Here’s what you’ll learn when you read this story:
- Text generated by AI is difficult for traditional plagiarism detection, yet the best technologies are emerging and are still frequently employed.
- Technology, and AI in particular, is affecting the integrity of literature.
- No detection technology can be totally trusted, and most may be beaten by paraphrase or obfuscation.
For years, the university has talked about writing as if everyone agreed on what effective writing looked like, but in reality, there have always been generations of writing with evolving ideas and tools. There were rubrics, English standards, and plagiarism policies. There were similarity reports. Increasingly, there were AI-writing detectors. Whenever a new writing tool appeared, another warning followed: “Students can no longer be trusted.”
Then came another: AI can now imitate authentic writing. Eventually, it became clear that our old detection system was obsolete.
Catherine listened. She had spent years in education, but she didn’t consider herself an English guru. Her bosses knew English. The English people knew English. The people who talked confidently about plagiarism knew English. So when they said a piece of writing was plagiarized, she initially let them define what plagiarism was.
When they said a sentence sounded artificial, Fatima accepted that assessment. When they said a student’s writing was “not authentic,” she wondered whether she misunderstood what authentic writing was supposed to sound like. After all, they were the experts.
She was, in her own private joke, just a piece of junk. So she watched.
The First Generation: Where Did This Come From?
Long before generative AI became an everyday educational concern, plagiarism detection had a relatively concrete question: Does the submitted text correspond to material that already exists somewhere else?
Turnitin became one of the major systems built around that problem after its emergence in the late 1990s. A similarity report could identify matching material. But a match was not necessarily plagiarism. A quotation could produce a match. A properly cited passage could produce a match. A common phrase could produce a match. The report could show similarity. The human still had to determine what similarity meant.
Catherine understood that distinction only gradually. But she noticed something. The machine was not reading the student’s mind. It was comparing text.
Then the question changed.
Generative AI introduced a different problem. Suppose a student submitted an essay that was original and did not substantially copy any identifiable source. The student had not pasted an article. There might be no obvious source to match. But perhaps the student had asked an AI system to write the essay. Now the question shifts to: Can we determine whether this text was generated by AI? That was a different problem.
Similarity detection asked, “Where else does this text appear?” AI detection asked something closer to, “Does this text resemble patterns associated with AI-generated writing?”
Catherine began to notice that these were not interchangeable questions. Neither one, by itself, answered, “Did this student actually learn?”
Then ChatGPT entered the room.
Catherine eventually began using ChatGPT herself, not because she wanted a machine to think for her. It is quite the opposite. She wanted to understand what the machines were doing. Catherine tried asking questions. She challenged answers and asked for definitions. She asked the same question in different ways. Occasionally, she disagreed with the response. The machine misunderstood her. It produced something polished that sounded strangely familiar. And then she noticed something simple. ChatGPT could write. It could also explain. ChatGPT could restructure and correct grammar. It could change tone, could summarize, and could translate as well. It could produce an entirely new passage. Those were not necessarily the same activity. That distinction became important.
English is a language.
One afternoon, Catherine looked again at the arguments surrounding AI writing. She kept encountering phrases such as
- authentic voice
- natural writing
- human writing
- AI-like language
- proper English
- academic English
- and professional tone.
Something bothered her. English was being treated as though it were a measure of intellectual legitimacy. But English was a language. Nothing more. Nothing less.
A person could speak English brilliantly and think poorly. A person could speak English imperfectly and think brilliantly. A student could have an extraordinary idea and express it awkwardly. Another could produce beautifully polished prose while having little understanding of what the words meant. The language did not automatically reveal the quality of the thinking. And then Catherine encountered another idea: English itself contains genres.
Academic English is a genre. Legal English is a genre. Scientific English is a genre. Corporate English is a genre. Journalistic English is a genre. Bureaucratic English is a genre. Social media English is a genre. Every genre develops conventions. Certain words become preferred. Some sentence structures become normal. Certain ways of arguing become recognizable. Certain expressions become almost obligatory. And AI systems were extraordinarily adept at learning patterns.
Catherine paused. “Wait.” She pondered more.
If AI had learned enormous quantities of human language, including institutional and academic language, then perhaps some writing would sound “AI-like” partly because AI had become excellent at reproducing standardized human language.
The direction of causation suddenly looked less obvious. Maybe it wasn’t always: AI makes people sound alike. Sometimes it could also be that people have already been trained to sound alike, and AI has become excellent at imitating the pattern.
The Bosses’ Question
The next time the subject of plagiarism came up, Catherine listened differently.
School Director Peter said, “The new AI writing tools make the old systems obsolete.”
Administrator Carlos said, “Students are becoming better at hiding AI use.” Administrator Ernesto said, “We need a stronger authenticity detector.”
Catherine did not immediately argue. Instead, she asked, “What exactly are we trying to determine?”
Silence was deafening until Maria answered, “Whether the student used AI.” “Why?” Catherine asked.
There was another pause before Maria said, “Because we need to know whether the work is authentic.”
Catherine asked, “What does authentic mean here?”
The discussion became complicated very quickly.
Catherine further asked, “Did ‘authentic’ mean entirely human-generated? Did using a dictionary make writing less authentic? What about spell-checking and grammar correction? translation? Try asking someone to explain a difficult concept? What about asking an AI to suggest a clearer sentence? Did they ask an AI to rewrite the entire paragraph? What about generating the argument itself? The distinctions multiplied. And Catherine realized something as she paused; she thought:
The technological question had been easier than the educational question.
