Meta description: Several studies find AI-generated text scores a higher reading grade level than human writing. Here’s what the research shows, a worked example, and edits that bring the score down.
Run a paragraph of AI-generated text and a paragraph written by a person through a readability checker, and you will often see the AI version score higher. The longer sentences, the heavier transitions and the formal word choices push the grade level up, even when both paragraphs say roughly the same thing. This post covers what published research says about that pattern, where it doesn’t hold, and how to edit it.
What the research says
The pattern shows up in several published studies, though not in all of them:
- A 2025 study in Frontiers in Education compared 50 student essays with 50 AI-generated ones. The ChatGPT essays scored a Flesch-Kincaid grade level of 16.61, against 12.12 for the students. Interestingly, the students wrote longer sentences, so the difference came from vocabulary and structure, not sentence length alone.
- A PLOS ONE study of academic abstracts found AI-generated abstracts scored significantly higher on ARI, Flesch-Kincaid and Coleman-Liau, which the authors link to longer words and sentences.
- A Stroke journal analysis found AI essays at a median Flesch-Kincaid grade of 14.8, against 13.1 for human essays.
- A 2026 preprint on donation appeals found the model’s text read at a higher level and used noticeably longer words.
- A study of scientific summaries found LLM-written summaries scored harder than human-written ones on every readability metric tested.
The pattern also has limits. One Human Factors and Ergonomics Society study of passages written to an 8th-grade target found the ChatGPT passages could be read faster, with better comprehension, than human-written ones. In other words, the effect depends on the task and on what you ask for. Formal, “professional” requests tend to give the highest scores, and a request for plain language can give simple text.
Why it might happen (a hypothesis)
These are my own explanations, not proven causes:
- Training favors complete, well-formed text. Human casual writing is full of fragments and abrupt changes of thought, which a model is less likely to produce. A grammatically complete sentence naturally carries more clauses.
- The usual requests lean formal. Marketing copy, business email and reports are heavily written with AI tools. They already tend toward hedged, formal phrasing, and a request to sound “professional” and “thorough” invites longer sentences and heavier transitions.
A worked example
This is a paragraph I constructed to show how the score moves, not a measurement of any AI tool. The scores use the standard Automated Readability Index (ARI) formula, the same measure used by the reading level on the Open Character Counter home page.
| Version | Words | Sentences | Avg. sentence length | ARI |
|---|---|---|---|---|
| Plain: “We moved the launch to May. The new date gives the team two more weeks to test. If anything breaks, we will have time to fix it. Please share any concerns by Friday.” | 33 | 4 | 8.2 words | 1.0 |
| Formal: “Please note that the launch has been rescheduled to May in order to provide the team with an additional two weeks of testing time, which will allow us to identify and resolve any potential issues prior to release. It is important to note that any concerns with respect to the revised timeline should be communicated by Friday.” | 57 | 2 | 28.5 words | 14.9 |
| Formal, three filler phrases cut | 46 | 2 | 23.0 words | 12.8 |
| Formal, trimmed and split into three sentences | 35 | 3 | 11.7 words | 5.4 |
The plain version scores about a first-grade level and the formal one nearly 15. That is an extreme gap on purpose, and real text will differ. The edits show the useful lesson: deleting “in order to”, “it is important to note that” and “with respect to” lowered the score by about two grades. Splitting the long sentences into shorter ones lowered it by another seven. In ARI, average sentence length carries a lot of weight, so splitting is the bigger lever.
How to edit AI text so it reads better
- Split long sentences first. If a sentence has more than two commas, or you can’t say it in one breath, break it in two.
- Then cut filler phrases. Delete “in order to”, “it is important to note that”, “with respect to”, “prior to” and “additional” when a short word does the job.
- Read it aloud. Anything hard to say is hard to read.
- Check the score. Compare the edited version with your own first draft. A higher score isn’t wrong on its own, since technical writing sometimes needs it, but it should be a choice.
- Ask for the style you want. Telling a model to use short sentences and plain words often changes the result more than editing afterward.
A caution about scores
Readability formulas measure sentence and word length, not clarity, accuracy or quality. A low score can still be confusing, and a high score can be perfectly clear to the right audience. Use the numbers as a prompt to look again, not as a verdict.
Run your own test
If you want evidence beyond one example, collect several of your own prompts and compare the AI and human versions in a character counter that reports reading level. Note the number of samples and the prompts you used, and the result will be worth more than anyone’s impression, including mine.
Sources: Frontiers in Education (2025), “Linguistic complexity in student vs. AI-generated essays”; PLOS ONE, “Comparative analysis of text readability and writing styles in AI-generated vs. human-written academic abstracts”; Stroke (AHA Journals), “Scientific Writing in the Era of Large Language Models”; arXiv:2604.03202; arXiv:2512.23206; Human Factors and Ergonomics Society Proceedings, “Humanizing AI in Education” (2024).
About the author
Kent Williams is the founder and developer of Open Character Counter, a free browser-based tool for tracking character counts, word counts, and readability. He writes about writing constraints, platform limits, and SEO. Read more about Open Character Counter →
Before you publish
- Check the citations against the originals. I worked from search summaries, so open each study, confirm the numbers and add direct links. Fix the exact titles in the source line, because I paraphrased a few.
- Keep the first-person claims honest. I removed “almost every time” and the unsupported 9th-to-13th-grade figures. If you have actually compared samples, add the real counts and results.
- Rename the post. The old title promised “longer sentences”, which the research only partly supports.
- Your site’s real gap is credibility, not word count. Posts that cite sources, label hypotheses and show their data are the strongest answer to a “low value” review.
Leave a Reply