Artificial intelligence has changed how we write. Emails, applications, research papers, blogs and student assignments can now be drafted or refined within seconds.
That has brought real benefits. AI can lower barriers to writing, improve clarity and help people communicate ideas that might otherwise remain trapped behind uncertainty, language difficulty or limited time.
But another problem may be emerging. More writing is beginning to share the same rhythms: polished openings, balanced lists, smooth transitions and conclusions that arrive with almost suspicious neatness. I have started thinking of the reader's response to this as AI authenticity fatigue.
A working idea, not a diagnosis
I am using this phrase as a proposed concept, not as an established psychological diagnosis. By AI authenticity fatigue, I mean the scepticism that may develop when readers encounter so much polished, formulaic or apparently AI-assisted prose that they begin to distrust writing before engaging fully with it.
The reader starts asking: Did this person actually write this? Is this another generic AI-assisted application? Does the author really think this way?
The uncomfortable possibility is that genuinely human work may also become suspect simply because it resembles what readers have learned to recognise as an ‘AI voice’.
We judge the assumed author as well as the text
There is some evidence for this concern. In experimental studies, people have rated text less favourably when they believed it came from AI rather than from a human author. One study found lower perceived message and source credibility when a GPT-generated science article was labelled as AI-written. Another found lower ratings of competence and content quality when texts were attributed to ChatGPT rather than to a person.
This does not prove the existence of AI authenticity fatigue. It does show that assumed authorship can shape evaluation.
Our ability to recognise AI writing is also imperfect. In a study of medical abstracts, blinded reviewers correctly identified many ChatGPT-generated abstracts, but they also incorrectly classified some original human-written abstracts as AI-generated. The suspected AI abstracts were often described as vague or formulaic.
That false-positive problem matters. Once a particular style becomes associated with AI, human writers who use similar structures may be judged through the same suspicion.
The problem is convergence
AI is not the enemy here. The deeper issue is convergence. Large language models generate likely continuations from patterns in their training data. When many people use similar systems for similar tasks, certain structures and phrases become common. Personal experiences are tidied into generic lessons. Uncertainty becomes a polished set of three points. Distinct voices begin to meet in the middle.
AI can improve a sentence while quietly removing the features that made it belong to a particular person.
This creates a strange cycle. Writers use AI to sound more professional. Readers become familiar with the resulting style. Familiarity creates suspicion. Writers then polish even harder to prove that their work is credible—and may end up sounding even more like everyone else.
Why this matters for education
For educators, the problem goes beyond detecting AI use. If polished prose is no longer reliable evidence of independent thinking, then assessment must make the thinking more visible. We may need to place greater value on oral explanation, staged decisions, application to unfamiliar cases, reflection on revisions and the student's ability to defend a position.
Instead of asking only, ‘Did the student use AI?’, we might ask: Can the student explain the reasoning, evidence and choices represented in this work?
This does not mean that writing no longer matters. Clear writing remains important. But a fluent artefact cannot, by itself, tell us how the learner arrived there.
Why this matters for academics and applicants
The same tension may appear in research, recruitment and professional communication. A cover letter filled with broad claims but no concrete experience reveals very little, regardless of whether AI produced it. A manuscript with elegant transitions but weak methods remains weak scholarship. A perfectly structured reflection may still hide the absence of reflection.
Specificity may therefore become more valuable than polish. What happened? What decision did you make? What did you misunderstand at first? What evidence changed your position? What remains uncertain?
These details do not guarantee human authorship, but they make writing more accountable. They give the reader something more meaningful to evaluate than style alone.
How I would use AI without losing my voice
The answer is not to abandon AI. It is to use it with more intention. I would begin with my own observation or argument before asking AI to intervene. I would use AI to test the structure, identify ambiguity and question weak reasoning. I would reject vocabulary that I would never normally use. I would restore uncertainty where the evidence is uncertain. Most importantly, I would keep responsibility for every claim.
Personal voice is not created by adding deliberate mistakes. Nor should we romanticise poor grammar as proof of humanity. Authenticity comes from the relationship between the writer's experience, reasoning, language and accountability.
AI can help edit that relationship. It should not erase it.
A fairness problem we should not ignore
There is also an ethical danger in talking about an ‘AI voice’. Formulaic writing may come from disciplinary convention, language learning, translation, assistive technology or a writer who has been taught to follow a rigid template. Judging AI use from style alone could unfairly affect multilingual writers, students with disabilities and anyone who relies on legitimate writing support.
The goal should not be to reward people who perform humanity in a particular style. It should be to evaluate ideas fairly and ask for appropriate evidence of process when authorship matters.
The research questions ahead
AI authenticity fatigue is still a hypothesis. That is what makes it interesting. We need to ask whether repeated exposure to formulaic AI-assisted prose changes reader trust. We need to identify which cues trigger suspicion and how often those cues produce false accusations. We need to know whether transparent disclosure, oral explanation or evidence of process can restore trust.
Artificial intelligence has changed how humans write. The next challenge may be preserving enough voice, context and accountable reasoning for readers to know what—and whom—they can trust.
Perhaps the question is no longer only, ‘Can AI write like a human?’ It may also be: Can humans still sound like themselves without being mistaken for AI?
Sources informing this article
- Gao et al. (2023), blinded review of AI-generated and original scientific abstracts: doi:10.1038/s41746-023-00819-6.
- Lermann Henestrosa and Kimmerle (2024), assumed AI versus human authorship and credibility: doi:10.3390/journalmedia5030069.
- Proksch et al. (2024), assumed authorship, competence and content-quality judgements: doi:10.3389/frai.2024.1412710.