabulary. In a guide released in 2025, Pangram lists numerous words that, based on their analyses, are frequently found in AI-produced content. Terms such as “to delve,” “landscape,” “vibrant,” “crucial,” “intricate,” “environment,” “innovative,” “nuances,” “significant,” “transformative,” “robust,” and “to reveal” are among those highlighted.
Additionally, the company identifies common phrases that recur within this type of writing, including “it is important to note,” “in summary,” “generally speaking,” “in conclusion,” “crucial role,” “significant impact,” “shed light on,” “highlights the importance,” “serves as,” “is a testament to,” “provides valuable insights,” “in a world where,” and constructions like “not only…, but also.” While these expressions serve as a helpful starting point, Pangram acknowledges that not all AI-generated texts will feature them. Thus, they recommend examining various examples to identify recurring characteristics.
It is also essential to consider the language in use, as the words and phrases prevalent in AI-generated English texts may not directly correlate with those in Spanish. Therefore, these indicators should not be interpreted literally across different languages.
Examining Structure and Style
In addition to vocabulary, Pangram advises examining sentence construction. Human writing typically blends simple and complex sentences, maintaining a varied rhythm. In contrast, AI-generated texts often display multiple sentences with similar length and structure, leading to repetitive constructions across paragraphs.
Spelling and grammar also offer hints. Pangram notes that it is uncommon for AI models to make spelling errors unless specifically instructed to do so. A lengthy text devoid of typos, incomplete sentences, shifts in verb tense, or other minor inconsistencies may raise suspicions, especially when paired with a uniform tone and syntax.
Punctuation marks can likewise provide insights. AI tends to favour the use of em dashes to introduce clarifications or pauses, yet it often avoids semicolons and parentheses. However, these tendencies may vary between models and languages.
Moreover, the company observes that AI-generated texts often feature similarly sized paragraphs and adhere to a predictable structure: an introduction that sets the theme, several paragraphs organised in a similar fashion, and a conclusion that reiterates previously discussed points.
Assessing Specificity and Detail
Pangram highlights that AI models often resort to generalities when not given detailed instructions. Statements like “technology is changing our lives” or “this advancement will have a significant impact on society” may sound accurate but fail to clarify what has changed, whom it affects, or the basis of such assertions.
This lack of specificity extends to examples, characters, or experiences. A piece authored by a human typically includes concrete details related to the topic, such as names, locations, dates, memories, situations, or particular decisions. While an AI can generate such specifics if prompted, it may produce vague examples when given general instructions.
This phenomenon is particularly evident in texts purporting to recount personal experiences. Although a model can write in the first person, it lacks actual memories or lived experiences. Consequently, phrases like “this taught me the importance of perseverance” or “that experience changed my worldview” can appear suspicious if they lack detailed explanations or specifics about the experience in question.
Identifying Tone and Consistency
The tone of the text can also be a critical element to consider. When AI does not receive detailed guidance, it tends to maintain a formal language throughout the piece. Pangram further notes that it often employs positive or non-critical formulations.
While these characteristics alone do not definitively indicate whether a text was generated by AI—since they can also appear in human writing—the convergence of multiple factors, such as the repetition of ideas and phrases, a lack of specific details, and an overly uniform structure, can serve as red flags suggesting potential use of these tools.
Shifting Writing Habits in Response to AI
These patterns do have their limitations. As some users noted in the aforementioned art post, many writing techniques now associated with AI were in common use long before these tools emerged in 2022. This suspicion has even led some writers to alter their writing styles. Reports indicate that individuals are now incorporating typos, simplifying their vocabulary, or rephrasing sentences to avoid the appearance of AI-generated content, complicating the task of determining whether a text was produced by a human or a machine.
Furthermore, AI models are continuously evolving and may respond differently based on the instructions they receive. As a result, the signals that appear easily recognisable today may lose their clarity over time, making it increasingly challenging to discern when a text has been crafted by AI.