The Writing Assistant Problem
Catherine returned to her experience with ChatGPT. She had understood it initially as a writing assistant. That did not necessarily mean, “Write everything for me.” It could mean, “Help me see whether this sentence is understandable.” Could it explain why this grammar is incorrect? Catherine thought, “Provide me another way of expressing this idea.” Or: “Show me where my argument is unclear.” But there was a boundary.
If the learner supplied the idea, reasoning, evidence, decisions, and revisions, then language assistance was doing something different from replacing the intellectual work. If the machine supplied the thesis, evidence, reasoning, structure, and final answer, the learner’s role could become fundamentally different.
The important question was therefore no longer simply, “Was AI involved?” It became, “What intellectual work did the learner actually do?”
Student’s Essay
Trixie, a university student, submitted an essay. Her English was grammatically inconsistent. She had drafted the essay herself. Trixie used a dictionary. She used spelling assistance. She asked an AI tool to identify grammatical errors. Then she revised the essay herself.
Professor Maria submitted it to an AI detector. The result suggested a high probability of AI-generated writing. The professor looked at Trixie and asked, “Did you use AI?”
Trixie answered, “Yes.”
The professor looked concerned. “So you admit it?” Maria stared at Trixie.
Trixie looked confused. “I used it for grammar,” she clarified.
Professor Maria opened the report. “But the detector says the writing is AI-like,” she told Trixie.
Trixie replied, “Does that mean my grammar belongs to the machine?”
The room became quiet. The professor had no easy answer.
The Missing Evidence
Instead of looking only at the final essay, the professor asked Trixie to explain her process. Trixie produced:
- her original draft
- her handwritten notes
- her sources
- her revisions
- her grammar questions
- her explanations of the argument
- and the changes she had accepted or rejected from the writing assistant.
The professor asked, “Why did you use this source?”
Trixie explained.
“Why did you change this paragraph?” Professor Maria
Trixie explained.
“What did you disagree with when the AI suggested this sentence? Professor Maria asked Trixie. She explained.
“Can you defend your conclusion without looking at the paper?” the professor requested, and Trixie did.
The detector had provided one piece of information. But it had not provided the answer to the educational question Trixie had.
The Uncomfortable Discovery
Catherine began to wonder whether institutions had gradually confused several different things: language proficiency with intellectual ability, then writing similarity with plagiarism, then AI-like language with AI authorship, and finally AI authorship with lack of learning.
The arrows between those concepts were not necessarily justified. They were assumptions. Sometimes useful assumptions. Sometimes dangerous ones. But assumptions nonetheless.
Decision Point
The university’s academic committee must now revise its AI-writing policy. Three proposals are placed before the committee.
Option A — Strengthen Detection
Purchase the newest AI detector whenever the current detector becomes outdated.
Require instructors to investigate high AI-probability scores.
Option B — Keep Detection, Change Its Meaning
- Continue using similarity and AI-detection systems, but classify their results explicitly as screening evidence, not proof of misconduct.
- Require human review and additional evidence.
Option C — Change What Counts as Evidence
- Reduce dependence on final-product detection.
- Assess learning through combinations of:
- Drafts
- Revision history
- Source evaluation
- Discussion
- Oral explanation
- Reflection
- Demonstrations
- Learner decisions
- The ability to defend the submitted work.
The committee chair asks one final question: “If our real concern is whether the student learned, which of these three options gives us evidence about learning rather than merely evidence about text?”
Catherine looks at the committee. For the first time, she does not feel compelled to let the English experts define the entire problem for her. She has discovered something more fundamental:
- English is a language.
- Writing is a human activity.
- A genre is not a status symbol.
- Similarity is not automatically plagiarism.
- AI-likeness does not automatically mean AI authorship.
- And authorship is not identical to learning.
ChatGPT had not given her the answer. It had helped her see the questions that had been hiding inside the answers. And that, she realized, was perhaps the more compelling use of a writing assistant.
Discussion Questions
- At what point does language assistance become intellectual substitution?
- Can a student produce highly standardized academic English without using AI? What evidence would distinguish that from AI-generated writing?
- Is “AI-like” a linguistic description, an authorship claim, or an educational judgment?
- When an institution says “authentic writing,” whose language norms define authenticity?
- How might multilingual or code-switching students complicate a detector’s assumptions about “normal” writing?
- If a detector produces a probability rather than a fact, what should the institution do with that probability?
- What evidence can demonstrate learning that a final essay cannot?
- When does a writing assistant help learners take control of their learning—and when does it replace that control?
- What did Fatima initially surrender when she assumed that the “English gurus” automatically had the correct definition of plagiarism?
- Was ChatGPT useful because it knew the answer—or because it made it possible for Trixie to interrogate the question?
Central tension
- The institution wants certainty about authorship.
- The learner needs space to demonstrate agency.
- The machine can provide signals.
- But who decides what those signals mean?
DISCLAIMER
This writing is for educational purposes and encourages discussions and debate. It aims to improve critical thinking in several domains, including philosophical psychology and ethics. It is concerned with perception, judgment, and decision-making. Any mention is for debate purposes, not an accusation of wrongdoing.
The information is inconclusive. Readers are cautioned to distinguish fact from interpretation, to hold themselves accountable to proper procedure and permitted findings, and to be responsible in their research and discussion. The discussion around learning’s true value can be deeply personal, so take what resonates and explore further. As you do so, you may form your view about learning’s true value.
RELATED LINKS
© 2026 CLEVERPENS
Leave a Reply