The anxiety: Widespread dishonesty with AI in workplace and education. The truth: It’s intricate.

Overall, it seems that educational and professional tasks are not being completely delegated to ChatGPT. A recent working document by David Deming, Danoff Dean of Harvard College, reveals the more routine realities of individuals’ AI practices.

“It’s more constructive and pragmatic than I anticipated,” stated Deming, a labor economist who concurrently holds the title of Harvard Kennedy School’s Isabelle and Scott Black Professor of Political Economy. “I think that’s a positive narrative. Yet, it’s likely disappointing if you believe this technology will dominate everything. It’s also not particularly encouraging for those forecasting tremendous productivity advancements.”

Deming’s comprehensive research, co-developed with economists at OpenAI, examines both the demographics and the methods behind ChatGPT’s global usage. The primary findings indicate that swift adoption has mitigated or even eliminated demographic disparities related to location and gender. A different analysis utilized a vast array of anonymized communications to more precisely contextualize the technology’s everyday function as a researcher and sounding board.

“Individuals have realized that having an assistant, a consultant, and a mentor is invaluable,” Deming remarked. “Certainly, you can employ it for automation, but effective prompting is crucial, and you must engage in dialogue with it. Conversely, seeking its advice or feedback incurs very little effort.”

Deming, who remains optimistic about college graduates’ employment opportunities in the 21^st century,  recently unveiled two significant studies regarding AI integration and labor market changes. He was sharing his findings last spring at the Bay Area headquarters of OpenAI, the firm that created ChatGPT, when Ronnie Chatterjee, the company’s chief economist, suggested collaborating on a third study using internal data.

This occurred prior to Deming being announced as the dean of Harvard College.

“A few months into the initiative, I told the team, ‘I’ve accepted this new position. But don’t fret. I am still dedicated to the paper,’” he recounted with a chuckle.

Their outcomes demonstrate ChatGPT surpassing Google’s traditional growth rate. As Deming pointed out in a recent Substack post, it took Google eight years to achieve 1 billion daily messages after its public launch in 1999. ChatGPT, which debuted in November 2022, reached that milestone in under two years.

As of July 2025, they discovered that roughly 10 percent of the global adult population is utilizing the technology, with individuals aged 18 to 25 accounting for almost half of the platform’s 2.6 billion daily messages.

“I anticipated that young individuals would be more frequent users,” noted Deming, clarifying that the analysis excluded minors. “However, the magnitude is astonishing. It indicates this generation will be genuinely AI native.”

Deming and colleagues pondered whether the well-established gender disparity in ChatGPT adoption rates was diminishing as the product expanded. They approached the inquiry by examining whether users had conventionally masculine or feminine names, utilizing various data sets including the Social Security Administration’s yearly ranking of popular names for boys and girls.

In early 2023, just months post the ChatGPT launch, the economists observed that about 80 percent of weekly active users had traditionally male names. By July 2025, users with traditionally female names made up a little over half of all users.

“That crossover happened in the last few months,” Deming remarked.

Additionally surprising was the observation that individuals in middle-income nations, such as South Korea and Chile, are now embracing the technology at a faster pace than those in the wealthiest countries.

“There no longer exists a significant disparity in usage between individuals in Brazil and the U.S.,” stated Deming.

Other insights pertain to how individuals are utilizing the technology and the variations in usage across different demographics. This section of the research, conducted with a strong focus on preserving user privacy, involved creating a classification system for different types of ChatGPT prompts.

The task of categorizing nearly 1 million messages sent between May 2024 and June 2025, ultimately fell to — what else?
“““html

— ChatGPT-5.

“We inquired of the extensive language model whether each communication was work-related — or if it was soliciting tutoring or instruction, inquiring about product purchases, or seeking personal counsel,” Deming elucidated.

As of June 2024, the statistics indicate an equal division between professional and personal messages. A year later, personal usage had significantly surpassed any work-related interactions — particularly among younger adults — and comprised nearly three-quarters of all communications sent through ChatGPT’s consumer packages.

To confirm the validity of this research approach, findings were juxtaposed with classifications made by humans. Further assessments revealed that the technology correctly categorized submissions taken from WildChat, a public repository of voluntarily provided ChatGPT conversations.

Approximately 80 percent of communications were classified into three categories. Those identified as “seeking information” expanded from 14 to 24 percent between July 2024 and July 2025.

“This is pretty much akin to a Google search,” Deming noted. “However, it’s perhaps slightly more convenient since you aren’t required to navigate through numerous websites. It simply provides you the answer.”

Another favored category titled “practical guidance” remained consistent, comprising roughly 29 percent of messages during the same timeframe.

“These messages are somewhat more personalized,” Deming expounded. “It could be something along the lines of: ‘I’m a 65-year-old with a left hamstring injury. Suggest some stretches for me.’

“And if you don’t find the response satisfactory,” he added, “you can simply express that. You’re engaging in a dialogue with the chatbot that adapts to your inquiries. That’s something a conventional web search simply cannot achieve.”

“Writing,” the primary application for work-related messages, decreased from 36 to 24 percent of communications during the observed period.

“In reality, much of the ‘writing’ usage isn’t solely about drafting,” Deming clarified. “It involves summarizing documents, evaluating opinion pieces, condensing 1,000 words down to 800, or translating a text into Farsi.”

Work-related messages, particularly prevalent among educated users in high-earning professions, underwent further analysis. These inputs were meticulously linked to job activities recorded in the U.S. Department of Labor-supported Occupational Information Network (O*NET) database, with “documenting/recording information” and “making decisions and solving problems” emerging as predominant messaging categories by users in nearly every profession.

As Deming articulates, white-collar workers across sectors are utilizing ChatGPT for a similar array of tasks.

“If you examine educators, the primary case of use isn’t labeled ‘training and teaching others,’” he remarked. “It’s ‘documenting/recording information.’ And in sales roles, the foremost task isn’t ‘selling or persuading others.’ It’s ‘making decisions and solving problems.’

“The manner in which individuals are employing it is so broad that it pertains to every job,” wrapped up Deming, who’s collaborating with another group of colleagues to initiate an AI tracker, featuring regular snippets of data-driven insights regarding usage and labor market influences in the U.S. “It makes me even more doubtful of the assertion that AI is taking over entry-level positions.”

“`