Writing Styles of AI

Writing Styles of Report Summaries Generated by Artificial Intelligence

Michael Alley, Hala Hobrom, Andrea Khouzami, and Suryansh Sijwali
First published 1 August 2026; updated 22 September 2026*

An AI writing tool such as ChatGPT or Claude has a distinct writing style for each type of documents that the tool generates. Moreover, these styles are changing as new versions of each tool emerge [1]. As shown in Figure 1, the tool's output (word choices and sentence constructions) depend on the user's input (prompt) and the tool's algorithms. These algorithms appear to follow writing principles arising from what the tool's coders have selected and from the documents on which the tool is trained. While these writing principles are not published, our team has deduced several such principles for report summaries generated by different AI tools. We deduced these styles by analyzing the grammar and style of the outputs for a test bed of documents [2]. Presented in Table 1 and Table 2 are results of these tests for both ChatGPT 5.3 and Claude Sonnet 5.5.

Figure 1. Relationship of writing principles of an AI tool to its input, algorithms, and output.

So far, one of our most important results is that for AI-generated summaries of trusted reports, both ChatGPT and Claude do a good job of identifying the most important details of the reports. However, both AI tools have similar stylistic weaknesses:

word lengths and sentence lengths that are significantly longer than human writing, thereby making the reading more difficult.
lack of variety in sentence openers (essentially two types used by ChatGPT and three types used by Claude). This weakness limits the ways that the AI tool can connect sentences.

Besides these weaknesses and strengths, we noticed distinctive aspects of the AI-generated text occurred that were neither a strength nor a weakness:

almost no use of the first person (we or our), even though all the reports judiciously used those words;
higher percentage (11 to 12%) of sentences openers being a participial phrase; and
for ChatGPT, a higher percentage of sentences (28%) ending with a participial phrase.

A next step in our research is to add stylistic directives to our baseline prompt to address these weaknesses. Possible directives for the baseline prompt are as follows:

Keep average sentence length below 20 words.
Avoid long and pretentious nouns such as "identification" and "portability." Instead, use shorter words such as "identify and "portable" in the sentences.
Vary the types of sentence openers (including adverb, dependent clause, and infinitive phrase) to strengthen connections between sentences.

Refresher for grammar terms used to analyze AI writing [3].

References

Editors, "How to spot AI writing," Culture: Paper Trails, The Economist (30 July 2026).
Michael Alley, Hala Hobrom, and Suryansh Sijwali, “Three Stylistic Weaknesses of AI Writing in Engineering Documents and Prompts to Address Those Weaknesses,” extended abstract, 2026 ASEE St. Lawrence Regional Conference (Ithaca, NY: ASEE St. Lawrence Regional Conference, 25 April 2026).
Michael Alley, "Grammar Glossary," The Craft of Scientific Communication, http://www.craftscicom.org (2026).

* AI was not used to generate text for this webpage.