expulp Verdict

Verdict blog, Research, explained, week of 2 November 2026

What makes writing original? What research can and cannot show

We apply revision prompts drawn from three studies to one flat paragraph and show where they add specificity but not proven originality.

What makes writing original: the short answer

Research has not settled what makes writing original, and none of the three studies discussed here measured originality directly. What they do suggest is that surprising, visual and emotional content helps separate great science journalism from typical articles, and that students notice when drafts lack personal voice and thought.

By originality we mean something specific: a piece adds an idea, observation or connection that a reader would not already find in existing texts on the subject. The studies measured nearby things instead, namely perceived greatness, students' self-reported impressions and correlations with performance. Each one shows a different gap between what was measured and originality, and the sections below take them in turn.

Great writing is not the same as original writing

Louis and Nenkova, 2013, writing in Transactions of the Association for Computational Linguistics, built a corpus of great science journalism alongside typical articles from the genre. Their abstract lists the features they implemented: surprising, visual and emotional content, plus general features of discourse organisation and sentence structure. They report that great and typical articles could be told apart fairly accurately and that the entire spectrum of features contributed.

The abstract gives no accuracy figure, only the phrase "fairly accurately", so readers who want the number should consult the full paper at the link above. The first limit this study illustrates is perceived greatness versus originality. A great article can be beautifully organised without saying anything new, and a surprising detail can be surprising without being original. The study also covers one genre only.

Personal voice in AI drafts: wishes, not tested outcomes

Wang et al., 2024, published in Education Sciences, is the source of the 47-student sample: students in a United States university writing course used generative AI tools such as ChatGPT to draft an application letter or personal statement, then answered five open-ended questions analysed with thematic and sentiment analysis.

Students described the drafts as general, robotic and lacking voice and critical thinking. The more distinctive finding is what they wanted to do about it: add personal stories, connections to the posting, feelings and thoughts, and delete repetitive language.

The second limit is self-reported wishes versus tested outcomes. Nobody in the study checked whether those additions made the letters better or more original. They are what students believed would help.

Writing beliefs: correlation, not cause

Villalón et al., 2013, in Educational Psychology, studied high school students' conceptions of writing and their writing self-efficacy beliefs. It found that writing self-efficacy beliefs and gender helped predict writing performance, and that students' writing conceptions moderated performance.

Its relevance here is narrow but real: what writers believe writing is appears to be bound up with how they perform, which plausibly affects whether they attempt anything beyond the expected. The third limit is correlation versus cause. The study reports associations with performance, not originality, and it does not show that changing a belief changes the writing.

A worked example: revising one flat paragraph

Here is a paragraph we wrote to resemble a generic personal statement. It is our illustration, not data from any study. Before: "I have always been passionate about environmental science. Throughout my studies I have developed strong research skills and a commitment to making a difference. I believe this program would help me achieve my goals."

Step one applies the prompt from Louis and Nenkova's feature list: add a surprising detail. "The summer I was fifteen, I counted dead bees on my grandmother's porch every morning and logged them on the back of a cereal box." Step two applies the students' wish from Wang et al.: add personal experience and thought. "By August the numbers had dropped, and I could not tell whether the bees were recovering or simply gone. That uncertainty, not knowing what a falling count means, is the question I want to learn to answer."

The result is clearly more specific, more visual and more personal. Is it more original? Honestly, we cannot say. A childhood observation leading to a career is a familiar structure, and many applicants could tell a version of it. The prompts changed the surface and supplied a voice, but they did not supply a new idea. The one line that comes closest, framing a falling count as ambiguous evidence, came from thinking about the material, which is the critical thinking students said AI drafts lacked and which neither prompt asks for directly.

What would count as evidence of originality

To test whether a revision is original, researchers would need measures the three studies did not use: expert ratings of novelty, systematic comparison against existing texts on the same topic, or measures of reader surprise. One workable design would collect generic drafts, randomly assign half to revision with the two prompts and half to an ordinary edit, then have raters who do not know which is which score novelty and overlap with a reference set of similar documents.

Until then, the prompts are reasonable revision habits rather than a tested formula. The useful question to ask of any draft is whether it adds something a reader could not find elsewhere, not just whether it is more detailed, and that distinction between detail and contribution is the one we care about at Expulp Verdict.

Questions people ask

What makes writing original according to research?

None of these three studies measured originality directly. Louis and Nenkova (2013) found that features including surprising, visual and emotional content helped distinguish great science journalism from typical articles, but greatness and originality are not the same thing.

How accurately can a computer tell great writing from typical writing?

Louis and Nenkova (2013) report that great and typical science journalism could be distinguished fairly accurately, with all their features contributing. The abstract does not give an accuracy figure, and the work covers only one genre.

Does adding personal stories make AI-assisted writing original?

In Wang et al. (2024), 47 university students said they wanted to improve AI drafts by adding personal stories, feelings and thoughts. These were their stated wishes; the study did not test whether such changes made the writing better or more original.

Do a writer's beliefs about writing affect their work?

Villalón et al. (2013) found that high school students' writing self-efficacy beliefs and gender helped predict writing performance, and that their writing conceptions moderated performance. These are associations with performance, not proof that beliefs cause better or more original writing.

Sources