In section headings, AI chatbots strongly tend to capitalize all main words
AI chatbots may display various phrases in boldface for emphasis in an excessive, mechanical manner. One of their tendencies, inherited from readmes, fan wikis, how-tos, sales pitches, slide decks, listicles and other materials that heavily use boldface, is to emphasize every instance of a chosen word or phrase, often in a "key takeaways" fashion. Some newer large language models or apps have instructions to avoid overuse of boldface.
AI chatbots output often includes vertical lists formatted in a specific way: an ordered or unordered list where the list marker (number, bullet, dash, etc.) is followed by an inline boldfaced header, separated with a colon from the remaining descriptive text.
Instead of proper wikitext, a bullet point in an unordered list may appear as a bullet character (•), hyphen (-), en dash (–), hash (#), emoji, or similar character. Ordered lists (i.e. numbered lists) may use explicit numbers (such as 1.) instead of standard wikitext. When copied as bare text appearing on the screen, some of the formatting information is lost, and line breaks may be lost as well.
While human editors and writers often use em dashes (—), LLM output uses them more often than nonprofessional human-written text of the same genre, and uses them in places where humans are more likely to use commas, parentheses, colons, or (misused) hyphens (-) and en dashes (–). LLMs especially tend to use em dashes in a formulaic, pat way, often mimicking "punched up" sales-like writing by over-emphasizing clauses or parallelisms.[12][10]
This sign is most useful when taken in combination with other indicators, not by itself. It is much more common on discussion pages than in article text. Also, because LLMs' use of em-dashes has become somewhat notorious, some AI companies have attempted to make their newer chatbots suppress their use, most notably OpenAI's GPT-5.1.[16]
In rare cases, some AIs may create unnecessary small tables that could be better represented as prose.
Curly quotation marks and apostrophes
ChatGPT and DeepSeek typically use curly quotation marks (“...” or ‘...’) instead of straight quotation marks ("..." or '...'). In some cases, AI chatbots inconsistently use pairs of curly and straight quotation marks in the same response. They also tend to use the curly apostrophe (’), the same character as the curly right single quotation mark, instead of the straight apostrophe ('), such as in contractions and possessive forms. They may also do this inconsistently.
Curly quotes alone do not prove LLM use. Directional quotation marks (curly or typographer) are often used in published works written and edited using the Chicago Manual of Style.[17] Microsoft Word has a "smart quotes" feature that converts straight quotes to curly quotes. So does the default system-wide configuration on macOS and iOS devices, except on some applications (or if turned off, as may be necessary for programming). Grammar correcting tools such as LanguageTool may also have such a feature. Curly quotation marks and apostrophes are common in professionally typeset works such as major newspapers. Citation tools like Citer may repeat those that appear in the title of a web page: for example,
McClelland, Mac (2017-09-27). "When ‘Not Guilty’ Is a Life Sentence". The New York Times. Retrieved 2025-08-03.
Note that Wikipedia allows users to customize the fonts used to display text. Some fonts display matched curly apostrophes as straight, in which case the distinction is invisible to the user. Additionally, Gemini and Claude models typically do not use curly quotes.