The reality of wholesale fraud with AI

Overall, school and job allocations seem to be completely outsourced to chatgpt. A new working paper by David Deming, Danoff Dean of Harvard University, reveals the more common reality of people’s AI habits.

“It’s healthier and more practical than I expected,” said Deming, a labor economist who also serves as a political and economics professor at Harvard Kennedy School’s Isabel and Scott Black. “I think that’s a good story. But if you think this is taking over the world, it’s probably a disappointment. And it’s not a very good story for those who predict a huge productivity boost.”

Deming’s large-scale study has been co-authored with Openai’s in-house economists, examining both ChatGpt usage around the world. Key findings show that they have even eased or even erased demographic gaps related to geography and gender. An individual set of analyses was drawn into a vast sample of anonymized messages, which more accurately positioned the daily role of technology as researchers and conducted gut checks.

“People have found it great to have assistants, advisors and guides,” Deming said. “Of course you can use it to automate things, but the prompts are important and you really need to go back and forth between them. On the other hand, there’s a very low friction just to ask for advice or feedback.”

Remaining bullish with the career prospects of 21st century college graduates, Deming recently released two famous enquiries about AI adoption and labor market disruptions. He was publishing his research last spring at Bay Area Headquarters, the artificial intelligence company that developed ChatGpt.

This was before Deming was presented as Harvard dean.

“A few months after the project, I told the team, “I’ve decided to take on this new job. But don’t worry. I’m still committed to the paper,” he remembered with a laugh.

Their findings show ChatGpt outweigh Google’s historic growth. As Deming said in his recent subsack, it took Google eight years to reach a billion messages every day, following the release in 1999 and then ChatGPT released in November 2022.

As of July 2025, they discovered that around 10% of the world’s adult population is using the technology, with adults ages 18 to 25 being responsible for almost half of the platform’s 2.6 billion daily messages.

“We knew young people would become heavier users,” Deming provided, and the analysis ruled out minors. “But the scale is amazing. It suggests that this generation will really become AI natives.”

Deming and her colleagues wondered whether the well-documented gender gap in ChatGPT adoption rates has been shut down as product growth. They approached the question by studying whether users traditionally had masculine or female names, according to a variety of datasets, including rankings of popular baby names for social security officials and girls.

In early 2023, just a few months after the launch of the CHATGPT product, economists discovered that around 80% of weekly active users traditionally have male names. As of July 2025, users with traditionally female names make up more than half of all users.

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

What’s also surprising is the fact that people in middle-income countries, including South Korea and Chile, are currently adopting technology faster than the technology of the wealthiest economies.

“There is no significant difference in use between people in Brazil and the US,” Deming said.

Other findings concern how people use technology and how their use differs depending on demographics. This part of the study was completed with care to protect user privacy and meant developing a taxonomy of various types of ChatGPT prompts.

What else is the job of categorizing nearly 1 million messages sent between May 2024 and June 2025?

“We asked if each message was work-related, or if they wanted tutoring or education, if they were asking for products to buy, or if they wanted personal advice,” Deming explained.

As of June 2024, data shows that there is evenly split between work and personal messages. A year later, personal use far outweighed work-related ones, particularly among younger adults, accounting for almost three-quarters of all messages sent via ChatGPT’s consumer plan.

To verify the accuracy of this research method, results were compared with human classification. Additional testing demonstrated the technique to accurately classify submissions drawn from WildChat, a public database of voluntarily submitted ChatGPT messages.

Approximately 80% of messages were divided into three categories. Those classified as “requesting information” increased from 14% to 24% between July 2024 and July 2025.

“This is basically the same as Google search,” Deming said. “But it’s a bit easier because you don’t have to scroll through a lot of websites. It just gives you the answer.”

Another popular category called “practical guidance” is stable, receiving a total of around 29% of messages over the same period.

“These messages are a little more customized,” Deming explained. “It could be something like this: “I’m 65 years old, hurting my left hamstring. Give me some stretches.”

“And if you don’t like the answer,” he added, “You can say so. You’re having a conversation with a chatbot that adapts to your requirements. That’s something that traditional web search can’t do.”

The best “writing” for work-related messages fell from 36% to 24% of messages during the survey period.

“In fact, most of the use of ‘writing’ is more than just writing,” Deming revealed. “It’s a summary of the document, a critique of Op-Eds, reducing 1,000 words to 800 or translating the text into Farsi.”

The work-related messages, much more common among well-paid, professionally educated users, were further scrutinized. These inputs are specifically mapped to work activities listed in the U.S. Labor Support Office’s Occupational Information Network (O*NET) database, with “Documentation/Recorded Information” and “Decision-making and Problem Resolution” appearing as top messaging categories for users in almost all occupations.

As Deming says, white-collar experts in the industry are applying ChatGpt to similar task sets.

“If you look at educators, top use cases aren’t in the category of ‘training and teaching others’,” he said. “It’s ‘documentation/recording of information.’ And when it comes to sales professions, the number one task is not “selling or affecting others.” It’s about “making decisions and solving problems.”

“The way people use it is so common that it applies to all jobs,” concludes Deming, who works with another colleague to launch an AI tracker, regularly serves insights wrapped in data on US use and the impact of the labor market.

The story is published courtesy of Harvard Gazette, the official Harvard newspaper. For additional university news, please visit harvard.edu.