By Elizabeth Steere
Since OpenAI launched ChatGPT in 2022, educators have been grappling with the problem of how to recognize and address AI-generated writing. The host of AI-detection tools that have emerged over the past year vary greatly in their capabilities and reliability. For example, mere months after OpenAI launched its own AI detector, the company shut it down due to its low accuracy rate.
Understandably, students have expressed concerns over the possibility of their work receiving false positives as AI-generated content. Some institutions have disabled Turnitinโs AI-detection feature due to concerns over potential false allegations of AI plagiarism that may disproportionately affect English-language learners. At the same time, tools that rephrase AI writingโsuch as text spinners, text inflators or text โhumanizersโโcan effectively disguise AI-generated text from detection. There are even tools that mimic human typing to conceal AI use in a documentโs metadata.
While the capabilities of large language models such as ChatGPT are impressive, they are also limited, as they strongly adhere to specific formulas and phrasing. Turnitinโs website explains that its AI-detection tool relies on the fact that โGPT-3 and ChatGPT tend to generate the next word in a sequence of words in a consistent and highly probable fashion.โ I am not a computer programmer or statistician, but I have noticed certain attributes in text that point to the probable involvement of AI, and in February, I collected and quantified some of those characteristics in hopes to better recognize AI essays and to share those characteristics with students and other faculty members.
I asked ChatGPT 3.5 and the generative AI tool included in the free version of Grammarly each to generate more than 50 analytical essays on early American literature, using texts and prompts from classes I have taught over the past decade. I took note of the characteristics of AI essays that differentiated them from what I have come to expect from their human-composed counterparts. Here are some of the key features I noticed.
AI essays tend to get straight to the point.
Human-written work often gradually leads up to its topic, offering personal anecdotes, definitions or rhetorical questions before getting to the topic at hand.
AI-generated essays are often list-like.
They may feature numbered body paragraphs or multiple headings and subheadings.
The paragraphs of AI-generated essays also often begin with formulaic transitional phrases.
As an example, here are the first words of each paragraph in one essay that ChatGPT produced:
โFirstlyโ
โIn contrastโ
โFurthermoreโ
โOn the other handโ
โIn conclusion.โ
Notably, AI-generated essays were far more likely than human-written essays to begin paragraphs with โFurthermore,โ โMoreoverโ and โOverall.โ
AI-generated work is often banal.
It does not break new ground or demonstrate originality; its assertions sound familiar.
AI-generated text tends to remain in the third person.
Thatโs the case even when asked a reader responseโstyle question. For example, when I asked ChatGPT what it personally found intriguing, meaningful or resonant about one of Edgar Allan Poeโs poems, it produced six paragraphs, but the pronoun โIโ was included only once. The rest of the text described the poemโs atmosphere, themes and use of language in dispassionate prose. Grammarly prefaced its answer with โIโm sorry, but I cannot have preferences as I am an AI-powered assistant and do not have emotions or personal opinions,โ followed by similarly clinical observations about the text.
AI-produced text tends to discuss โreadersโ being โchallengedโ to โconfrontโ ideologies or being โinvitedโ to โreflectโ on key topics.
In contrast, I have found that human-written text tends to focus on hypothetically what โthe readerโ might โsee,โ โfeelโ or โlearn.โ
AI-generated essays are often confidently wrong. Human writing is more prone to hedging, using phrases like โI think,โ โI feel,โ โthis might mean โฆโ or โthis could be a symbol of โฆโ and so on.
AI-generated essays are often repetitive.
An essay that ChatGPT produced on the setting of Rebecca Harding Davisโs short story โLife in the Iron Millsโ contained the following assertions among its five brief paragraphs: โThe setting serves as a powerful symbol,โ โthe industrial town itself serves as a central aspect of the setting,โ โthe roar of furnaces serve as a constant reminder of the relentless pace of industrial production,โ โthe setting serves as a catalyst for the charactersโ struggles and aspirations,โ โthe setting serves as a microcosm of the larger societal issues of the time,โ and โthe setting โฆ serves as a powerful symbol of the dehumanizing effects of industrialization.โ
AI writing is often hyperbolic or overreaching.
