"Across 1.1 million conversations dating from May 2024 to June 2025, a full 28 percent dealt with writing assistance in some form or another, OpenAI said. That rises to a whopping 42 percent for the subset of conversations tagged as work-related (by far the most popular work-related task), and a majority 52 percent of all work-related conversations from users with "management and business occupations.""
"OpenAI is quick to point out, though, that many of these users aren't just relying on ChatGPT to generate emails or messages from whole cloth. The percent of all conversations studied involves users asking the LLM to "edit or critique" text, at 10.6 percent, vs. just 8 percent that deal with generating "personal writing or communication" from a prompt. Another 4.5 percent of all conversations deal with translating existing text to a new language, versus just 1.4 percent dealing with "writing fiction.""
"In June of 2024, about 14 percent of all ChatGPT conversations were tagged as relating to "seeking information." By June of 2025, that number had risen to 24.4 percent, slightly edging out writing-based prompts in the sample (which had fallen from roughly 35 percent of the 2024 sample). While recent GPT models seem to have gotten better about citing relevant sources to back up their information, OpenAI is no closer to solving the widespread confabulation problem that makes LLMs a dodgy tool for retrieving facts."
Across 1.1 million conversations from May 2024 to June 2025, 28 percent involved text assistance. That proportion rises to 42 percent for conversations tagged as work-related and to 52 percent for users in management and business occupations. Editing or critiquing existing content accounts for 10.6 percent of conversations, while generating personal communication represents 8 percent. Translation of existing content appears in 4.5 percent versus 1.4 percent for fiction composition. Information-seeking increased from about 14 percent in June 2024 to 24.4 percent in June 2025. Models show improved source citation but continue to suffer confabulation, reducing factual reliability. Work-related information-seeking remains lower at 13.5 percent.
#chatgpt-usage #workplace-text-assistance #editing-and-translation #information-seeking #llm-confabulation
Read at Ars Technica
Unable to calculate read time
Collection
[
|
...
]