The quotes above describe a โpowerful symbol,โ for example. AI essays frequently describe even the most mundane topics as โgroundbreaking,โ โvital,โ โesteemed,โ โinvaluable,โ โindelible,โ โessential,โ โpoignantโ or โprofound.โ
AI-produced texts frequently use metaphors, sometimes awkwardly.
ChatGPT produced several essays that compared writing to โweavingโ a โrichโ or โintricate tapestryโ or โpaintingโ a โvivid picture.โ
AI-generated essays tend to overexplain.
They often use appositives to define people or terms, as in โMargaret Fuller, a pioneering feminist and transcendentalist thinker, explored themes such as individualism, self-reliance and the search for meaning in her writings โฆโ
AI-generated academic writing often employs certain verbs.
They include โdelve,โ โshed light,โ โhighlight,โ โilluminate,โ โunderscore,โ โshowcase,โ โembody,โ โtranscend,โ โnavigate,โ โfoster,โ โgrapple,โ โstrive,โ โintertwine,โ โespouseโ and โendeavor.โ
AI-generated essays tend to end with a sweeping broad-scale statement.
They talk about โthe human condition,โ โAmerican society,โ โthe search for meaningโ or โthe resilience of the human spirit.โ Texts are often described as a โtestament toโ variations on these concepts.
AI-generated writing often invents sources.
ChatGPT can compose a โresearch paperโ using MLA-style in-text parenthetical citations and Works Cited entries that look correct and convincing, but the supposed sources are often nonexistent. In my experiment, ChatGPT referenced a purported article titled โPoe, โThe Fall of the House of Usher,โ and the Gothicโs Creation of the Unconscious,โ which it claimed was published in PMLA, vol. 96, no. 5, 1981, pp. 900โ908. The author cited was an actual Poe scholar, but this particular article does not appear on his CV, and while volume 96, number 5 of PMLA did appear in 1981, the pages cited in that issue of PMLA actually span two articles: one on Frankenstein and one on lyric poetry.
AI-generated essays include hallucinations.
Ted Chiangโs article on this phenomenon offers a useful explanation for why large language models such as ChatGPT generate fabricated facts and incorrect assertions. My AI-generated essays included references to nonexistent events, characters and quotes. For example, ChatGPT attributed the dubious quote โHalf invoked, half spontaneous, full of ill-concealed enthusiasms, her wild heart lay out thereโ to a lesser-known short story by Herman Melville, yet nothing resembling that quote appears in the actual text. More hallucinations were evident when AI was generating text about less canonical or more recently published literary texts.
This is not an exhaustive list, and I know that AI-generated text in other formats or relating to other fields probably features different patterns and tendencies. I also used only very basic prompts and did not delineate many specific parameters for the output beyond the topic and the format of an essay.
It is also important to remember that the attributes Iโve described are not exclusive to AI-generated texts. In fact, I noticed that the phrase โIt is important to โฆ [note/understand/consider]โ was a frequent sentence starter in AI-generated work, but, as evidenced in the previous sentence, humans use these constructions, too. After all, large language models train on human-generated text.
And none of these characteristics alone definitively point to a text having been created by AI. Unless a text begins with the phrase โAs an AI language model,โ it can be difficult to say whether it was entirely or partially generated by AI. Thus, if the nature of a student submission suggests AI involvement, my first course of action is always to reach out to the student themselves for more information. I try to bear in mind that this is a new technology for both students and instructors, and we are all still working to adapt accordingly.
Students may have received mixed messages on what degree or type of AI use is considered acceptable. Since AI is also now integrated into tools their institutions or instructors have encouraged them to useโsuch as Grammarly, Microsoft Word or Google Docsโthe boundaries of how they should use technology to augment human writing may be especially unclear. Students may turn to AI because they lack confidence in their own writing abilities. Ultimately, however, I hope that by discussing the limits and the predictability of AI-generated prose, we can encourage them to embrace and celebrate their unique writerly voices.
Elizabeth Steere is a lecturer in English at the University of North Georgia.
